The AI Extraction Investigation: The Complete Research File (October 2026)

READER'S MAP -- FOR HUMANS, START HERE

This file is the complete research record of the AI Extraction Investigation.

It is raw material for people or AI to investigate -- records, not verdicts.

It is long on purpose. You are not expected to read it end to end.

If you want the short version first, read this:

https://ai-memory-ark.blogspot.com/2026/10/the-ai-extraction-investigation.html

Where things live:

- Parts 1-6: the question, the method, and the historical window (verified Epstein-AI correspondence, Levels A through D).

- Parts 7-9: the money ledgers -- reported figures and what verification found.

- Parts 10-22: the investigation architecture, datasets, and execution rounds.

- Parts 23-24: the thesis -- AI pointed at people, not for them; the Massachusetts deployment layer.

- After Part 24: the transaction ledger, the Massachusetts ledger, and the full Epstein-AI correspondents file.

How to read the confidence grades (the file’s locked definitions): A = directly established (executed agreement, filing, dataset). B = strongly supported (multiple primary sources, or a primary source plus corroboration). C = reported but incompletely corroborated. D = analytical inference. U = unresolved. (Corrected 2026-10-06: this map previously gave different definitions for B and C; these now match the file’s own evidence hierarchy.)

A relationship between two names is not automatically a finding of anything.

Names appear because the records name them.

THE AI EXTRACTION INVESTIGATION: THE COMPLETE RESEARCH FILE

Frozen against the investigation ledger as of October 5, 2026.

This is everything. Every number we checked, every ledger entry, every

open question, every dead end. It is published as one file because the

point is not to convince you of anything. The point is to give you --

person or AI -- the raw material to investigate for yourself.

Questions work better than accusations. What it all means is not up to

us to decide.

HOW TO READ THIS FILE

- Records, not verdicts. We document; we do not accuse.

- Confidence grades: A = directly established. B = strongly supported.

  C = reported but incompletely corroborated. D = inference.

  U = unresolved.

- Evidence tiers: Tier 1 = filings, contracts, government records.

  Tier 2 = presentations, calls, testimony. Tier 3 = reputable

  secondary reporting. Tier 4 = estimates, aggregators, unnamed sources.

- "Unresolved" means the public record does not establish it. It is a

  finding, not a failure.

- Financial circularity is not the same as commercial circularity is not

  the same as demand circularity. A relationship between two companies is

  not automatically a circularity finding.

The plain-language guide to this investigation is published separately:

https://ai-memory-ark.blogspot.com/2026/10/the-ai-extraction-investigation.html

================================================================

THE INVESTIGATION (PARTS 1-24)

================================================================

THE AI EXTRACTION INVESTIGATION

Opened October 5, 2026. Status: active.

Question at the center: who captures the productivity gains from AI, who

bears the costs, and who owns the infrastructure in between?

Entry point: Epstein's AI correspondence (it was what came to mind).

Organizing principle: the extraction question, not Epstein.

This file is the case record. Plain text, dated entries, records not

verdicts. Nothing goes in the personal-cases section except what Ricky

directs in his own words.

PART 1: THE HISTORICAL WINDOW (VERIFIED)

Jeffrey Epstein's AI correspondence, verified against the DOJ January 30,

2026 Epstein Transparency Act releases. Full master list: 75 entries across

four levels (A: Epstein personally discusses AI; B: someone discusses AI

with Epstein; C: foundation and funding material; D: network background

only), plus a forward trace of money and people to 2026.

Full detail lives in: epstein-ai-correspondents.txt (same folder).

Load-bearing findings:

- Epstein personally proposed fully funding "Synthetic General Intelligence"

gatherings in August 2009 (EFTA01816129), alongside a "Conference on Power"

covering reputation, awe, trust, deception, and reciprocity.

- His exact verified words to Ben Goertzel, December 2013: "i want to see

the deception character, how do we send the funds?" The stated context was a

game-environment demo.

- The December 2013 money chain: 45,000 dollars routed via Humanity+ to Hong

Kong Polytechnic University; Humanity+ returned it for unstated reasons; a

60,000 dollar STC-to-Novamente wire was discussed ("tomotw") with completion

unconfirmed.

- Goertzel wrote in August 2014 that the Epstein Foundation never actually

funded iCog Labs, contradicting the foundation's own press materials.

- Every documented Epstein AI funding line ended by 2016. The forward trace

supports independent continuation with no centralized coordination: the

network survived him, nobody replaced him.

- What the records do NOT show: no AI built for exploitation, blackmail, or

intelligence operations; no secret Epstein AGI; the ASU 2017 workshop

happened without him; the PrimeAGI deck was Goertzel and Rutt's, sent to

him, not his.

PART 2: THE STRUCTURAL PICTURE (LEADS, NOT YET VERIFIED)

The following claims come from cross-checks against public reporting and

are NOT yet independently verified. Each is a lead to check, not a finding.

- Census-based research (Nov 2025-Jan 2026): about 18 percent of U.S. firms

using AI in at least one business function, 32 percent employment-weighted;

50-60 percent among very large firms in information, professional services,

finance.

- Census 2026 working paper: employment of 22-24 year olds fell 12 percent

over ten quarters after ChatGPT's introduction in the most AI-exposed

industry/state combinations, driven by reduced hiring, not layoffs.

- Federal Reserve Governors Barr and Cook (Sept 2026): AI may be limiting

entry-level opportunities in exposed sectors; limited evidence of

economy-wide displacement so far.

- IMF 2026: about 2.7 trillion dollars per year in labor-time value saved by

AI, gains concentrated in higher-paid occupations.

- ILO 2026: task-level gains of 10-70 percent not translating into clear

aggregate productivity growth; concentration in large digitally advanced

firms.

- IEA: data-center electricity up 17 percent in 2025; AI-focused data

centers up about 50 percent; roughly 485 TWh in 2025 toward 950 TWh by 2030.

- Lawrence Berkeley National Lab analysis: U.S. data centers could consume

11.8 percent of all U.S. electricity by 2030.

- FTC (July 2026): seeking comment on personalized/surveillance pricing.

- Senate hearing (Aug 2026): "Your Data, Their Profit: The Consumer Cost of

AI Surveillance Pricing."

- OECD (July 2026): competition risks from concentration of compute, data,

and skills; three largest cloud providers about 74 percent of global cloud

share in 2023; NVIDIA about 90 percent of GPU market.

- NVIDIA/Apollo/BlackRock/Blackstone/Brookfield/Goldman/KKR: platforms to

mobilize over 500 billion dollars for AI infrastructure; compute described

as an "investable asset class."

- NIST AI Agent Standards Initiative (Feb 2026): interoperability,

open-source protocols, agent security, identity, authorization.

- Linux Foundation: open Agent Name Service. ITU: trust and identity work

for humans and autonomous agents.

PART 3: THE AGENT IDENTITY THREAD (RESEARCHED OCTOBER 5, 2026)

The one thread where the historical documents touch present infrastructure:

Epstein's 2009 "reputation, trust, deception" agenda on one end; 2026 agent

identity, reputation, and authorization standards on the other. Findings

below carry evidence levels. No continuity is claimed between 2009 and 2026;

the thematic overlap is an open question only.

NIST. The Center for AI Standards and Innovation launched the AI Agent

Standards Initiative on February 17, 2026. It is convening and facilitating,

not regulating: three pillars are industry-led standards plus U.S.

leadership in international bodies, community-led open-source protocol

development, and research in agent security and identity. Published so far:

the launch announcement; a Request for Information on AI Agent Security

(closed March 9, 2026); NIST AI 800-5, the summary analysis of RFI responses,

published May 18, 2026; and an NCCoE concept paper, "Accelerating the

Adoption of Software and AI Agent Identity and Authorization" (February 5,

2026, comment closed April 2, 2026, now reviewing comments), covering

identification, authorization, auditing, non-repudiation, and

prompt-injection controls. No binding standard has issued. Evidence:

primary and reputable reporting.

ITU. A Focus Group on Trust and Identity for Humans and Agentic AI (TIDA)

was announced July 9, 2026 at the AI for Good Global Summit. It reports to

ITU-T Study Group 17; co-chairs are Debora Comparin and Amir Banifatemi;

first meeting Paris November 2026. Scope: terminology, identity models,

credential formats, assurance levels, lifecycle controls, cross-border

mutual recognition. Delegation doctrine: mandates specific, time-bound,

traceable, revocable; identity separated from trustworthiness. No normative

output yet; pre-standardization. Evidence: primary ITU text read in full,

reputable reporting.

Linux Foundation Agent Name Service. Intent announced June 23, 2026; not yet

operational. An open standard for agent identity, verification, and

discovery built directly on DNS, with no proprietary registry. Supports DIDs

and LEIs. Originated as a GoDaddy and Infoblox IETF draft. Backed by

Cloudflare, GoDaddy, Cisco, Salesforce, and Infoblox. Evidence: primary

press release read in full.

MIT. The Media Lab's NANDA/AGNI projects (PI Ramesh Raskar) specify the

missing infrastructure as: agent identity, naming, and discovery;

interoperable agent-to-agent communication; authentication, authorization,

and secure delegation; provenance, trust signals, and policy-aware privacy;

multi-agent coordination, negotiation, and orchestration; reputation;

agent-mediated commerce, pricing, and incentive design; governance of

systemic risks. Raskar's warning: "The window to keep this web of agents

open is closing soon." Evidence: primary MIT source read live.

WHO CONTROLS EACH LAYER TODAY

Foundation models: locked (OpenAI, Anthropic, Google dominate agent

underpinnings per the MIT Agent Index).

Cloud and compute: concentrated (AWS, Azure, GCP; NVIDIA silicon).

Enterprise agent identity: proprietary (Microsoft Entra Agent ID, Okta;

tenant-bounded).

Open identity primitives: open or proposed-open (Agent Name Service

proposed, ERC-8004 live January 29, 2026, DIDs/LEIs).

Settlement and payment rails: going open (x402 Foundation under the Linux

Foundation, operational launch July 14, 2026, 40 founding members including

Visa, Mastercard, Amex, Stripe, Google, AWS, Coinbase, Circle).

Authorization mandates: proprietary (Google AP2, Visa Trusted Agent

Protocol, Mastercard Agent Pay, Stripe MPP).

Commerce and discovery: proprietary (OpenAI ACP, Google UCP).

Interop protocols: open (MCP, A2A, AGENTS.md).

The pattern: card networks are moving to own authorization while conceding

settlement to open rails. Evidence: reputable reporting.

THE REPUTATION LAYER: LEAST SETTLED, MOST CONSEQUENTIAL

Documented open questions: who sets and verifies scores and how their

independence is assured; gaming and farming of scores; market-shaping (one

vendor's trust-based price multipliers run 1.00x to 0.60x: once reputation

changes pricing it stops being a safety mechanism and becomes a

price-setter); appealability (no deployed appeal mechanism documented);

algorithmic blacklisting (trust bands gating autonomy are the intended

function; contestability unresolved); portability (ERC-8004 makes records

portable; no consensus on whether trust transfers across versions);

ownership of the behavioral record (operator logs vs on-chain vs federated

registries; no consensus).

Academic work, 2025-2026: Hu, Rong, Van Kleek, "Dissociative Identity"

(arXiv:2605.30169, FAccT 2026): reputation-as-sanction is structurally

inapplicable to language-model agents, which lack persistent identity and

behavioral continuity; recommends observability-based harnesses instead.

Kumar Surapani et al., "Authorization Architectures for Tool-Using AI

Agents" (arXiv:2609.15906v1, September 2026): almost no deployed system

meets traceability to a human principal, bounded delegation, and

contestability end to end; runtime enforcement unresolved. OpenID

Foundation, "Identity Management for Agentic AI" (October 2025): revocation

across delegation chains largely unsolved. Bottom line: agents are being

given real-world authority before the identity and responsibility problems

are solved. Evidence: primary abstracts and reputable reporting.

HISTORICAL ANCHOR

Verified document EFTA01816129, read in full: email from Jeffrey Epstein to

John Brockman, August 11, 2009, 9:38 PM, subject "Re: CONFIDENTIAL."

Item 2: "Synthetic General Intelligence.. encompassing, signal processing,

statistical and machine learning, virtual world robots." Item 3:

"Conference on Power -- Its definition political financial, intellectual,

physical, includes reputation, awe, trust, deception, reciprocity."

Reported strictly as a documented anchor. No continuity claimed.

PART 4: PERSONAL CASES (RICKY'S TO DIRECT)

Empty until he fills it. One case per entry, in his words. For each case we

trace five layers: the experience, the institution involved, the decision

mechanism, the technology used, the financial incentive. Then the honest

column: AI involved, not involved, or unproven. Finding "AI isn't involved"

counts as a result.

No case is added without his direction. His elaboration rule stands: he can

correct or expand any entry rather than having it frozen.

PART 5: STANDING QUESTIONS

1. Who benefits from AI productivity gains?

2. Who bears the costs (money, water, electricity, land, health)?

3. Who owns the infrastructure (compute, models, data, identity rails)?

4. Who makes automated decisions about individuals?

5. Can individuals see and challenge those decisions?

6. Can workers capture any of the productivity gains?

7. Are public resources being used to build private AI infrastructure, and

what does the public get back?

8. Is AI increasing competition or increasing concentration?

9. Is AI giving ordinary people more agency, or making institutions more

powerful relative to individuals?

PART 6: METHOD

Records, not verdicts. Falsify the hypothesis, don't feed it: actively look

for evidence that AI is NOT the cause, that a public-interest path leads

nowhere near the Epstein network, that a funding line simply ended.

Evidence levels on every claim: primary document, financial record,

reputable reporting, inference, speculation. Documented fact, strong

inference, and speculation stay in separate piles. A cold trail is also an

answer. Timed correlations are presented as records, not verdicts.

Sixth layer, added October 5: every power map needs "who is being extracted

from." Supply-side maps (compute, models, data, agents, distribution)

without the humans reproduce the erasure this archive exists to fight.

Three competing hypotheses are investigated simultaneously, not one

conclusion defended: (A) productive revolution -- AI as gigantic

infrastructure investment with real expected productivity gains, disruption

inevitable in transition; (B) financial bubble -- capital chasing AI on

greater-fool expectations, supplier financing and guarantees disguising

the risk, losses propagating through AI companies to infrastructure to

banks to funds to investors if revenue disappoints; (C) extractive

platform economy -- AI concentrating data, IP, compute, capital, labor

productivity, and pricing power, the public supplying infrastructure and

research while few companies capture the upside. All three can contain

some truth. The investigation does not start with a predetermined

conclusion.

A dead-ends register is kept alongside the evidence: leads that looked

suspicious but did not pan out are recorded as dead ends (a person in a

document with no later connection found; an incentive that was ordinary

and disclosed). This protects the investigation from becoming a machine

that only collects confirming evidence.

PART 7: THE MONEY LEDGER (REPORTED, VERIFICATION RUNNING)

A cross-check produced a transaction-level ledger of AI infrastructure

finance, 2025-2026. Its accounting discipline is the real contribution and

is adopted here: a capital investment, a future spending commitment, a lease

obligation, a guarantee, a credit facility, and an announced project value

are DIFFERENT THINGS. They are not added together as cash spent.

Entries below are REPORTED, not verified. Verification of the major entries

against SEC filings, company announcements, and government documents is

running. Status will be marked confirmed, corrected, or unconfirmed.

Reported entries:

1. OpenAI-Amazon, Feb 27, 2026: 50B Amazon investment (15B initial, 35B

conditional); AWS agreement expansion +100B over eight years on top of

existing 38B (about 138B total); about 2GW Trainium capacity.

2. Amazon-Anthropic, Apr 20, 2026: >100B AWS commitment over ten years; 5B

Amazon investment (facility to 20B); up to 5GW additional compute; 8B

previously invested.

3. Anthropic IPO filing, reported Sept 29, 2026: at least 518B in

infrastructure commitments over about a decade, about 80 percent

non-cancelable. Reported split: Broadcom 161.2B lease obligations, Google

111.1B, Amazon 110B, Microsoft 31.4B, xAI/NVIDIA capacity to 84.5B, AMD 5B

equity + 20B capacity. NOTE: via press reporting of a confidential filing,

not the public filing itself.

4. NVIDIA with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs,

KKR, announced Aug 10, 2026: platforms to mobilize >500B in third-party

capital for AI compute; compute as an "investable asset class."

5. NVIDIA-SB Energy-OpenAI, Ohio, Aug 2026: 4.25GW initial (+~3.8GW option);

OpenAI customer for 8GW on a 20-year lease; NVIDIA guarantee to 105B per

SEC filing; 1.5B NVIDIA investment in SB Energy; 80M community-benefits fund. NOTE: a guarantee is contingent exposure, not cash spent.

6. Broadcom, June 8, 2026 SEC filing: backstop guarantee with maximum 29B

exposure on AI rack purchases.

7. FT, Sept 2026: to ~300B in off-balance-sheet residual-value guarantees

for AI data-center and chip debt (Meta, NVIDIA, Broadcom named); special

purpose vehicles keep exposure off ordinary balance sheets.

8. Stargate, Jan 21, 2025: OpenAI/Oracle/SoftBank/MGX announced 500B target

over four years; 100B initial deployment. NOTE: announced intention, not

cash in an account.

9. Stargate Michigan, June 1, 2026 (Saline Township): multi-billion-dollar

campus; equity from Related Digital + Blackstone-affiliated funds; long-term

debt anchored by PIMCO-managed funds.

10. Meta: 2025 capex 69.69B; 2026 guidance raised to 130-145B; about 103.77B

in lease obligations not yet commenced at end of 2025.

11. Alphabet: 2025 capex 91.4B; 2026 expected 175-185B; Google Cloud backlog

240B.

12. Amazon: 2025 capex 128.3B; H1 2026 capex 96.3B.

13. Oracle: 638B remaining performance obligations; 75B of large AI

contracts with customer prepayment; FY2026 debt financing 43B.

14. DOE Paducah, Kentucky, July 29, 2026: >100B private investment; 1.8GW

AI/HPC campus; 2GW new gas generation; DOE/Brookfield/NextEra involvement.

Federal land and infrastructure entering the buildout.

15. Savannah River Site, July 20, 2026: NNSA selected Amentum to negotiate a

phased lease for a 1GW AI data center with onsite generation.

16. FTC, Aug 2026: proposed enforcement policy on personalized/surveillance

pricing. Senate, Aug 2026: "Your Data, Their Profit" hearing. DOT closed

its airline privacy investigation Oct 5, 2026 without penalties, while

warning against unlawful discriminatory dynamic pricing.

Key structural observations to test, not conclusions:

- The same corporations appear as investor + infrastructure supplier +

creditor + customer + competitor (notably Amazon with both OpenAI and

Anthropic). That is interdependence, not evidence of coordination.

- Hardware is becoming collateral: leases, debt, guarantees, residual-value

assumptions around chips and racks (NVIDIA, Broadcom).

- Finance is moving upstream: from funding AI companies to financing

compute itself as an asset class.

- Government is becoming a node: federal land, DOE involvement, state tax

incentives, utility agreements. Terms matter more than existence.

- Announcement aggregates (e.g., a 3.1T White House project table) mix

announced intentions with committed capital; they are an index, not a total.

What the ledger does not prove: that AI is a deliberate extraction scheme;

that companies are secretly coordinating; that ordinary Americans are being

intentionally harmed; that AI is the primary cause of U.S. economic

problems. What it establishes: an extraordinary financial and

infrastructure system being built very fast, with enough contracts and

dollar amounts to investigate without speculation.

PART 8: MASTER LEDGER V2 (REPORTED, VERIFICATION RUNNING)

Supersedes Part 7 with a fuller transaction ledger from the cross-check.

Same discipline: investment, commitment, lease, guarantee, facility, and

announced project value are different things. Status marks below are AS

REPORTED by the cross-check, not independently verified: Primary means the

cross-check claims an SEC filing, government document, or company

announcement; Reported means credible press reporting of a document not yet

in hand; Flag means an investigative lead, not an allegation. Verification

of the major entries is running; entries will be marked confirmed,

corrected, or unconfirmed.

A. OPENAI / STARGATE

1. Jan 21, 2025 -- OpenAI, SoftBank, Oracle, MGX -- 500B / 10 GW --

announced Stargate vehicle -- Primary as reported.

2. Jan 21, 2025 -- same -- 100B -- initial deployment announced -- Primary

as reported. Distinguish from the 500B target.

3. Jul 22, 2025 -- OpenAI, Oracle -- 4.5 GW -- additional U.S. data-center

capacity -- Primary as reported.

4. Jul 22, 2025 -- OpenAI/Oracle -- >300B / 5 yrs -- reported investment

with expansion -- Primary as reported.

5. Sep 23, 2025 -- OpenAI, Oracle, SoftBank -- >400B / 3 yrs -- nearly 7 GW

planned -- Primary as reported.

6. Oct 30, 2025 -- OpenAI/Oracle -- >450B / 3 yrs -- >8 GW planned --

Primary as reported.

7. Jan 9, 2026 -- OpenAI, SoftBank, SB Energy -- 1B (500M each) -- invested

in SB Energy -- Primary as reported. AI lab directly financing an

infrastructure developer.

8. Jan 9, 2026 -- OpenAI/SB Energy -- 1.2 GW -- long-duration

infrastructure relationship -- Primary as reported.

9. Jan 9, 2026 -- SB Energy/Ares -- 800M -- redeemable preferred equity --

Primary as reported.

10. Aug 17, 2026 -- NVIDIA/SB Energy -- 1.5B -- NVIDIA investment in SB

Energy -- Primary as reported. Chipmaker becomes infrastructure investor.

11. Aug 17, 2026 -- OpenAI/SB Energy/NVIDIA -- 4.25 GW -- initial

PORTS-Pike (Ohio) capacity -- Primary as reported.

12. Aug 17, 2026 -- OpenAI/SB Energy -- ~8 GW -- OpenAI as customer --

Primary as reported. Long-duration customer lock-in.

13. Aug 17, 2026 -- NVIDIA/SB Energy/OpenAI -- 105B -- NVIDIA residual-value

guarantees -- Primary as reported, FLAG. Huge contingent exposure, not cash

spent.

14. Aug 17, 2026 -- NVIDIA/SB Energy -- ~3.8 GW -- additional capacity

option -- Primary as reported.

15. 2026 -- OpenAI/Oracle -- 638B remaining performance obligations at

Oracle -- large cloud-contract ecosystem -- Primary as reported. Not all

OpenAI; must not be attributed wholesale.

B. OPENAI / AMAZON

16. 2026 -- Amazon to OpenAI -- 50B -- investment commitment -- Primary as

reported.

17. Q1 2026 -- Amazon to OpenAI -- 15B -- initial investment, cash --

Primary as reported.

18. Q2 2026 -- Amazon to OpenAI -- 13.7B -- additional investment -- Primary

as reported, via Amazon filing.

19. 2026 -- Amazon/OpenAI -- 38B -- existing AWS commitment -- Primary as

reported.

20. 2026 -- Amazon/OpenAI -- +100B / 8 yrs -- AWS expansion -- Primary as

reported. About 138B total with the existing agreement.

21. 2026 -- Amazon/OpenAI -- ~2 GW -- Trainium capacity -- Primary as

reported.

22. 2026 -- AWS/OpenAI Frontier -- exclusive third-party cloud distribution

-- Primary as reported. FLAG: Amazon as investor + cloud provider + chip

provider + distribution channel simultaneously.

C. AMAZON / ANTHROPIC

23. Q2 2026 -- Amazon to Anthropic -- 5B -- Series G preferred stock --

Primary as reported.

24. 2026 -- Amazon to Anthropic -- to 20B -- financing facility -- Primary

as reported.

25. Q2 2026 -- Amazon to Anthropic -- 15B remaining -- facility after the 5B

investment -- Primary as reported.

26. 2026 -- Amazon/Anthropic -- >100B / 10 yrs -- AWS commercial commitment

-- Primary as reported.

27. 2026 -- AWS/Anthropic -- 5 GW-class -- compute infrastructure --

Reported. FLAG: Amazon financing both OpenAI and Anthropic, two

competitors, while supplying both with compute.

D. ANTHROPIC'S CONTRACTUAL WEB (via reported Sept 29, 2026 IPO filing)

28. Google -- 111.1B -- Reported.

29. Amazon -- 110B -- Reported.

30. Microsoft -- 31.4B -- Reported.

31. Broadcom -- 161.2B -- lease obligations -- Reported.

32. xAI/NVIDIA capacity -- to 84.5B -- Reported.

33. AMD computing capacity -- 20B -- Reported.

34. AMD equity -- 5B -- Reported.

35. Total Anthropic infrastructure plan -- 518B+ over ~10 yrs, ~80 percent

reportedly non-cancelable or usage-independent -- Reported. NOTE: via press

reporting of a confidential filing, not the public filing.

New: Broadcom reportedly agreed to lend Anthropic to 42B to finance its TPU

infrastructure obligations -- Reported. Supplier finances customer: Anthropic

needs chips, Broadcom supplies chips, Broadcom finances the buyer.

E. NVIDIA AND THE FINANCIALIZATION OF COMPUTE

36-41. Aug 10, 2026 -- NVIDIA with Apollo, BlackRock, Blackstone,

Brookfield, Goldman Sachs, KKR -- each part of >500B -- AI infrastructure

financing platforms -- Primary as reported. NVIDIA: turning compute and

full-stack infrastructure into an "investable asset class."

42. Aug 2026 -- NVIDIA -- 105B -- SB Energy/OpenAI residual-value

guarantees -- Primary as reported (see A13).

43. 2026 -- NVIDIA -- 3.5B -- other AI-cloud land/power/shell guarantees --

Primary as reported.

New: NVIDIA reportedly working with insurers (Howden Re named) on chip

depreciation data and risk structures so NVIDIA hardware can collateralize

loans to smaller cloud companies -- Reported. Total NVIDIA maximum

guarantee exposure across AI-cloud guarantees reported at 108.5B -- Reported.

F. COREWEAVE: DEBT-BACKED AI INFRASTRUCTURE

44. Mar 30, 2026 -- 8.5B -- HPC/GPU-backed delayed-draw facility -- Primary

as reported.

45. Apr 14, 2026 -- 1.75B -- senior notes at 9.75 percent -- Primary as

reported.

46. May 15, 2026 -- 3.1B -- GPU/infrastructure delayed-draw facility, called

a publicly syndicated HPC-infrastructure-backed financing vehicle --

Primary as reported.

47. Jun 11, 2026 -- 3.5B -- senior notes -- Primary as reported.

48. Aug 7, 2026 -- 2.6B -- additional delayed-draw facility -- Primary as

reported.

49. Jun 30, 2026 -- 35.6B -- total CoreWeave indebtedness -- Primary as

reported. FLAG: the question is how much future AI revenue must arrive for

debt-financed GPU infrastructure to earn its return.

G. META

50. Q1 2026 -- 19.0B -- servers, data centers, network infrastructure --

Primary as reported.

51. 2026 -- 130-145B -- expected annual capex -- Primary as reported.

52. 2026 -- 24.9B -- debt proceeds in first half -- Primary as reported.

53. 2026 -- 27B -- venture's estimated development costs (per SEC filing) --

Primary as reported.

54. 2026 -- 12.31B -- initial lease commitments -- Primary as reported.

55. 2026 -- 28B -- residual-value guarantee threshold -- Primary as

reported. FLAG: same structure as elsewhere -- construction, venture/SPV,

lease, residual-value guarantee.

H. MICROSOFT

56. FY2026 Q4 -- 41B quarterly capex -- AI/cloud infrastructure -- Primary

as reported.

57. FY2026 -- 34.6B -- committed construction and building improvements --

Primary as reported.

58. FY2026 -- 215.9B -- servers and network equipment at cost -- Primary as

reported.

59. FY2026 -- 35.6B -- Q4 cash PP&E expenditure -- Primary as reported.

60. FY2026 -- 5.6B -- finance leases in Q4 -- Primary as reported. About

two-thirds of quarterly capex in short-lived assets (CPUs, GPUs), the rest

long-lived infrastructure, per Microsoft.

I. ALPHABET / GOOGLE

61. 2025 -- 91.4B -- capex -- Primary as reported.

62. 2026 -- ~180B -- planned capex -- Primary as reported.

63. 2026 -- 175-185B -- earlier projected range -- Primary as reported.

64. 2026 -- >70B annualized -- Google Cloud revenue run rate -- Primary as

reported.

65. 2026 -- 111.1B -- reported Anthropic commitment (see D28).

J. ORACLE

66. May 31, 2026 -- 638B -- remaining performance obligations -- Primary as

reported.

67. FY2026 -- 43B -- debt financing raised -- Primary as reported.

68. FY2026 -- 5B -- equity financing raised -- Primary as reported.

69. FY2027 -- ~40B -- expected additional financing -- Primary as reported.

70. 2026 -- 75B -- AI contracts involving prepaid or customer-supplied GPUs

-- Primary as reported. FLAG: prepayment changes the financing equation --

customers fund the build, Oracle's capital risk falls, customers lock into

the ecosystem.

K. GOVERNMENT, PUBLIC LAND, ENERGY

71. Apr 3, 2025 -- DOE -- 16 federal sites -- AI/data-center development RFI

-- Primary as reported.

72. Mar 20, 2026 -- DOE/DOC/SB Energy/AEP Ohio -- 10 GW generation --

Portsmouth public-private project -- Primary as reported.

73. Mar 20, 2026 -- same -- 10 GW data center -- on DOE land -- Primary as

reported.

74. Apr 7, 2026 -- DOE/SB Energy -- world's-largest AI data center announced

on DOE land -- Primary as reported.

75. Jul 29, 2026 -- DOE/Brookfield/NextEra and others -- >100B -- Kentucky

AI/data-center investment, privately funded per DOE -- Primary as reported.

76. Jul 29, 2026 -- same -- 1.8 GW -- initial data-center campus -- Primary

as reported.

77. Jul 29, 2026 -- same -- 2 GW -- planned generation -- Primary as

reported.

78. Jul 29, 2026 -- same -- to 2.6 GW -- battery storage -- Primary as

reported.

79. Mar 2026 -- DOE/SB Energy -- 9.2 GW gas -- part of Ohio generation plan

-- Primary as reported.

FLAG: the federal government is putting land and infrastructure into the

development model. The question is not whether this is good or bad but:

what economic value transfers from public assets to private AI operators,

under what contractual terms, and who bears the downside risk?

L. CONSUMER EXTRACTION AND PRICING

80. Aug 2026 -- FTC -- personalized pricing using personal data -- Primary

as reported.

81. Aug 2026 -- FTC -- proposed enforcement policy on surveillance pricing

-- Primary as reported.

82. 2026 -- consumer companies -- individual willingness-to-pay modeling --

Reported.

83. 2026 -- data ecosystem -- browsing, purchase, and location data

potentially used in pricing -- Reported.

Separate branch from infrastructure finance: AI and data analytics

extracting value directly from consumers through individualized decisions.

The two branches should not be conflated.

PATTERNS HARD TO DISMISS (NOT PROOF OF MISCONDUCT)

1. Compute is being turned into a financial asset. NVIDIA says so

explicitly. CoreWeave finances GPUs through syndicated debt. Residual-value

guarantees from NVIDIA, Broadcom, and Meta. Hardware to collateral to debt

to infrastructure to revenue.

2. The industry is becoming circular. AI company commits future spending;

cloud/infrastructure company borrows to build; chip company supplies

equipment and finances or guarantees the customer; financial institutions

fund the infrastructure; AI company receives compute; future AI revenue is

expected to repay everyone. A BIS analysis reported by the Financial Times

independently identifies circular financing in AI as an emerging systemic

concern: suppliers financing customers, customers financing suppliers,

financing alongside existing commercial relationships.

3. Government is becoming part of the physical infrastructure layer:

federal land, energy infrastructure, public-private partnerships,

permitting, redevelopment. Terms matter more than existence.

RED-FLAG MATRIX (LEADS, NOT ALLEGATIONS)

Circular financing: evidence strong. Supplier financing customer: strong.

Residual-value guarantees: strong. AI infrastructure debt: strong.

Investor/provider/customer overlap: strong. Government land in AI

infrastructure: strong. Energy-cost allocation: needs project-level work.

Consumer personalized pricing: real regulatory concern. AI deliberately

designed to extract from citizens: not established. Epstein network

directly controlling today's AI finance: not established. Coordinated secret

plan: not established. The financial network is sufficiently real that no

Epstein connection is needed to explain it.

THE TOP INVESTIGATIVE QUESTION

Who bears the downside if AI demand fails to justify the infrastructure

being built? For every major project, trace five things: the debt holder,

the equity holder, the government and public contribution, the guarantee

provider, and the ultimate asset owner. Then compare AI company revenue

required to service the system against actual contracted AI revenue, debt

and lease obligations, guaranteed residual values, and public

infrastructure expenditure. That distinguishes a normal infrastructure boom

from a highly leveraged but rational buildout from something more fragile.

PART 9: LEDGER VERIFICATION RESULTS (OCTOBER 5, 2026)

All 15 major entries were checked against SEC filings, company

announcements, government documents, and reputable reporting. Verdict: all

dollar figures check out, but the dominant pattern is confirmed -- most

headline numbers are future spending commitments, non-cancelable service

obligations, contingent guarantees, or announced project values, NOT cash

invested. Actual cash deployed is a small fraction of headline totals.

CONFIRMED, WITH CORRECTIONS THAT MATTER

1. Anthropic 518B. The Reuters September 29, 2026 reporting checks out:

at least 518B over about ten years, about 80 percent non-cancelable or

payable regardless of usage. Corrections: the xAI 84.5B piece is largely

cancelable on 90 days notice; the Broadcom 161.2B is equipment lease

obligations; the prospectus itself is not public, so this rests entirely on

Reuters' reporting of a confidential filing. No primary document exists.

2. NVIDIA's 500B+ platforms. Confirmed as announced August 10, 2026 -- but

they are NON-BINDING memorandums of understanding, not committed capital

and not NVIDIA cash. NVIDIA may offer residual-value support up to 25

percent of an individual opportunity. "Investable asset class" is NVIDIA's

own characterization. One report notes the MOUs may never become definitive

agreements.

3. Ohio SB Energy campus. The 105B NVIDIA guarantee is confirmed as a CAP in

the SEC 8-K, covering about 4.25GW initial load, effective on lease

commencement, terminating on events including OpenAI achieving a

satisfactory credit rating; OpenAI indemnifies NVIDIA. Earlier July press

reports of 250B were superseded by the 8-K. NVIDIA's total maximum exposure

is 108.5B (105B Ohio + 3.5B other cloud-partner lease guarantees).

Correction: NVIDIA's actual cash into SB Energy is 3B total (1.5B private

placement + 1.5B prepaid forward), not 1.5B. The 80M community fund is

confirmed. The 1.5M GPUs per generation figure is a company statement

quoted in press, not in SEC filings.

4. OpenAI-Amazon. Confirmed: Amazon up to 50B (15B initial + 35B

conditional), part of a 110B OpenAI round at 730B pre-money (SoftBank 30B,

NVIDIA 30B). Amazon's 10-Q confirms 28.7B invested in H1 2026; the 50B

completed around August 2026. AWS side confirmed: existing 38B expanded by

100B over eight years (about 138B total), about 2GW Trainium capacity, AWS

as exclusive third-party cloud distributor for OpenAI Frontier. The

circular structure is documented: Amazon's cash flows to OpenAI, OpenAI contracts 100B+ of AWS spend back to Amazon.

5. Amazon-Anthropic. Confirmed: 5B immediately plus to 20B conditional on

milestones, on top of 8B previously invested; >100B AWS spend over ten

years; up to 5GW capacity. Amazon's Q2 10-Q records a 10B Anthropic

preferred investment in Q2 2026.

6. Stargate. Confirmed: January 21, 2025 White House announcement, up to

500B over four years, 100B initial deployment -- an announced target, only

about 100B initially committed. Michigan campus (Saline Township, June 1,

2026, "The Barn"): about 16B construction and development plus about 40B

Oracle compute fit-out; press reports about 2B equity and about 14B debt

with PIMCO around 10B in bonds. Financing confirmed as project-level equity

plus debt, per Oracle's press release.

7. Meta. 2025 capex 69.69B confirmed via SEC XBRL compilation (some outlets

cite ~72.2B; definitional discrepancy flagged). 2026 guidance raised to

130-145B confirmed. Update: Meta's Q2 2026 10-Q shows unstarted leases of

about 278.99B at June 30, 2026, plus about 68B more in July 2026 -- far

above the 103.77B at end of 2025.

8. Alphabet. 2025 capex 91.4B confirmed. 2026 guidance raised mid-2026 to

180-190B (from 175-185B). Google Cloud backlog 240B confirmed -- note this

is contracted future revenue (RPO), not spend.

9. Amazon. 2025 capex 128.3B confirmed. H1 2026 cash capex 96.3B vs 55.6B

prior-year period confirmed via 10-Q. Full-year 2026 guided about 200B,

raised to about 220B by September 2026.

10. FT 300B off-balance-sheet. Confirmed as the FT's estimate (flag it as

an estimate, not a filing total). Components: Meta Hyperion about 28B

residual-value guarantee backing about 27B debt (SPV "Beignet Investor,"

Blue Owl 80 percent / Meta 20 percent); Broadcom about 29B; NVIDIA 105B;

Alphabet data-center lease guarantees 16.9B rising to 43.8B in six months.

Type: maximum potential liability, not debt or cash spent.

11. DOE Paducah, Kentucky. Confirmed via DOE announcement: ">100B

privately funded investment" is DOE's wording -- an announced private

investment target, not federal spend and not committed cash. 1.8GW campus,

2GW new gas generation, to 2.6GW battery storage, completion 2031, power

agreement subject to Kentucky PSC approval.

12. Savannah River. Confirmed: NNSA selected Amentum to enter NEGOTIATIONS

for a phased lease -- pre-lease stage, not a signed lease, not funded

construction.

13. Oracle. Confirmed: 638B remaining performance obligations at end of

FY2026, up 363 percent year over year; 75B of large AI contracts prepaid or

customer-supplied hardware; FY2026 raised 43B debt + 5B equity. Note: RPO

is contracted future revenue for Oracle, not spend.

14. Broadcom backstop. Confirmed with a critical correction: the 29B is a

lease-payment BACKSTOP (contingent guarantee), NOT a loan and NOT debt. The

42B convertible-note facility for Anthropic TPU leasing is a SEPARATE,

undrawn credit facility. Do not add 29B + 42B; they are different

instruments on the same stack.

15. FTC, Senate, DOT. Confirmed: FTC voted 2-0 on August 19, 2026 to publish

a proposed enforcement policy statement on personalized ("surveillance")

pricing -- a proposal, not a ban; comment extended to September 25, 2026.

Senate Judiciary subcommittee hearing August 4, 2026, "Your Data, Their

Profit," chaired by Hawley. DOT closed its 2024 airline privacy review with

no enforcement actions; notice posted September 4, 2026; Reuters reported

the memo October 5, 2026.

DO-NOT-SUM OVERLAPS (COUNTING THESE TOGETHER DOUBLE-COUNTS)

- Anthropic's 518B vs the April 2026 Amazon-Anthropic deal: the 110B Amazon

component is the same underlying relationship as the >100B ten-year AWS

commitment.

- Anthropic-Broadcom stack: the 161.2B lease obligations encompass the TPU

racks in the Apollo/Blackstone XPV platform (35B initial); the 29B backstop

and 42B note facility are guarantees and credit support ON that stack, not

additional obligations.

- NVIDIA Ohio 105B vs the FT 300B: the 105B is a component of the 300B

aggregate.

- NVIDIA 30B OpenAI equity (part of the 110B round) vs the 105B Ohio

guarantee: separate instruments, same counterparty -- equity vs contingent

guarantee.

- Amazon 50B into OpenAI vs OpenAI's 100B+ AWS commitment: cash flows one

way, spending commitment flows back. Same two companies, opposite

directions.

- Oracle 638B RPO vs Stargate 500B target: Oracle's RPO includes large

OpenAI AI contracts -- same underlying demand in both; RPO is Oracle's

contracted future revenue, 500B is announced program spend.

- Alphabet 240B backlog, Amazon ~496B backlog, Meta 103.77B unstarted

leases: backlog is future customer revenue; unstarted leases are

obligations. Different sides of the ledger.

- NVIDIA Aug-10 six-firm MOU platform vs Broadcom Apollo/Blackstone XPV:

parallel vehicles with overlapping arrangers (Apollo, Blackstone) but

distinct; both largely non-binding.

- OpenAI 110B round: Amazon 50B + SoftBank 30B + NVIDIA 30B = 110B at 730B

pre-money, confirmed consistent across outlets.

COULD NOT VERIFY

- Meta FY2025 capex exact figure: 69.69B confirmed in XBRL compilation;

some outlets cite ~72.2B; the definitional difference could not be pinned

down.

- The 1.5M GPUs per generation figure: company statement in press only.

- The Broadcom June 8 filing text itself: confirmed via multiple secondary

reports citing it; the filing was not pulled directly.

- XPV 35B initial transaction, PIMCO 10B bond tranche, Michigan 16B + 40B

split: press-reported financing figures, no primary filing confirmed.

- The Anthropic confidential prospectus: no public document; entirely

dependent on Reuters' reporting.

BOTTOM LINE

The ledger's numbers check out, and the fine print matters more than the

headlines. The system being built is: AI companies make enormous future

compute commitments; infrastructure companies borrow and build against

those commitments; chip companies guarantee or finance the equipment;

private capital funds the infrastructure; governments contribute land,

energy, and frameworks; AI companies receive the compute; the whole

structure depends on future AI demand arriving at the assumed scale.

Whether that is a productive infrastructure boom, a leveraged but rational

buildout, or something more fragile is now a balance-sheet question, and

the balance sheets are partially visible in the filings above.

PART 10: EXPANDED INVESTIGATION ARCHITECTURE (BLANK-AI PROPOSAL, OCT 5 2026)

The blank cross-check AI, given only the ledger and the extraction

question, independently proposed broadening the investigation into

interconnected maps rather than deepening the Epstein thread first. Its

central move: follow the money, compute, data, energy, and political

influence OUTWARD into what they flow into, not only inward into who

supplies them. Recorded here because it converged with the investigation's

own direction without knowledge of the user's work. Several of its maps

(e.g., data extraction, decision power, productivity capture) overlap

material the user already developed independently in the Extraction

Machine -- noted where relevant.

MAP 1: CAPITAL. For every major AI infrastructure project: equity

investors, debt lenders, bondholders, private-equity funds, asset managers,

insurers, credit facilities, guarantees, leases, SPVs, beneficial owners,

refinancing and maturity dates, collateral, default provisions,

termination rights, residual-value assumptions. Question: who actually

owns the risk? Who supplied the money, who borrowed it, who guarantees it,

what secures it, who gets paid first if things go wrong, who takes the

loss. Status: begun (Parts 7-9); needs project-level tracing.

MAP 2: CIRCULAR MONEY. A weighted directed financial graph: every

transaction an arrow, every arrow carrying date, amount, instrument,

contractual obligation, and source. Then ask mathematically which entities

occupy the most central positions -- more informative than counting

appearances. Status: the circular pattern is documented in filings (Part

9); the formal graph is not yet built.

MAP 3: GOVERNMENT SUBSIDY. A separate database of public inputs into

private AI infrastructure: federal and state land, tax abatements,

property-tax and sales-tax exemptions, infrastructure grants, loans and

loan guarantees, electricity subsidies, transmission/road/water/sewer

construction, workforce subsidies, economic-development grants,

accelerated depreciation, energy credits, federal procurement commitments.

Then compute the public contribution per project. A 10B announced data

center with 0 public money and one with 2B are economically different

projects. Status: not yet built; DOE land entries (Part 8, section K) are

the start.

MAP 4: ELECTRICITY. For each major AI campus: data center to utility to

generation to transmission to rate structure to customer. Who supplies

electricity, who owns generation, who pays for new generation and

transmission, special tariffs, interruptible or discounted power, whether

costs are socialized across ratepayers, dedicated plants and their owners,

what happens if the data center shuts down. Core question: do residential

customers pay any portion of infrastructure built primarily to serve AI

load. Answerable with utility filings. Status: not yet built.

MAP 5: WATER. Data center to water utility to withdrawal to consumption to

wastewater to local ratepayer. Gallons per day, source, price,

infrastructure required and who paid, drought restrictions, municipal

agreements, wastewater arrangements. Then compare economic value created

per unit of water against public infrastructure cost. Status: not yet

built.

MAP 6: LABOR. Measure before and after AI deployment: employment, wages,

productivity, occupations, hours, headcount, wage distribution, contractor

use, layoffs, job creation, management layers, outsourcing, gig and

temporary labor. Separate jobs created BY AI infrastructure from jobs

displaced BY AI applications -- different ledgers. A 10B data center may

create thousands of construction jobs and few permanent ones while the

systems running there restructure tens of thousands elsewhere. Status: not

yet built.

MAP 7: DATA EXTRACTION. What exactly is being extracted from people:

search, location, purchases, browsing, communications, workplace activity,

biometrics, medical, financial, social graphs, behavioral predictions,

inferred preferences, inferred willingness to pay. For each: who collects,

who owns, who processes, who buys access, who decides from it -- and can

the individual opt out. Overlaps the user's existing Extraction Machine

material. Status: partially covered there; needs systematic mapping.

MAP 8: THE INFERENCE ECONOMY. Separate observed data from inferred data.

AI creates new information about people (e.g., inferred financial

desperation from purchases, browsing, timing, debt, device behavior) that

the person never disclosed. Who owns the inference is potentially one of

the biggest economic questions of the AI era. Status: not yet built; hard

-- inference ownership is a legal gray area.

MAP 9: DECISION POWER. Catalogue where algorithms make consequential

decisions: employment, credit, insurance, housing, healthcare, education,

policing, benefits, immigration, transportation, pricing, advertising,

financial services. For each: who built it, who bought it, who trained it,

what data, what decision, can a human override, can the affected person

appeal, who is legally responsible for an error. If an algorithm denies

someone something important, where does accountability terminate. Overlaps

the user's existing material. Status: partially covered there.

MAP 10: LOBBYING AND POLICY. For every major AI company and infrastructure

investor: lobbying expenditure, firms, former officials hired, officials

who later joined companies, campaign contributions, PACs, trade

associations, think tanks, nonprofits, regulatory comments, legislation

supported and opposed. Build a revolving-door timeline documenting the

sequence without assuming corruption. Status: not yet built.

MAP 11: PROCUREMENT. Every major federal AI contract: agency, contractor,

subcontractor, value, award date, duration, competitive or not, sole-source

justification, option years, modifications, actual spending, performance

requirements, data rights, IP rights, classification. Follow the money

down through subcontractors -- the headline contractor may not perform the

work. Status: not yet built.

MAP 12: OWNERSHIP. For every major project: land owner, data-center owner,

SPV, debt holder, equity holders, operator, AI customer, chip supplier,

financier -- down to ultimate beneficial ownership, not the brand on the

press release. Especially important for private funds. Status: not yet

built; partially constrained by private-fund disclosure limits.

MAP 13: PATENTS AND IP. AI and semiconductor patents, licensing and

cross-licenses, exclusive licenses, university IP, government-funded

research, patents transferred into private companies. Question: how much of

the commercialized technology originated in publicly funded research, and

who privatizes the resulting value. Status: not yet built.

MAP 14: THE UNIVERSITY AND RESEARCH PIPELINE. Trace government research to

universities to laboratories to researchers to startups to venture capital

to corporations: grants, DARPA, NSF, DOE, NIH, university licensing,

corporate sponsorship, endowed chairs, research institutes. Then compare

against the historical Epstein network -- AFTER building the modern map.

Status: not yet built.

MAP 15: THE EPSTEIN OVERLAY. A separate layer, not the foundation. For

every person in the Epstein/AI corpus: correspondence date, topic,

financial relationship, organization, subsequent career, boards, companies,

grants, patents, government appointments, venture investments. Ask: did

the relationship produce something observable later. If the chain stops

after the correspondence, mark it DEAD END. If it continues, document

exactly how. Status: the corpus exists (separate file); the overlay

against the modern map is not yet built.

MAP 16: THE BENEFIT LEDGER. Stop measuring AI success by market

capitalization alone. A social return ledger per major AI deployment:

costs (capital, electricity, water, land, public infrastructure,

environmental, labor displacement, privacy, security, regulatory) against

benefits (productivity, wages, new businesses, medical outcomes, scientific

discovery, reduced costs, new services, time saved, public-sector

efficiency). Then private return vs public return vs externalized cost.

Overlaps the user's existing material. Status: partially covered there.

MAP 17: WHO CAPTURES PRODUCTIVITY GAINS. If AI raises productivity 20

percent, where does it go: higher wages, lower prices, shorter workweeks,

higher profits, executive pay, stock valuations, lower headcount,

shareholder distributions. Get actual numbers. If productivity rises while

wages stagnate, employment falls, prices do not decline, and margins rise,

that is evidence of distributional capture -- stronger than saying "AI is

extracting." Status: not yet built; the sharpest single empirical test in

the proposal.

MAP 18: THE CITIZEN BALANCE SHEET. The average person's AI balance sheet:

benefits (cheaper services, better diagnosis, education, productivity,

accessibility, employment) against costs (data surrendered, surveillance,

job displacement, algorithmic discrimination, energy demand, tax

expenditure, reduced bargaining power, personalized pricing). Is the median

citizen a net beneficiary. Status: not yet built; requires counterfactuals

and is the hardest map to do honestly.

MAP 19: THREE COMPETING HYPOTHESES. Investigate simultaneously: (A)

productive revolution -- gigantic infrastructure investment with real

expected productivity gains, disruption inevitable in transition; (B)

financial bubble -- capital chasing AI on greater-fool expectations, with

supplier financing and guarantees disguising the risk, losses propagating

through AI companies to infrastructure to banks to funds to investors if

revenue disappoints; (C) extractive platform economy -- AI concentrating

data, IP, compute, capital, labor productivity, and pricing power, with

the public supplying infrastructure and research while few companies

capture the upside. All three can contain some truth. Added to the method

in Part 6.

MAP 20: THE TIMELINE. A chronological spine, roughly 1990 to today:

government AI funding, research breakthroughs, Epstein correspondence, AI

companies founded, major investments, semiconductor developments, cloud

infrastructure, government contracts, regulatory changes, data and privacy

legislation, acquisitions, infrastructure financing, energy projects, AI

layoffs, consumer pricing controversies. Look for inflection points:

chronology answers WHEN the system changed, which graphs cannot. Status:

not yet built.

THE DEAD-ENDS REGISTER. A standing record of leads that did not pan out:

a person in an Epstein document with no later connection found; a

government incentive that was ordinary and disclosed. Protects the

investigation from becoming a machine that only collects confirming

evidence. Added to the method in Part 6.

PROPOSED FINAL STRUCTURE: a research graph of linked databases -- PEOPLE,

CAPITAL, COMPANIES, PROJECTS, GOVERNMENT, FINANCE, DATA, ENERGY, PUBLIC

MONEY, AI SYSTEMS, CITIZENS/WORKERS, BENEFITS/COSTS -- every connection

carrying a primary source; plus a separate EPSTEIN HISTORICAL OVERLAY and

a separate UNSUPPORTED/DISPROVEN/DEAD-END database. The questions it would

answer: who supplies the capital, who controls the infrastructure, the

data, the models, the electricity; who receives government assistance; who

bears the financial risk; who captures the productivity gains; who absorbs

the external costs; and ultimately whether economic and political power are

becoming more or less concentrated because of AI.

FEASIBILITY NOTE (added by the investigator, not the blank AI): the maps are not equally buildable. Most documentable now: capital (SEC filings),

government subsidy (DOE, utility commissions, state incentive disclosures),

procurement (federal contract databases), electricity (utility filings).

Partially constrained: ownership (private funds do not disclose

beneficial owners), lobbying (disclosure regimes have gaps), patents (the

public-funding lineage is traceable but laborious). Hardest: the inference

economy (ownership of inferences is legally unsettled), the citizen

balance sheet (requires counterfactuals). Build in that order.

PART 11: NEW DATASETS AND METHOD ADDITIONS (BLANK-AI, OCTOBER 5, 2026)

The blank cross-check AI answered "can you search all of this and find the

answers we need" with: yes, a surprisingly large amount, divided into

three levels -- what SEC filings document well, a network made clearer by

an independent BIS study, and datasets deeper than the current ledger. New

entries and method rules recorded below. All as reported by the blank AI,

not independently verified.

NEW LEDGER ENTRIES (REPORTED)

84. Akamai / Anthropic: Akamai disclosed an Anthropic agreement worth

approximately 11.6B over seven years for dedicated cloud capacity, per SEC

filing. Implication: the reported 518B Anthropic total is not necessarily

the whole universe of its infrastructure commitments.

85. CoreWeave / Meta: a 21B Meta agreement running through 2032, per SEC

filing. This is the customer-revenue side of CoreWeave's debt stack (Part

8, section F).

86. Galaxy Helios financing documents, per SEC: 400 MW additional capacity;

1.63 GW approved electrical capacity; 800 MW leased capacity; CoreWeave as

tenant; approximately 10.4B in minimum contracted lease payments; NVIDIA

owning approximately 11.5 percent of Galaxy and obligated to purchase up

to 6.3B of residual cloud capacity. This is the full chain in one

transaction: AI demand to cloud contract to data-center asset to lease to

financing to NVIDIA equity to NVIDIA residual obligation.

87. Oracle: explicitly planned to raise 45-50B in 2026 to expand OCI

capacity for customers including AMD, Meta, NVIDIA, OpenAI, TikTok, and

xAI, per Oracle investor relations.

88. NVIDIA: AI-cloud commitments of 36B as of July 26, 2026, per NVIDIA's

own filing -- alongside the financing partnerships intended to mobilize

more than 500B of third-party capital.

THE BIS STUDY (AS REPORTED)

Bank for International Settlements researchers independently studied the

phenomenon this investigation has been mapping. Dataset: 1,246 AI firms.

Findings as reported: more than half of AI-company financing from 2021-25

came from other AI companies; nearly half of those AI-to-AI financing

relationships involved firms that already had commercial relationships.

The BIS researchers describe this as potentially "self-referencing

financing." Three mechanisms identified: (1) a company receives financing

AND business from the same counterparty; (2) a supplier finances its

customer so the customer can buy more of its products; (3) a customer

finances a supplier to secure scarce inputs. The BIS is not alleging

fraud; it warns these structures can make genuine underlying demand harder

to determine and can create hidden interconnected exposures. This is

independent confirmation of the circular-financing pattern, and it is more

defensible than any allegation: the structure obscures demand and hides

interconnection.

ANTHROPIC'S OWN WARNING (AS REPORTED)

Anthropic itself warns that Amazon, Google, and Microsoft simultaneously

occupy roles as investors, infrastructure providers, distributors, and

competitors. When the company at the center of the obligations names the

concentration itself, the lead is documented, not inferred.

THE RISK CALCULATION (METHOD)

The question is not how much money is going into AI. It is how much money

is being committed based on expected future AI demand, and who absorbs the

loss if that demand does not materialize. The calculation: contracted

revenue, minus non-cancelable compute obligations, minus debt service,

minus lease obligations, minus capital expenditure, minus guarantees,

minus minimum purchase commitments -- compared against actual AI revenue.

That is how a genuine infrastructure boom is distinguished from an

unstable financing structure.

THE FIELD DISCIPLINE (METHOD, ADOPTED)

Never convert an announced commitment into money spent. Every entry keeps

separate fields: announced, contracted, committed, drawn, spent,

guaranteed, contingent, cancelable, non-cancelable. This extends the

accounting discipline already in Parts 7-9.

THE PROJECT RULE (METHOD, ADOPTED)

"Follow the obligation, not the headline." A 500B announcement is not

necessarily 500B of money. A 500B commitment is not necessarily 500B of

debt. A 500B debt obligation is not necessarily 500B of economic loss. A

guarantee is not a payment. An equity investment is not a grant. A

government incentive is not automatically a subsidy. Preserving these

distinctions is what makes the investigation hard to dismiss.

ENTITY EXPANSION

The blank AI's recommended starting set adds to the ledger's entities: AMD,

Crusoe, Lambda, Applied Digital, Digital Realty, Equinix, Galaxy,

Microsoft/BlackRock infrastructure vehicles, major utilities, DOE, state

economic-development agencies, major universities, and DARPA/NSF/DOE

research programs -- then a people layer (CEOs, founders, directors,

major shareholders, financiers, officials, former officials, lobbyists,

researchers, board members), with the Epstein correspondence as an overlay

only after that map exists.

WHAT CANNOT CURRENTLY BE KNOWN (AS REPORTED)

Beneficial ownership of every private vehicle; exact terms of private

credit facilities; undisclosed side agreements; internal transfer pricing;

private-company cash flows; all government negotiations; confidential IPO

disclosures; exact economics of some guarantees; what percentage of

announced commitments actually gets drawn; whether projected AI demand

materializes. The BIS researchers' warning applies here: many companies

are private, and transactions can combine equity, purchases, long-term

contracts, and guarantees, making the structure difficult to monitor.

Unknowns are recorded as unknowns, not filled in.

PART 12: THE DEMAND-ORIGIN QUESTION (OCTOBER 5, 2026)

A transaction ledger now exists as a separate file

(ai-ledger-transactions.txt): 35 transaction rows with instrument types,

cash/contractual/guarantee fields, do-not-sum overlaps, a projects table,

and a seeded people/control table. It is built to compute one target

metric: the amount of AI infrastructure whose economics depend on other AI

companies.

Two additions from the blank AI's latest pass are recorded here as

findings-in-progress, not conclusions.

First, AI financing increasingly combines five things that used to be

more separate: equity investment, long-term purchase contracts, debt,

supplier financing and guarantees, and physical infrastructure -- with the

same company participating in several categories at once. That is why

headline numbers mislead: a single relationship can appear as investment,

revenue, and contingent exposure simultaneously.

Second, the sharpest unanswered question, now the center of the

investigation: where did the original economic demand come from? If the

answer is millions of independent customers paying for useful AI, that is

one thing. If a significant portion is AI company to investment to

infrastructure commitment to supplier financing to infrastructure expansion

to more purchases from the supplier to higher projected AI revenue, then

the system has a self-reinforcing financing component. That hypothesis is

now to be tested, not assumed -- and the ledger is the instrument for

testing it. The current evidence justifies investigating the question

seriously. It does not establish fraud or deliberate exploitation, and

those two statements stay separate.

PART 13: THE MASSACHUSETTS BRANCH (OCTOBER 5, 2026)

A dedicated Massachusetts ledger now exists as a separate file

(ai-ledger-massachusetts.txt): 21 entries across the executive branch,

federal delegation, state legislature, and regulators, with new status

codes for tax expenditure, infrastructure cost, ratepayer cost, and

projected benefit.

Why Massachusetts matters to this investigation: the Commonwealth is

writing data-center cost-allocation rules BEFORE large load arrives --

EO 658 (confirmed, September 8, 2026) requires >25 MW data centers to

secure community-benefits agreements, procure clean energy for their load

or pay into a Ratepayer Protection Fund, and bars NDAs with permitting

agencies; the DPU is directed to build large-load rate schedules and

clear speculative projects from interconnection queues. Separately,

Warren's federal investigation found Big Tech would not disclose the

utility contracts, rates, and infrastructure costs needed to verify who

pays. The standing question for the branch: does the actual money trail

match the stated policy -- exactly how much of the AI/data-center

infrastructure bill is paid by the operator, and how much lands elsewhere.

Two items need primary pulls: the DPU's reported September 2026

interconnection-queue disclosure order to Eversource, National Grid, and

Unitil (the EO 658 DPU directives are confirmed; the specific disclosure

order is not yet), and DOR records quantifying the paused data-center tax

exemption. The Pressley-tracked $33.4B AES acquisition by BlackRock's

Global Infrastructure Partners connects the capital map (BlackRock is

already a node) to the ratepayer layer.

This branch is distinct from the separate Auditor-anchored Massachusetts

money-trail investigation, which concerns different subject matter.

PART 14: PRIMARY-RECORD UPGRADES (OCTOBER 5, 2026)

The blank AI tested the research brief against primary sources. Three

upgrades matter.

BIS PUBLICATIONS (via the blank AI's pull; the papers themselves not

yet read directly -- treat numbers as reported until read):

- Bulletin 137, "Circular relationships among AI firms," October 1,

  2026: 28.7% of AI firms' investment deals by value (2021-2025)

  involved another AI firm; 55.2% of incoming investment into AI firms

  came from other AI firms; 46.4% of AI-to-AI deal value also involved

  commercial supply-chain relationships between investor and target.

  BIS: these relationships increase opacity and create macroeconomic

  risks.

- Working Paper 1367, "The AI investment race," July 14, 2026: models

  the AI investment boom at roughly 1.5x the socially efficient level

  (baseline), up to about 3x where demand is less elastic; models debt

  and circular equity ties; a single firm's failure can propagate

  through the network.

- Annual Economic Report 2026: circular financing described as

  hyperscalers/chipmakers taking equity positions in AI labs or

  neoclouds while those companies commit purchases back to the

  investors; arrangements can obscure risk and potentially involve the

  same asset pledged multiple times.

COREWEAVE 2025 10-K (via the blank AI's pull):

- $60.7B remaining performance obligations at December 31, 2025.

- About 67% of 2025 revenue came from Microsoft alone.

- OpenAI committed up to approximately $6.5B through May 2031.

- Meta committed approximately $14.2B through 2031 under the 2025

  order; on March 31, 2026 Meta entered a NEW order of approximately

  $21B extending through December 2032 (SEC filing).

- More than 98% of 2025 revenue from committed contracts, generally

  take-or-pay structures; customer prepayments generally 15-25% of

  contract value; $21.6B of debt at December 31, 2025.

- CORRECTION: the "$21B Meta agreement" belongs to 2026, not the 2025

  filing. Date every finding; stale is worse than absent.

- Infrastructure primarily financed through asset-level debt supported

  by take-or-pay customer contracts, plus corporate debt and equity:

  customer commitment -> revenue visibility -> debt capacity ->

  construction -> GPUs -> compute. That chain is testable.

ORACLE $75B (via the blank AI's pull):

- Oracle's FY2026 disclosure: $75B of large-scale AI contracts involve

  customer prepayment for GPUs OR customers purchasing and supplying

  GPUs themselves; Oracle says this reduces the capital it must raise.

- Do NOT assume the $75B is cash prepayments. Cash prepayment and

  customer-supplied equipment are different instruments sharing one

  headline. The split is an open question.

METHOD REFINEMENTS ADOPTED:

- Four ledgers before any netting: (1) equity, (2) debt/credit,

  (3) commercial commitments, (4) guarantees/collateral/support. Net

  positions only after separation -- otherwise a $21B commercial

  commitment gets added to a $21B loan as if they were equivalent.

- New missions: A7 (follow the same dollar twice -- trace economic

  circulation, not circular accounting accusations); A8 (revenue

  dependency -- who pays whom, % AI-native vs hyperscaler vs

  enterprise, prepaid/take-or-pay/cancelable splits, concentration);

  A9 (follow the MW -- dollars to contracts to GPUs to MW to data

  centers to electricity to ratepayers; where the financial and

  Massachusetts ledgers intersect).

- Hypothesis ladder H1-H6 maps onto the standing three hypotheses:

  H1 (enormous investment, enormous expected demand) ~= productive

  revolution; H2-H4 (debt financing, reciprocal commitments,

  forward demand overstated) ~= bubble mechanics -- BIS is already

  investigating close variants; H5-H6 (risks transferred outward,

  fragility if revenue disappoints) ~= extraction/fragility -- H5 is

  where the Massachusetts evidence matters. The ladder is adopted as

  the working version because it is testable step by step.

PART 15: ROUND-3 UPGRADES (OCTOBER 5, 2026)

The blank AI adopted the research brief as the master standard and

tested its premises against primary sources. Upgrades recorded here.

BIS (primary publications now pulled, not via press):

- Working Paper 1367, "The AI investment race," by Phurichai

  Rungcharoenkitkul (July 14, 2026): 1.5x socially efficient

  investment (baseline model), up to ~3x under less-elastic demand;

  models failure propagation. CAUTION: 1.5x is model output under

  stated assumptions, not a finding that 50% of spending is

  fraudulent or unnecessary.

- Bulletin 137, "Circular relationships among AI firms" (October 1,

  2026), by Frost, Kansal, Rishabh, Shreeti, Zhang: 28.7% of AI

  firms' investment deals by value (2021-2025) involved another AI

  firm; 55.2% of incoming AI-firm investment came from other AI

  firms; 46.4% of AI-to-AI deal value also involved commercial

  supply-chain relationships. BIS PROVIDES AN EXCEL FILE WITH THE

  UNDERLYING GRAPH DATA -- top priority to pull and analyze

  directly.

- Annual Economic Report 2026: circular financing as equity-for-

  purchase-commitment swaps; arrangements can obscure risk and

  potentially involve the same asset pledged multiple times

  (the rehypothecation question -- needs precision, not

  characterization).

COREWEAVE TIMELINE (SEC filings):

- September 2025: Meta committed ~$14.2B through December 2031,

  with an option to expand materially through 2032.

- April 2026: expanded to ~$21B through December 2032.

- The "$21B Meta agreement" is a 2026 number; the 2025 number was

  $14.2B. Transaction evolution belongs in the ledger.

- Reframed question: who finances CoreWeave's capacity while

  Microsoft (67% of 2025 revenue) and Meta commit to buying it,

  and what happens to the economics if those customers reduce

  demand?

ORACLE DECOMPOSITION (three numbers, not one):

- $75B: large AI contracts with prepaid OR customer-supplied

  hardware (mixed).

- $4.6B: customer prepayments containing a significant financing

  component in FY2026 (10-K).

- $43B debt + $5B equity raised in FY2026.

Never conflate the three.

ANTHROPIC IPO (method note):

- A confidential IPO filing does NOT appear in EDGAR as an

  ordinary S-1. Distinguish: public S-1, confidential submission,

  amended confidential submission, Form D/private filings, press

  reports of an alleged submission.

- Caught: "Augurey Ventures QP Fund, LLC -- Series Anthropic SP-1"

  (Sept 28, 2026) is a Form D/A for an investment vehicle, NOT

  Anthropic's IPO. Do not mistake it.

MASSACHUSETTS LOAD TAXONOMY (adopted):

- Project stages: concept -> application -> queue -> study ->

  agreement -> construction -> energized.

- MW measures: requested -> studied -> contracted -> actual.

- Never call a proposed load a committed load.

NEW MISSIONS ADOPTED:

- B8: follow the Massachusetts dollar all the way to the ratepayer

  (developer -> interconnection -> infrastructure -> who pays

  initially -> rate mechanism -> who ultimately bears cost ->

  tax/subsidy -> actual employment/investment -> actual

  consumption). Distinguishes "MA spends to attract AI" from

  "residents bear identifiable AI costs."

- C2 strengthened: all threshold effects (1/10/20/25/50/100 MW) --

  does crossing change legal/regulatory/tariff/reporting/tax/

  interconnection treatment? Then look for projects clustered at

  boundaries (e.g., 40 projects at 24-25 MW after a 25 MW rule).

- C5: trace circular economic value -- for each major relationship,

  test whether the loop exists (A invests in B -> B buys from A ->

  A finances B -> B commits purchases to A -> commitments support

  A's valuation/debt -> A finances more B capacity), then ask how

  much of the loop is genuine external demand vs circulating

  capital. Empirical question, not assumption.

PRIORITY ORDER (blank AI's, adopted):

A2 (BIS data) -> A3 (CoreWeave map) -> A5 (Oracle) -> Q1

(Microsoft/CoreWeave) -> Q3 (BIS model) -> Q5 (CoreWeave debt vs

revenue) -> B1 (DPU filings) -> B2 (EOED certs) -> B5 (S.3178) ->

B4 (OpenAI spend) -> B6 (OCPF) -> C1 (stranded costs) -> C2

(thresholds) -> C3 (states) -> C4 (Amazon map) -> C5 (loops).

STANDING NOTE: the blank AI states the investigation no longer

depends on Epstein -- the contemporary structure is investigable through SEC filings, BIS research, utility proceedings, procurement

records, legislative documents, and campaign-finance records alone.

Recorded because it converges with this file's standing position

(Part 1: Epstein as historical window, not organizing principle).

PART 16: EXECUTION PHASE (OCTOBER 5, 2026)

The blank AI adopted the brief as the master research standard and

moved to execution. Recorded here: the schema, the phased plan, the

analytical frames, and new leads.

MASTER TABLE SCHEMA (15 fields): date | source entity | recipient |

instrument | face amount | cash actually transferred | future

commitment | term | cancellation | take-or-pay | collateral |

revenue dependency | physical dependency | DOWNSIDE BEARER |

primary source | confidence. The downside-bearer column is the

public-interest core of the table.

FOUR-WAY DISTINCTION ON "CIRCULAR FINANCING" (adopted -- analytical

hygiene): (1) normal vertical integration -- invest in suppliers for

capacity; (2) strategic financing -- fund a customer/supplier to

secure business; (3) circular financial exposure -- balance sheets

and contracts interdependent; (4) artificial demand / financial

engineering -- transactions making demand or strength appear greater

than independent end-user economics justify. 1-3 can exist without

fraud. 4 requires much stronger evidence. The job is to see where

the evidence stops.

ECONOMIC CUSHION FRAME (adopted -- the best analytical frame yet):

per major node, committed future revenue minus required

infrastructure spending minus debt service minus operating costs =

economic cushion. Then scenarios: Microsoft -25%, OpenAI -25%,

Meta delayed, GPU residual values fall, electricity costs rise,

construction delayed, utilization below contract. Then: who absorbs

the loss -- investors, lenders, AI companies, cloud customers,

equipment makers, utilities, taxpayers, ratepayers? This replaces

"is the system good or bad" with a stress test.

A6 (new mission): follow the dollar twice -- flow-of-funds test per

major transaction (A invests in B -> B buys from A -> A finances B

-> B commits purchases to A -> commitments support A's

valuation/debt -> A finances more B capacity). Documents the loop

if it exists; does not assume it.

PHASED PLAN (adopted): Phase 1 -- financial network (BIS data,

CoreWeave, Oracle, Microsoft/CoreWeave, BIS model, CoreWeave debt

vs revenue). Phase 2 -- physical demand (ISO-NE CELT, DPU filings,

EOED certs, stranded-cost precedents, threshold engineering,

state comparison). Phase 3 -- political/public-money layer

(procurement rubric, OpenAI spend, S.3178, OCPF, H.5175/fund).

Politics LAST -- otherwise confirmation bias leads.

NEW LEADS:

- Oracle CEO: $67B of AI infrastructure contracts signed "in that

  quarter," bringing the prepaid/BYOH total to $75B (Oracle Blogs).

  Which quarter, which counterparties -- open (R4-Q4).

- ISO-NE 2026 CELT Report exists with downloadable XLSX

  (2026-2035 forecast) -- pullable now (R4-Q6).

- Meta/CoreWeave date refinement: April 9, 2026 announcement of

  the ~$21B agreement (not just "April 2026").

- MA load taxonomy adopted in the ledger: project stages and MW

  measures kept separate (Part 15); B8 adopted (follow the MA

  dollar to the ratepayer).

PART 17: BIS EXCEL PARSED (OCTOBER 5, 2026)

The blank AI located but could not parse the BIS Excel file. This

investigation pulled and parsed it directly: "Circular relationships

among AI firms," BIS Bulletin 137 (Frost, Kansal, Rishabh, Shreeti,

Zhang). File saved at

~/workspace/research_notes/bis-bulletin-137/circular-relationships-among-ai-firms-data.xlsx.

Sources per the file: Compustat; PitchBook (Morningstar);

Rishabh and Shreeti (2026).

STRUCTURE: 5 sheets. Readme (citation). Graph 1: "AI giants have grown

both organically and through acquisitions" -- 7 "AI giant" rows with

an organic-growth series plus 374 named acquisition rows, in USD

millions, across periods 2000-2004 through 2025. Graph 2: no data

(typology only). Graph 3: the headline percentages (28.7 / 55.2 /

16.1 / 46.4). Graph 4: circular deals by supply-chain layer --

136 deals by count.

GRAPH 1 FINDINGS:

- Largest named acquisitions in the dataset: Microsoft->Activision

  Blizzard $75.4B; Broadcom->VMware $69B; Microsoft->Aligned Data

  Centers $40B; Nvidia->Aligned Data Centers $40B; Alphabet->Wiz

  $32B; Microsoft->LinkedIn $27.1B.

- FLAG: the two $40B Aligned Data Centers rows (Microsoft and

  Nvidia) are surprising and NEED VERIFICATION against primary

  sources before use. They may reflect announced commitments rather

  than closed deals, or a data quirk. Do not cite them as fact

  until checked.

- No AI-lab targets (OpenAI, Anthropic, xAI, etc.) appear among

  named acquisition rows: the AI-to-AI deals are minority

  investments and partnerships, not acquisitions, so they do not

  appear in this sheet.

- The 7 organic-growth rows (Apple, Microsoft, Nvidia, Alphabet,

  Amazon, Meta, Broadcom) carry very large values; the file does

  not define the organic measure in the sheet itself -- read the

  bulletin text before interpreting or citing.

GRAPH 4 FINDINGS (circular deals by layer, counts):

- By source layer: Infrastructure 54, Compute 45, Applications 25,

  Data tools 10, Models 2.

- By target layer: Data tools 40, Models 35, Applications 29,

  Infrastructure 17, Compute 15.

- Largest cells: Infrastructure->Models 21, Infrastructure->Data

  tools 16, Infrastructure->Applications 15, Compute->Infrastructure

  13, Compute->Data tools 11, Compute->Compute 11.

- Reading: the circularity flows downstream -- infrastructure and

  compute firms invest into models, data tools, and applications

  with which they also have commercial relationships.

WHAT THE FILE CANNOT ANSWER:

- It contains NO firm-level rows for the 972 investment

  relationships. R3-Q1's concentration test (mega-deals vs

  broad-based) cannot be run from this file. The underlying

  deal-level data sits with PitchBook or the authors.

- Graph 2's typology has no data behind it in the file.

BOTTOM LINE: the Excel confirms the bulletin's structure and adds

the layer matrix, but the firm-level dataset behind the 55.2% is

not public in this file. The $40B Aligned Data Centers rows are

either a significant find or a data artifact -- verification

required either way.

PART 18: PROOF-OF-CONCEPT EXECUTION -- COREWEAVE (OCTOBER 5, 2026)

Pulled primary filings directly from SEC EDGAR (CIK 0001769628):

Q2 2026 10-Q (filed 2026-08-12) and the April 9, 2026 8-K (Meta

order form). Files saved at

~/workspace/research_notes/coreweave-filings-20261005/.

R5-Q2 RESOLVED -- DEBT RECONCILIATION:

$21.6B | total indebtedness | CoreWeave | Dec 31, 2025 | 10-K

$35.6B | total indebtedness | CoreWeave | Jun 30, 2026 | 10-Q

Both correct. The $14B gap is REAL new borrowing in H1 2026, not

a definitional difference. Tranche drivers: 2031 9.75% Senior

Notes $2,750M (new, Oct 2031, 10% eff.); 2032 9.625% Senior Notes

$1,250M (new, Jul 2032, 10% eff.); 2032 EUR Senior Notes $2,279M

(new, Jul 2032, 9% eff.); 2032 Convertible Senior Notes $4,000M

(new, Oct 2032, 2% coupon); DDTL 5.0 $1,101M (new, Nov 2031, 9%);

DDTL 4.0 $2,837M (new, non-recourse, Mar 2032, 7%); DDTL 3.0 drew

$340M -> $2,215M. Offsets: DDTL 2.0 paid down $5,037M ->

$3,190M; revolver $1,000M -> $0 drawn; convertible promissory

$168M matured.

Full stack at Jun 30, 2026: recourse $31,405M net; non-recourse

$3,663M net (DDTL 4.0 + OEM); current portion $7,513M combined

($6,235M recourse + $1,278M non-recourse) due within 12 months.

Undrawn availability $10.0B. Effective rates 9-15% on DDTLs, 10%

on senior notes, 2% on converts, 7% on DDTL 4.0 and revolver,

12% Magnetar loan. This is expensive debt -- a market signal in

itself (feeds R6-Q1).

DURATION LADDER (R6-Q2, first pass):

Debt maturities: Mar 2028 (DDTL 1.0, $1,300M, 15%) -> Jun 2030

(2030 notes) -> Aug 2030 (DDTL 2.0/3.0) -> Feb/Oct 2031 (notes,

DDTL 2.1/5.0) -> 2032 (notes, converts, DDTL 4.0).

Contract durations: Meta through Dec 20, 2032 (plus option to

Apr 2032); OpenAI through May 31, 2031.

Reading: contracted durations EXTEND BEYOND most debt

maturities -- the favorable direction for refinancing risk. The

risk is rate, not duration: refinancing 9-15% paper in 2028-2031

depends on credit conditions then, not on contract coverage.

META $21B -- EXACT TERMS (8-K, Apr 9, 2026):

$21B | customer commitment (order form, NOT equity, NOT cash

received) | Meta -> CoreWeave | Mar 31, 2026 (announced Apr 9) |

8-K Item 8.01. Order form under MSA dated Dec 10, 2023. Meta

"initially committed to pay" ~$21B, INCLUSIVE of (i) new capacity

through Dec 20, 2032 and (ii) exercise of an existing option

through Apr 10, 2032. Subject to delivery and availability-of-

service requirements -- CoreWeave must deliver to earn it.

Termination for cause only, either party. "Initially" leaves room

for change. The $21B bundles old option + new order; it is not

$21B of incremental new commitment on top of everything prior.

RPO CORRECTION:

$103.7B | remaining performance obligations | CoreWeave | Jun 30,

2026 | 10-Q (SEC filing). This REPLACES the $104.2B presentation

figure -- the filing number is RPO as defined (transaction price

allocated to undelivered performance obligations, net of

estimated variable consideration including availability credits,

delivery-delay risk, and resalable capacity). Variable

consideration means it is NOT guaranteed revenue.

CUSTOMER CONCENTRATION TREND (10-Q, anonymized as Customer A/B):

Customer A: 36% of Q2 2026 revenue (was 71% Q2 2025); 40% H1

2026 (was 72% H1 2025). Customer B: 26% Q2 2026, 23% H1 2026

(was under 10% in 2025, unlisted).

CORRECTION (Oct 5, adversarial review): Customer A = Microsoft

is NOT established for 2026. CoreWeave's filing warns the A/B/C/D

labels may represent different customers across periods. The

10-K named Microsoft at ~67% of FY2025 revenue (A, historical).

The 71%->36% decline is the anonymized label's share (A on the

numbers); attributing it to Microsoft is inference (D) or

unresolved identity (U). Do not write "Microsoft's share fell"

without the caveat. Customer B is unidentified in the filing;

timing is consistent with Meta's March 2026 order but the filing

does not say so (D).

Reading (narrowed): the top-customer concentration is falling

fast as the book diversifies -- genuine improvement in the risk

picture, whoever Customer A is.

PROOF-OF-CONCEPT DEMAND CHAIN: CoreWeave -> Microsoft -> ?

Node 1: CoreWeave -> Microsoft (Customer A). 36% of Q2 2026

revenue | commercial revenue | financial-statement evidence |

classification: INTERMEDIATE (hyperscaler/aggregator).

Node 2: Microsoft -> downstream. CoreWeave's filings do not

identify what Microsoft does with the capacity. Microsoft's

filings disclose aggregated Azure/AI revenue, not CoreWeave-

sourced capacity by end customer. Classification: UNRESOLVED.

The chain terminates at node 2 under the locked rule -- unknown

is not promoted to independent.

Complication: Microsoft is simultaneously CoreWeave's largest

customer, OpenAI's largest backer, and OpenAI is itself a

CoreWeave customer ($6.5B commitment). The triangle

Microsoft->CoreWeave / Microsoft->OpenAI / OpenAI->CoreWeave

means "Microsoft demand" cannot be classified without knowing

which downstream dollars are OpenAI's vs enterprise Azure

customers'. This is exactly the AI-loop vs independent ambiguity

the recursion rule was built for.

MEASURED RESULT: the first end-to-end trace terminates at

UNRESOLVED after one hop. That is not a failure; under the

locked methodology it feeds the unresolved ratio. It also tells

us the binding constraint on the whole demand layer: hyperscaler

10-Ks do not disclose per-supplier capacity allocation by end

customer, so every chain through a hyperscaler will terminate at

unresolved unless the hyperscaler discloses it (R4-Q7, still

open).

PART 19: R7-Q1 -- OPENAI $6.5B MSA TERMS (OCTOBER 5, 2026)

From SEC EDGAR, primary filings.

$6.5B | customer commitment (order form, ceiling) | OpenAI ->

CoreWeave | Sep 23, 2025 (8-K filed Sep 25, 2025) | 8-K Item

1.01. Confidence: A (terms), C (economics below).

Terms, verbatim in substance:

- Order Form entered September 23, 2025, under MSA dated May 8,

  2025, between CoreWeave, Inc. and OpenAI OpCo, LLC.

- OpenAI "committed to pay the Company up to approximately $6.5

  billion through May 31, 2031 under the Order Form."

- Subject to termination provisions AND "satisfaction of delivery

  and availability of service requirements" -- CoreWeave must

  deliver working capacity to earn it.

- MSA remains until all outstanding orders expire/terminate or

  the MSA is terminated per its terms.

- Either party may terminate for cause.

- Full MSA text filed as Exhibit 10.1 -- WITH PORTIONS REDACTED

  under Reg S-K 601(b)(10).

Material findings:

1. The May 8, 2025 MSA was NOT separately 8-K'd. The May 2025

   8-Ks are earnings (Item 2.02) and other matters. The MSA only

   became disclosable when the September order form attached a

   dollar figure. Implication: the original MSA carried no

   disclosed dollar commitment; the $6.5B arrived with the

   September order form.

2. "Up to approximately $6.5 billion" -- a CEILING, not a floor,

   not cash received.

3. Filed as Item 1.01 (material definitive agreement) -- a

   stronger materiality treatment than the Meta $21B order form,

   which went out as Item 8.01 (other events) + Item 7.01 (Reg

   FD). The company judged the OpenAI MSA itself material; the

   Meta disclosure was framed as an announcement.

4. TAKE-OR-PAY NOT ESTABLISHED. The 8-K describes "reserved

   capacity orders" with payment "subject to delivery and

   availability." Whether the commitment is take-or-pay, and on

   what conditions payment is excused, sits in the REDACTED

   portions of Exhibit 10.1. Every prior characterization of

   CoreWeave's contracts as "generally take-or-pay" (including

   in this investigation's earlier parts) rests on company

   description, not on visible contract language. Flag as

   analytical inference (confidence D), not filing fact.

5. Comparison, Meta vs OpenAI: Meta $21B = order form under Dec

   2023 MSA, bundles prior option + new capacity, "initially

   committed," Item 8.01. OpenAI $6.5B = order form under May

   2025 MSA, single order, "up to," Item 1.01 with MSA exhibit

   filed (redacted). Different vintages, different filing

   postures, same structural shape: pay-if-we-deliver,

   terminable for cause.

R7-Q1 status: PARTIALLY RESOLVED. Commercial terms extracted

(A). Payment unconditionality / take-or-pay mechanics: REDACTED

in exhibits (U). Do not assert take-or-pay as filing fact.

PART 20: R7-Q3 -- ORACLE $75B DECOMPOSITION (OCTOBER 5, 2026)

From SEC EDGAR: FY2026 10-K (filed Jun 22, 2026), Q1 FY2027

10-Q (filed Sep 11, 2026), Q4 FY2026 earnings press release

(8-K exhibit, Jun 2026), Q1 FY2027 earnings press release (8-K

exhibit, Sep 2026). Files at

~/workspace/research_notes/oracle-filings-20261005/.

THE $75B -- EXACT MEANING PINNED:

$75B | prepaid cash + customer-supplied hardware PORTIONS of

large AI contracts | customers -> Oracle | Q4 FY2026 (June

2026) | earnings press release, 8-K exhibit (Tier 2).

Confidence: B.

Verbatim substance: "Most of the RPO increase in both Q3 and Q4

were large scale AI contracts where the customer prepaid Oracle

for the purchase of the GPUs, or the customer bought and

supplied the GPUs to Oracle. The prepaid and customer supplied

hardware portions of our large AI contracts now total $75

billion."

Critical nuance: the $75B is the customer-financed PORTION

carved out of the contracts -- NOT the contracts' total value.

The contracts themselves are larger (they drove RPO to $638B).

WHAT IS NOT DISCLOSED:

- The split between prepaid cash and customer-supplied hardware:

  NOT disclosed anywhere found.

- Customer names: NOT disclosed.

- Per-customer amounts: NOT disclosed.

R7-Q3 status: PARTIALLY RESOLVED. Meaning pinned (B);

decomposition undisclosed (U).

ORACLE'S OWN FRAMING (demand layer relevance):

"This substantially reduces the amount of capital Oracle must

raise to build out our AI datacenters." (Jun 2026)

"Based on the structuring of those new contracts, the Company

confirms there is no incremental impact on its plans to raise

capital." (Sep 2026)

Oracle states outright: customer money substitutes for Oracle's

own capital raising. The demand layer and the capital layer

collapse into each other -- the customer IS the financier. This

is the AI-loop pattern in Oracle's own words.

ACCELERATION -- CORRECTED (Oct 5, adversarial review):

$4.6B | customer prepayments with significant financing

component | customers -> Oracle | FY2026 (full year) | 10-K

(Tier 1). Confidence: A. (Zero in FY2025 and FY2024.)

$11.363B | same | customers -> Oracle | Q1 FY2027 (one quarter)

| earnings press release (Tier 2). Confidence: A on the event.

The filing does NOT say the $11.4B is AI-related, and one

quarter does not establish a run-rate. "Accelerating trend":

DOWNGRADED to D/U. Treat as a lead (one large quarterly event),

not a trend line.

RPO: $638B | May 31, 2026 (10-K, Tier 1) -> $664B | Aug 31,

2026 (press release, Tier 2). +$30B AI cloud contracts in Q1

FY2027.

FY2026 CAPITAL RAISED: $43B debt + $5B equity (press release,

Tier 2). Free cash flow -$23.7B despite $32.0B operating cash

flow -- the capex burn.

OPERATIONS: 300,000+ GPUs delivered to AI cloud customers since

end of Q4 FY2026; "almost triple the capacity delivered in Q4

FY26." $20B common stock sold via ATM in Q1 FY2027.

THE $67B: NOT found in either filed earnings release. Remains

attributed to CEO remarks / Oracle blog (Tier 2/3, unverified

by this investigation). Do not cite as established. The filed

Q1 FY2027 signings figure is $30B -- do not conflate.

PART 21: R7-Q5 -- ANTHROPIC COUNTERPARTY CHECK (OCTOBER 5, 2026)

Question: do Google, Amazon, or Microsoft disclose the

Reuters-reported Anthropic commitments ($111.1B / $110B / $31.4B) in primary filings? Pulled from SEC EDGAR. Files at

~/workspace/research_notes/anthropic-counterparties-20261005/.

GOOGLE/ALPHABET:

- FY2025 10-K (filed Feb 5, 2026): 0 mentions of Anthropic.

- Q2 2026 10-Q (filed Jul 23, 2026): 0 mentions of Anthropic.

- Total purchase commitments disclosed: $149.1B at Dec 31,

  2025 ("energy take-or-pay, licenses, technical infrastructure

  and inventory") -- no Anthropic breakout.

$111.1B | commitment | Google -> Anthropic | 2026 | Reuters

(Tier 3). Confidence: C. Counterparty corroboration: NONE.

AMAZON:

- FY2025 10-K (filed Feb 6, 2026): 12 mentions of Anthropic --

  ALL on the investment side, NONE on a $110B commitment.

- Disclosed: $5.3B convertible notes (Q3 2023-Q4 2024); $1.3B

  new note Q2 2025; $1.4B new note Q4 2025; portion converted to

  nonvoting preferred in Q3 2025 (~$2.3B gain reclassified).

- Dec 31, 2025 balance sheet: nonvoting preferred ~$14.8B;

  convertible notes fair value ~$45.8B. $15.2B of 2025 net income

  largely unrealized gains on the Anthropic position.

- Total commitments disclosed: $439.7B across all categories --

  no Anthropic-specific commitment line.

$110B | commitment | Amazon -> Anthropic | 2026 | Reuters

(Tier 3). Confidence: C. Investment side corroborated (A);

commitment figure: NO counterparty corroboration.

MICROSOFT:

- FY2026 10-K (filed ~Jul 2026): 0 mentions of Anthropic.

$31.4B | commitment | Microsoft -> Anthropic | 2026 | Reuters

(Tier 3). Confidence: C. Counterparty corroboration: NONE.

DEMAND CHAIN Anthropic -> Google:

Node 1 (Anthropic -> Google, $111.1B): observed transaction =

Reuters report only; Google's filings silent. The edge itself

is Tier 3. Under the recursion rule the chain cannot proceed on

primary evidence. Classification: UNRESOLVED at node 1.

This is a weaker evidentiary position than CoreWeave ->

Microsoft (node 1 was filing-grade there).

MEASURED RESULT (CORRECTED Oct 5, 2026 -- adversarial review):

The original "zero dollars corroborated" claim was too absolute.

Amazon's Q2 2026 10-Q (filed Jul 31, 2026, verified from EDGAR)

states: "In Q2 2026, AWS and Anthropic announced an expansion

of the strategic collaboration and existing multi-year

commitment by more than $100.0 billion over 10.0 years, which

includes contractual obligations related to the performance of

AWS chips."

Corrected grades:

- >$100B / 10 years, Anthropic-AWS, incl. AWS-chip obligations:

  A (Amazon 10-Q, Tier 1, counterparty-disclosed).

- Exactly $110B: C (Reuters reporting of confidential filing).

- Anthropic -> Google $111.1B: C, no counterparty corroboration.

- Anthropic -> Microsoft $31.4B: C, no counterparty corroboration.

- $518B aggregate: C.

The AWS-chips detail matters for demand-chain analysis: this is

not generic cloud capacity but chip-performance obligations.

The exact stack remains uncorroborated, but "100%

Reuters-dependent" is no longer defensible for the Amazon piece.

R7-Q5 status: PARTIALLY RESOLVED. Counterparty check complete.

Amazon piece corroborated (A); Google/Microsoft pieces remain

Tier 3 (C). Downstream trace still blocked at unresolved.

PART 22: ROUND 7 COMPLETION -- OEM FINANCING + MAGNETAR (OCTOBER 5, 2026)

From the blank AI's filing-level execution, primary filings

checked. Treated as blank-AI-reported pending my verification

unless noted.

R7-Q2 -- $4.220B OEM/SOFTWARE FINANCING:

$4.220B | OEM + software-license financing, recourse |

financing to CoreWeave | Dec 2026-Jul 2030 maturities | 10-Q

(Tier 1). Confidence: A on amount/terms.

$882M | same, non-recourse (subsidiaries) | Aug 2026-Aug 2028 |

10-Q (Tier 1). Confidence: A.

$347M | software-license component within $5.1B total

equipment/software balance | 10-Q (Tier 1).

Rates: 11% recourse, 9% non-recourse. Secured by the financed

equipment. OEM terms generally 1-3 years; software under 5 years

(10-K description).

(Restored 2026-10-06: a Blogger HTML-parsing fault swallowed the

passage below in the published post -- the original phrasing used

a less-than sign before "5 years," which Blogger read as an HTML

tag opening, eating everything up to the ">" in "MagAI Ventures

->". Rewritten as "under 5 years" and the missing text restored

from the investigation file.)

KEY NEGATIVE FINDING: the filing deliberately names no

counterparties -- "original equipment manufacturers" and "a

software license vendor." Dell, NVIDIA, Supermicro, HPE not

named. A January 2026 +$1.5B OEM addition (3-year terms) is

likewise anonymous.

Classification: supplier/vendor financing, not bank debt.

Next: subsidiary security agreements / UCC filings, not more

earnings material.

R7-Q2 status: PARTIALLY RESOLVED (amount/terms A; counterparty

identity U).

R7-Q4 -- MAGNETAR $189M -- STRUCTURALLY RESOLVED:

This was NOT originally a loan.

$230M | refundable capacity prepayment (deposit) | MagAI

Ventures -> CoreWeave | Aug 2024 agreement; 4-year initial term

+ 2-year extension option | 10-K (Tier 1). Refundable;

consumable by MagAI portfolio companies for cloud services.

Feb 2025: amended to add termination-for-convenience (both

sides). CoreWeave reclassified the deposit as in-substance debt

under ASC 470 (sale of future revenue).

$100M | partial settlement paid | CoreWeave -> MagAI | Jun 2026

| 10-Q (Tier 1).

$189M | remaining: unused refundable deposit + accrued

redemption premiums | Jan 2029 | 10-Q (Tier 1).

The 12% is CONTRACTUAL REDEMPTION ECONOMICS (multiplier on the

returned deposit if terminated) -- not a stated loan coupon,

not a market-implied credit spread. The filing does not explain

the 12% by risk; do not label it a credit spread.

No separate GPU collateral package identified for this

instrument; protection sits in the refundable/redemption

mechanism.

Separately: Magnetar-managed/advised funds held $106M of DDTL

2.0 at Dec 31, 2024 -- Magnetar appears twice in the capital

stack (lender + capacity prepayer).

R7-Q4 status: RESOLVED as to structure (A); pricing rationale U.

R7-Q5 -- CONFIRMED ASYMMETRY (see Part 21 correction):

Amazon 10-Q: >$100B/10yr AWS-Anthropic expansion, incl. AWS-chip

obligations (A). Alphabet 10-Q: $811B aggregate commitments, no

Anthropic breakout (C on the $111.1B). Blank AI additionally

reports: $10B Anthropic preferred investment by Amazon in Q2

2026 (see row T-004 -- reconciled against Amazon's Q2 2026

10-Q, 2026-10-06; the "unverified" tag below applied before

that reconciliation); separate financing facility up to $20B,

reduced to $15B after the equity investment, conditioned on

compute-delivery milestones. If verified: Amazon is

simultaneously infrastructure provider, equity investor, AND

financier -- keep the instruments separate. The Anthropic->AWS

edge remains AI-loop/intermediate under the recursion rule;

Amazon's financing does not make the commercial commitment

fictitious, and the commitment does not establish external end

demand.

PART 23: THE INSTITUTIONAL DEPLOYMENT LAYER (OCTOBER 5, 2026)

Ricky's question: Massachusetts has AI in hospitals, the state

is deploying it, police use it -- but not the way he does. What

does that tell us?

This is the missing wing: the investigation has mapped who

FINANCES AI infrastructure. It has not mapped who DEPLOYS AI

against people, with whose money, under what oversight.

CONFIRMED MASSACHUSETTS STATE DEPLOYMENTS (Tier 2/3, Oct 2026):

- Feb 2026: Healey administration deployed a ChatGPT-powered AI

  assistant via competitive-procurement contract with OpenAI to

  ~40,000 executive-branch employees, phased from EOTSS outward.

  First state with enterprise-wide deployment. Walled-off

  environment; use optional; human oversight "essential."

  Contract value and terms: NOT YET PULLED (public record).

- DESE: AI task force, K-12 guidance, literacy modules,

  multi-year roadmap; implementation 2025-26 school year.

- Pioneer Institute brief (Apr 2026), "Massachusetts State

  Government and Artificial Intelligence": 15 reported use cases

  across secretariats -- GrantWell (municipal federal funding),

  IT Ops Genie (HHS docs), One L (agency attorney contract

  review), HEKA highway engineer chatbot (MassDOT), BEACON

  (business grants), ABE procurement chatbot, MassHealth Helper

  (call-center eligibility policy), EOTSS HR chatbot, MassDOT

  RFI summarizer, environmental grants navigator, unemployment

  call-center assistant, education complaint report generator,

  Child Support Payment Predicter (A&F).

- Sep 2026: Gaming Commission ordered evaluation of AI/ML in

  sports betting after NYT's DraftKings investigation (bettor

  "elasticity" targeting).

CONFIRMED LAW ENFORCEMENT (Pioneer Valley):

- Springfield PD: ShotSpotter since 2008 (early adopter);

  body-worn cameras (policy bars facial recognition); councilors

  have pushed facial-recognition bans; PD says no plans to

  deploy it.

- Chicopee PD (Feb 2026): building "AI-integrated Real Time

  Information Center" -- regional; ingests ShotSpotter from

  Holyoke and Springfield; 500+ cameras including private

  business/resident cameras volunteered in; school camera

  integration modeled on Springfield's MoU. Official line:

  "everything is logged," "no expectation of privacy on the

  street."

ANALYTICAL FRAME -- THE ASYMMETRY THESIS:

Every institutional deployment points AI AT people, not FOR

people. The state gives 40,000 employees an assistant to make

government faster. The Child Support Payment Predicter predicts

payers. Cameras and gunshot sensors surveil residents.

DraftKings' AI identified who was likeliest to lose money.

In no deployment found does the citizen, patient, defendant, or

resident get an AI advocate of their own. Ricky's use --

AI as the prepared second mind against the system -- has no

institutional counterpart. The technology is the same; the

direction of power is opposite.

DEMAND-LAYER IMPLICATION:

Institutional procurement is taxpayer-funded AI demand. The

state's OpenAI contract, hospital AI spending, police tech

budgets -- these are public dollars flowing to AI companies.

Some "independent demand" is governments buying tools to manage

populations. This complicates the demand-origin question: the

dollar enters from the public purse and exits as institutional

leverage over the public.

PROCUREMENT OPACITY:

15 state use cases are named; vendors, contract values, training

data, and audit mechanisms are not in the public summary. The

Child Support Payment Predicter -- an algorithmic prediction

about child-support payers, a vulnerable population -- is

exactly the kind of deployment where the investigated pattern

lives: a person enters a system, the system scores them, the

score compounds. Vendor, model, recourse: unknown.

OPEN GAPS:

1. MA hospital AI specifics (MGB, Beth Israel Lahey, UMass

   Memorial): clinical decision support, ambient scribes,

   sepsis/risk models -- who bears the risk when wrong.

2. State OpenAI contract: value, terms, data provisions (public

   record, not yet pulled).

3. MassHealth: beyond the call-center chatbot -- any algorithmic

   eligibility, fraud, or risk scoring?

4. Child Support Payment Predicter: vendor, model, what a

   prediction triggers, recourse for the scored.

5. Springfield PD current tech stack + contract values

   (ShotSpotter, RTIC participation, camera network costs).

6. Federal law enforcement AI use in/around MA (FBI, DHS) --

   nothing gathered yet.

PART 24: STATE OPENAI CONTRACT + CHILD SUPPORT PREDICTOR (OCTOBER 5, 2026)

Ricky picked this thread. Findings from primary/secondary

sources; gaps marked.

THE OPENAI CONTRACT (Tier 2/3):

- 3-year contract, usage-based, tiered: $13/month/worker

  falling to $9/month as adoption expands (news-usa.today;

  corroborated by wheninyourstate.com).

- Procurement: Request for Quotes (RFQ), not full competitive

  bid; OpenAI selected (Pioneer brief). Google noted as the

  usual public-sector competitor.

- Up to 30,000 ChatGPT Enterprise licenses; "advanced units"

  may be added for certain use cases.

- Boston Consulting Group + Slalom Consulting: implementation,

  use-case development, training (both had prior OpenAI

  partnerships and prior Commonwealth work).

- Contract scope: secure connectors to agency systems, custom

  GPT agents, Codex coding model, prompt libraries, playbooks,

  champions network, employee training.

- Scale: 30,000 x $9 x 12 = ~$3.24M/yr at the 30k-seat tier, per the bid's own annual price (M-032; the $13/mo rate was the 10k-seat tier). Roughly $9.7M in license value over 3 years, plus unknown BCG/Slalom services value. (Corrected 2026-10-06: the earlier $4.7M/yr applied the 10k-seat $13/mo rate to 30,000 seats; at the 30k-seat $9/mo rate the annual figure is $3.24M. The old '$10-14M' range spanned the correct ~$9.7M lower bound and the erroneous ~$14M upper bound.)

- Timing note: the September 2026 federal GSA OneGov deal now

  gives ALL governments $0 license fees + 50% off usage.

  Massachusetts signed in February 2026 at $13-$9/mo, seven

  months before the $0 federal deal existed. Observation, not

  accusation -- but the renegotiation question is legitimate:

  does the state contract have a most-favored-terms clause?

- Sam Altman appeared remotely at the launch event.

CHILD SUPPORT PAYMENT PREDICTER:

- Owner: Executive Office of Administration & Finance (the

  Department of Revenue's Child Support Enforcement division

  sits under A&F).

- Won "Innovation Mastermind" at the first EOTSS AI Showcase

  (fall 2024); built in the EOTSS AI sandbox.

- What it predicts, what data it trains on, what a prediction

  triggers, what recourse a scored parent has: NO PUBLIC

  DOCUMENTATION FOUND. Searched mass.gov, EOTSS materials,

  the MassBuys AI presentation (lists only the award), news

  coverage. This is a finding, not an oversight: an

  award-winning predictive tool scoring child-support payers

  operates with zero public description.

- UK PARALLEL (publictechnology.net): the DWP deploys AI "to

  identify patterns that suggest when someone might stop paying

  child maintenance... helps caseworkers prioritise cases and

  collect evidence more easily." Same opacity: "It is not clear

  exactly what characteristics are considered to be indicators

  of risk, nor what actions can then be taken." The pattern is

  international.

CJIS / LAW ENFORCEMENT CLOUD:

- Massachusetts criminal-justice data (CJIS) is hosted on AWS

  COMMERCIAL cloud, not GovCloud, via EOTSS contract

  (govtech.com). The state says commercial meets CJIS security

  policy.

- The same move "gives Massachusetts law enforcement agencies

  access to AI-powered tools": data analysis, workflow

  automation, security threat detection. The state's crime data

  sits on commercial infrastructure with vendor AI tools

  attached.

WHAT THIS ADDS TO THE INVESTIGATION:

1. The asymmetry thesis (Part 23) now has dollars: the state's

   AI spending buys institutional leverage (faster government,

   scored payers, surveilled streets). No line item buys the

   citizen an advocate.

2. Demand layer: the OpenAI contract is taxpayer-funded AI

   demand -- real money, real use, flowing to the same

   companies in the financing investigation.

3. Procurement opacity is the through-line: a $10M+,

   3-year, 40,000-employee AI deployment via RFQ; an

   award-winning predictor with no public spec; crime data on

   commercial cloud. The public cannot see what was bought,

   what it does, or what it costs in full.

4. The predictor is the sharpest thread: it sits exactly where

   Ricky's pattern lives -- a vulnerable person enters a

   system, the system scores them, the score compounds. Vendor

   unknown, model unknown, recourse unknown.

================================================================

TRANSACTION LEDGER

================================================================

AI TRANSACTION LEDGER v1

Cutoff: October 5, 2026

Rule: follow the obligation, not the headline. An announced commitment is

not counted as cash spent. Unlike figures are never added together.

Status codes:

  CASH      money actually invested or paid

  COMMIT    contractual or announced spending obligation

  DEBT      borrowing or financing

  GUARANTEE contingent exposure, not current cash

  RPO       accounting measure of future contracted revenue, not cash

  PUBLIC    government land, infrastructure, or resource involvement

  FLAG      relationship worth investigating; not evidence of wrongdoing

Evidence levels:

  confirmed   verified against SEC filing, company announcement, or

              government document

  reported    reputable press reporting, underlying document not public

              or not yet pulled

Columns: ID | Date | From | To | Amount | Type | Cash? | Contractual? |

Guarantee? | Duration | Asset | Government exposure | Source | Evidence

T-001 | 2026-02-27 | Amazon | OpenAI | $50B staged equity ($15B + $35B conditional) | CASH/COMMIT | $28.7B by 6/30/26, remainder after | yes | no | staged | OpenAI Series C preferred | none | Amazon 10-Q | confirmed

T-002 | 2026 | OpenAI | AWS | $100B over 8 years (expansion of existing $38B) | COMMIT | no | yes | no | 8 yrs | cloud compute | none | Amazon filing | confirmed

T-003 | 2026 | OpenAI | AWS | ~2 GW Trainium capacity | COMMIT | no | yes | no | multi-yr | AI chips/compute | none | Amazon filing | confirmed

T-004 | 2026-04-20 | Amazon | Anthropic | $5B immediate + up to $20B additional conditional (atop $8B prior) | CASH/COMMIT | $10B invested in nonvoting preferred stock during Q2 2026 | yes | no | staged | Anthropic preferred stock | none | Amazon 10-Q | reconciled

NOTE (2026-10-06): Reconciled against Amazon's Q2 2026 10-Q (filed 2026-07-31): the $10B invested during Q2 = $5B Series G + $5B Series H nonvoting preferred stock; the Series H $5B was drawn under the $20B facility announced April 20, reducing the available facility to $15B. The April announcement's '$5B immediate' and the filing's '$10B quarterly total' describe different scopes of the same transactions. Both figures retained per the contradiction rule.

T-005 | 2026-04-20 | Anthropic | AWS | >$100B over 10 years, to 5 GW | COMMIT | no | yes | no | 10 yrs | cloud compute | none | company announcements | confirmed

T-006 | 2026-09-29 | Anthropic | Google | $111.1B (Apr 2026-Jul 2033, shortfall payments) | COMMIT | no | yes | no | ~7 yrs | cloud/AI infrastructure | none | Reuters (non-public prospectus) | reported

T-007 | 2026-09-29 | Anthropic | Amazon | $110B (May 2026-Apr 2036) | COMMIT | no | yes | no | 10 yrs | cloud/AI infrastructure | none | Reuters (non-public prospectus) | reported

T-008 | 2026-09-29 | Anthropic | Microsoft | $31.4B (Nov 2026-May 2033, non-cancelable except breach) | COMMIT | no | yes | no | ~6.5 yrs | cloud/AI infrastructure | none | Reuters (non-public prospectus) | reported

T-009 | 2026-09-29 | Anthropic | Broadcom | $161.2B equipment lease obligations, largely non-cancelable | COMMIT | no | yes | no | multi-yr | TPU racks | none | Reuters (non-public prospectus) | reported

T-010 | 2026-09-29 | Anthropic | xAI | to $84.5B NVIDIA-based capacity through 2029 | COMMIT | no | largely cancelable (90-day) | no | to 2029 | GPU capacity | none | Reuters (non-public prospectus) | reported

T-011 | 2026-09-29 | Anthropic | AMD | >$20B compute + to $5B equity purchase | COMMIT | no | yes | no | multi-yr | compute + stock | none | Reuters (non-public prospectus) | reported

T-012 | 2026-10 | Broadcom | Anthropic | to $42B convertible notes for TPU leasing | DEBT | undrawn | facility | no | n/a | credit facility | none | press on IPO filing | reported

T-013 | 2026-06-08 | Broadcom | Anthropic (via Apollo/Blackstone XPV) | to $29B lease-payment backstop | GUARANTEE | no | contingent | yes (cap) | grows/shrinks with deployment | TPU racks | none | secondary reports on SEC filing | reported

T-014 | 2026-08-17 | NVIDIA | SB Energy PORTS-Pike | to $105B residual-value guarantee, ~4.25 GW initial (+~3.8 GW option) | GUARANTEE | no | contingent | phases from ~2028/2029 | Ohio data-center campus | DOE former uranium site land | NVIDIA 8-K | confirmed

T-015 | 2026-08-17 | NVIDIA | SB Energy | $3B ($1.5B placement + $1.5B prepaid forward) | CASH | yes | n/a | no | n/a | SB Energy equity | none | company announcements | confirmed

T-016 | 2026-08-17 | OpenAI affiliate | SB Energy | 20-year lease, 8 IT-GW customer | COMMIT | no | yes | no | 20 yrs | Ohio campus tenancy | DOE site | NVIDIA 8-K | confirmed

T-017 | 2026-08-10 | NVIDIA + Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, KKR | AI infrastructure | >$500B third-party capital platforms | COMMIT | no | MOUs, non-binding | over time | financing platforms | none | NVIDIA announcement | confirmed

T-018 | 2026-07-26 | NVIDIA | AI clouds | $36B AI-cloud commitments | COMMIT | no | yes | no | n/a | land/power/shell | none | NVIDIA filing | confirmed

T-019 | 2026-06-30 | lenders | CoreWeave | $35.551B total indebtedness | DEBT | drawn | yes | no | $4.413B due rest of 2026; $6.184B 2027; $4.416B 2028; $2.421B 2029; $3.221B 2030; $14.896B after | GPU/HPC infrastructure | none | CoreWeave SEC filing | confirmed

T-020 | 2026-H1 | lenders | CoreWeave | $8.5B DDTL + $3.1B DDTL + $2.75B 9.75% notes + $1.25B 9.625% notes + EUR 2B 8.5% notes + $4B converts; rates to 15% | DEBT | drawn | yes | no | various | GPU/HPC infrastructure | none | press on SEC filing | reported

T-021 | 2026 | Meta | Hyperion venture | $12.31B initial lease + $28B residual-value guarantee; Meta 20%, investors 80% | COMMIT/GUARANTEE | no | lease yes; guarantee contingent | leases from 2029, to 20 yrs | data-center venture (SPV Beignet Investor) | none | Meta SEC filing | confirmed

T-022 | 2026 | Meta | infrastructure | $130-145B capex guidance | COMMIT | guidance | n/a | no | 1 yr | servers, data centers, network | none | Meta results | confirmed

T-023 | 2026 | Alphabet | infrastructure | $180-190B capex guidance | COMMIT | guidance | n/a | no | 1 yr | servers, data centers, network | none | Alphabet results | confirmed

T-024 | 2026 | Amazon | infrastructure | ~$220B capex guidance | COMMIT | guidance | n/a | no | 1 yr | property and equipment | none | Amazon results | confirmed

T-025 | FY2026 | customers | Oracle | $638B remaining performance obligations; $75B prepaid or customer-supplied GPUs | RPO | $75B prepaid | yes | no | multi-yr | cloud contracts | none | Oracle 8-K | confirmed

T-026 | FY2026 | lenders/investors | Oracle | $43B debt + $5B equity raised; ~$40B more planned FY2027 | DEBT/CASH | raised | n/a | no | n/a | cloud infrastructure | none | Oracle filing | confirmed

T-027 | 2025-01-21 | OpenAI, SoftBank, Oracle, MGX | Stargate | to $500B over 4 years; $100B initial deployment | COMMIT | ~$100B initial | announced target | no | 4 yrs | US AI infrastructure | White House announcement | company announcements | confirmed

T-028 | 2026-06-01 | Related Digital, Blackstone funds, PIMCO | Stargate Michigan ("The Barn") | ~$16B construction + ~$40B Oracle fit-out; ~$2B equity + ~$14B debt | DEBT/CASH | partial | project financing | multi-yr | 1 GW+ campus, Saline Township MI | state/local involvement | Oracle press release + press | reported

T-029 | 2026-03-20 | DOE | SB Energy / AEP Ohio | land lease; 10 GW data center + 10 GW generation (9.2 GW gas); $4.2B transmission (SB Energy pays); $40M community benefits; $33.3B Japanese funding referenced | PUBLIC | n/a | partnership | multi-yr | Portsmouth Site, OH (former gaseous diffusion plant) | federal land lease; ratepayer-protection claim testable | DOE | confirmed

T-030 | 2026-07-29 | DOE + Brookfield, NextEra, utilities | Paducah | >$100B private investment target; 1.8 GW campus; 2 GW gas; to 2.6 GW storage | PUBLIC | target only | announced | to 2031 | Paducah Site, KY | federal site redevelopment | DOE | confirmed

T-031 | 2026-07-20 | NNSA | Amentum | negotiation for phased lease; 1 GW data center + ~2 GW generation (gas to nuclear) | PUBLIC | pre-lease | negotiation | n/a | Savannah River Site, SC | federal site | DOE/NNSA | confirmed

T-032 | 2026 | Akamai | Anthropic | ~$11.6B over 7 years, dedicated cloud capacity | COMMIT | no | yes | no | 7 yrs | cloud capacity | none | Akamai SEC filing | reported

T-033 | 2026 | Meta | CoreWeave | $21B through 2032 | COMMIT | no | yes | no | to 2032 | AI compute (revenue side) | none | press on SEC filing | reported

T-034 | 2026 | Galaxy / NVIDIA / CoreWeave | Helios project | ~$10.4B minimum lease payments; NVIDIA ~11.5% of Galaxy; NVIDIA to $6.3B residual cloud-capacity purchase | COMMIT/GUARANTEE | partial | yes | multi-yr | 400 MW new, 1.63 GW approved, 800 MW leased | none | Galaxy financing documents | reported

T-035 | 2026-08-19 | FTC | -- | proposed enforcement policy on personalized (surveillance) pricing; 2-0 vote; not a ban | regulatory | n/a | n/a | n/a | n/a | n/a | none | FTC/Reuters | confirmed

DO-NOT-SUM OVERLAPS

  T-006 through T-011 are components of the ~$518B Anthropic program; do not add to it.

  T-012 and T-013 are credit support ON the T-009 stack, not additional obligations.

  T-014 is a component of the FT's ~$300B off-balance-sheet estimate; do not add.

  T-001 and T-002 are opposite-direction flows between the same two companies (circular).

  T-004/T-005 overlap T-007 (same Amazon-Anthropic relationship).

  T-025 (Oracle RPO) includes OpenAI contracts also inside T-027 (Stargate target).

  T-017 (NVIDIA MOUs) and the Broadcom XPV platform are parallel, both largely non-binding.

TARGET METRIC (to compute as rows grow)

  The amount of AI infrastructure whose economics depend on other AI

  companies: for each row, flag whether the counterparty is itself an AI

  company or AI-financed vehicle, then sum COMMIT + DEBT + GUARANTEE rows

  with that flag, keeping instrument types separate.

PROJECTS TABLE (seed)

  P-01 | Stargate Abilene TX | first Stargate site | OpenAI/SoftBank/Oracle/MGX | power TBD | debt TBD | government: White House announcement

  P-02 | Stargate Michigan ("The Barn"), Saline Township MI | 1 GW+ | Related Digital, Blackstone funds (equity), PIMCO (debt) | ~$16B build + ~$40B Oracle fit-out | government: state/local

  P-03 | PORTS-Pike, Pike County OH | 4.25 GW initial (+~3.8 GW option), 8 IT-GW OpenAI | SB Energy builds/owns/operates; OpenAI 20-yr tenant; NVIDIA to $105B guarantee | government: DOE former uranium site, federal land lease

  P-04 | Portsmouth, OH | 10 GW data center + 10 GW generation (9.2 GW gas) | SB Energy, AEP Ohio | $4.2B transmission (SB Energy pays); $40M community benefits | government: DOE land lease, public-private partnership

  P-05 | Paducah, KY | 1.8 GW campus; 2 GW gas; to 2.6 GW storage | Brookfield, NextEra, Big Rivers Electric, Jackson Purchase, Paducah Power | >$100B private target | government: DOE site redevelopment

  P-06 | Savannah River, SC | 1 GW data center + ~2 GW generation | Amentum (negotiating lease) | pre-lease | government: NNSA site

  P-07 | Hyperion (Meta) | data-center venture | Meta 20%, Blue Owl-led 80% (SPV Beignet Investor) | ~$27B development; $12.31B Meta lease; $28B Meta residual guarantee | government: none reported

PEOPLE / CONTROL TABLE (seed; public roles only)

  Masayoshi Son | SoftBank | Chairman, Stargate | announced Jan 21, 2026 (White House)

  (to be expanded: CEOs, founders, directors, major shareholders, financiers,

   officials, former officials, lobbyists, researchers, board members --

   factual roles only, no associations presented as allegations)

================================================================

MASSACHUSETTS LEDGER

================================================================

MASSACHUSETTS LEDGER v1

Cutoff: October 5, 2026

Branch of the AI Extraction Investigation: the Massachusetts political,

regulatory, and ratepayer layer. Distinct from the separate

Auditor-anchored Massachusetts money-trail investigation (different

subject, different file).

Rule: follow the obligation, not the headline. Collect what officials DO

(dates, orders, bills, votes, dockets), not only what they SAY.

Status codes (inherited): CASH, COMMIT, DEBT, GUARANTEE, RPO, PUBLIC, FLAG.

Added for this branch:

  TAXEXP    tax expenditure: revenue the government forgoes

  INFRACOST cost of generation, transmission, distribution

  RATECOST  amount ultimately borne by electricity customers

  PROJBEN   projected benefit: promised jobs, tax revenue, development

Evidence levels: confirmed (checked against a primary or reputable

secondary source this investigation read directly); reported (blank-AI

cross-check material not yet independently pulled).

KEY CONTEXT (confirmed): Massachusetts does not currently have many large

data centers planned; the Healey administration describes EO 658 as a

preemptive framework. That makes MA a policy laboratory: the state is

writing the cost-allocation rules BEFORE the load arrives, which is

exactly when the rules matter most.

EXECUTIVE BRANCH

M-001 | 2026-06-25 | Gov. Maura Healey | Released responsible data-center

development framework; paused applications for the state data-center

sales/use-tax exemption | TAXEXP (amount not yet quantified; pull DOR and

application records) | framework confirmed via EO 658 coverage | reported

M-002 | 2026-09-08 | Gov. Maura Healey | Executive Order 658: data centers

over 25 MW peak demand must (a) conform to the June framework, (b) submit

a community-benefits agreement meeting OEJE standards, before any state

permit issues; no NDAs between state agencies and developers; must procure

enough new clean electricity to meet annual consumption or pay into a

Ratepayer Protection Fund (MassDEP to set up payment mechanism by Dec 31,

2026); DPU directed to finalize large-load rate schedules and to clear

speculative projects from interconnection queues via fees/deposits |

PUBLIC/regulatory | confirmed by multiple outlets (Beveridge & Diamond,

Mondaq, TechTarget, local press)

M-003 | 2026-03 | Gov. Maura Healey | Directed administration to pursue

10 GW of new energy resources by 2035 and 5 GW of new storage | COMMIT

(announced target) | do NOT assume this is an AI/data-center program;

question to test: how much of the planned buildout is attributable to

ordinary demand vs new data-center load | reported

M-004 | 2026-09-14 | Gov. Maura Healey | Called for stronger federal AI

safeguards and independent oversight (safety, economic effects, national

security) | statement | second MA track: computational governance, not to

be conflated with physical infrastructure | reported

M-005 | 2026-09-18 | Gov. Maura Healey | Called for independent evaluations

of powerful AI models and stronger serious-incident reporting in

MA economic-development legislation | statement | reported

FEDERAL DELEGATION

M-006 | 2025-12-16 | Sens. Warren, Van Hollen, Blumenthal | Opened

investigation into Google, Microsoft, Amazon, Meta, CoreWeave, Digital

Realty, Equinix over data-center effects on household electricity costs;

cited 13% national household electricity increase since Jan 2025 | inquiry

| reported

M-007 | 2026-01-22 | same | Released company responses; senators said

companies did not answer on utility contracts, actual rates paid,

infrastructure costs, or mechanisms preventing cost-shifting to households

| missing data | the confidential utility contracts are a priority

public-records target | reported

M-008 | 2026-01 | Sens. Warren, Markey | MA energy report: claimed federal

policy changes affected >$8.6B of MA investment and >16,700 jobs |

PROJBEN/claim | their estimates; trace methodology before using | reported

M-009 | 2026-03-26 | Sens. Warren, Hawley | Requested mandatory

data-center energy reporting: consumption, prices, upfront payments,

deposits, AI-server consumption, grid-upgrade costs | inquiry, national

scope | reported

M-010 | 2026-07-10 | Sen. Markey | AI Accountability Agenda: AI effects on

workers, communities, energy, environment, health, rights; proposes data

centers fund clean energy/storage for their capacity needs | proposals

| reported

M-011 | 2026-07-13 | Sen. Markey | Discussion draft, national data-center

framework: energy costs, pollution, fossil-fuel infrastructure, community

and health impacts | draft | reported

M-012 | 2026-06-09 | Sen. Markey, Rep. Beyer | Reintroduced AI

Environmental Impacts Act of 2026: AI data centers report environmental

and energy impacts; NIST to set measurement standards; EPA studies

lifecycle impacts | bill | reported

M-013 | 2026-03 | Sen. Markey + senators | Asked state utility regulators

to prevent data-center grid-cost shifting: large-load rate structures,

cross-subsidization, PUC policies, FERC coordination, load forecasting,

cost allocation | inquiry | reported

M-014 | 2026-01-28 | Sen. Markey + New England senators | Questioned

ISO-New England; cited +13% residential electricity prices in first 9

months of 2025, with data-center demand contributing to pressure | claim |

keep SEPARATE from M-006's national 13% figure; pull EIA/ISO-NE data

directly | reported

M-015 | 2026 | Rep. Pressley + bipartisan lawmakers | Letter on proposed

$33.4B AES acquisition by BlackRock Global Infrastructure Partners, EQT,

others: electricity rates, private-equity returns, utility

infrastructure, data-center demand, cross-subsidization risk | FLAG |

BlackRock/GIP is already a node in the capital map (T-017); this is the

ownership layer meeting the ratepayer layer | reported via Reuters

STATE LEGISLATURE

M-016 | 2025-12-05 | Rep. Rodney Elliott | Filed HD.5404: applies to data

centers >=10 MW; would bar residential rate increases or reduced

availability because of large data centers; requires utilities to give

regulators cost-attribution and interconnection records | bill |

identifies exactly the records to pull: utility cost attribution,

interconnection agreements, data-center electricity sales,

infrastructure costs, rate impacts | reported

M-017 | 2026-02-26 | MA Legislature | H.5175 energy-affordability bill:

requires utility/aggregator efficiency-plan changes producing at least

$1B in aggregate ratepayer savings (a PROPOSED requirement, not savings

achieved); would set a data-center electricity tariff and require DPU

filings | bill | reported

M-018 | 2026 | Joint Committee on Telecommunications, Utilities and Energy

| Senate: Barrett (chair), Brady (vice), Cronin, Cyr, Fernandes, Tarr

(ranking); House: Cusack (chair), Kushmerek (vice), MacGregor, Turco,

Robertson, Tyler, Higgins, Scarsdale, Reid, Jones (ranking), Ferguson |

committee node: sits over telecom, utilities, energy | reported

M-019 | 2026 | Joint Committee on Advanced Information Technology, the

Internet and Cybersecurity | Senate: Moore (chair), Payano (vice),

Barrett, Finegold, Mark, Durant (ranking); House: Farley-Bouvier

(chair), Hawkins (vice), Ouellette, Flanagan, Moakley, Ramos, Owens,

Vitolo, Meschino, Lombardo (ranking), Gaskey | AI policy/governance

branch, separate from electricity/infrastructure | reported

REGULATORY

M-020 | 2026 | MA Department of Public Utilities | Commissioners: Jeremy

McDiarmid (chair, from Oct 2025), Elizabeth Anderson (from Oct 2025),

Staci Rubin | the DPU, not Congress or the governor, sets the utility

rate structures | reported

M-021 | 2026-09-24 | DPU | "Directive to Electric Distribution Companies

regarding data center interconnection requests, pursuant to Executive

Order No. 658," signed by Andrew Strumfels (DPU Director of Government

Relations & Public Affairs), addressed to Britland (Eversource), Hon

(National Grid), Asbury (Unitil) | DOCKET: D.P.U. 26-DC (each utility

files a company-specific informational filing) | DEADLINE: close of

business Friday, October 23, 2026 -- no filings found as of October 5,

which is expected, not delinquent | the order is sweeping: every >25 MW

load project with facility type; every >1 MW data center with queue

position, status, estimated interconnection cost, timeline; full queue

data including financial commitments, system impacts with

customer-vs-ratepayer cost splits, project attrition; Excel workpapers

with formulae intact | INFRACOST dataset | WATCH: DPU File Room,

docket 26-DC, after October 23 | confirmed (primary directive read)

M-022 | 2024-11-20 | MA tax code | Data-center sales/use-tax exemption

effective (M.G.L. c. 64H section 6(zz)): covers equipment, software,

electricity consumed, construction/renovation; thresholds 100,000 sq ft,

$50M qualified costs, 100 MA jobs; certification to 20 years | TAXEXP:

Governor's FY2026 tax-expenditure budget estimated $17.0M for FY2025 and

$17.0M for FY2026 | DECISIVE FINDING (October 5, 2026): per a

public-records response obtained by The Springfield Republican, ZERO

applications were ever submitted before Healey paused the program on

June 25, 2026. Zero applications means zero certifications and zero

annual reports. The $17M/year estimate is therefore ENTIRELY

PROSPECTIVE -- an estimate of what the exemption would cost IF used,

not revenue actually forgone. The pause froze a program with no

participants. | confirmed (statute, budget, regulation); the zero-

applications finding via press reporting of a public-records response

WATCH LIST (dated; check after each date passes)

  - D.P.U. 26-DC utility filings -- due close of business Oct 23, 2026    (highest value)

  - EOED application resumption -- no date ("until further notice")

  - MassDEP clean-energy compliance payment mechanism -- due Dec 31, 2026

  - Municipal Guidance Document -- due Dec 31, 2026

  - First annual EEA/EOED/EOTSS governor's report -- September 2027

  - H.5175 conference committee -- no enactment as of Oct 5, 2026

M-023 | 2025-12 | Rep. Bradley Jones Jr. (+ Ferguson, Frost, Smola, Pease)

| H.83: special legislative commission to study electric-load growth

attributable to AI and data centers; favorably reported Dec 2025,

discharged to House Rules Jan 8, 2026 | investigative | tries to answer

how much future MA load is AI/data-center driven | reported

M-024 | 2026-06/07 | Sen. Michael O. Moore | Amendment: moratorium on

municipal/state approval of data centers >=20 MW until Nov 1, 2027, plus

a MA Data Center Coordination Council (EEA, economic development, AG

ratepayer advocate, DPU, Broadband Institute, DEP, utilities,

municipalities, environmental, labor, host-community resident) | NOT

ADOPTED | the political fork: regulate-and-condition (Healey) vs

stop-and-study (Moore) | reported

M-025 | 2026 | MA Legislature | Economic-development legislation: $325.1M

in bond obligations + $100M direct FY2026 appropriations | NOT

AI/DATA-CENTER SPECIFIC -- broader package; included only in case later

research shows recipient overlap with AI infrastructure | reported

M-026 | 2026-04-15 | Sens. Warren, Hawley | Secured EIA commitment to

develop a MANDATORY nationwide data-center energy survey (Paperwork

Reduction Act process; pilots in TX, WA, Northern VA/DC); target

completion Sept 30, 2026 -- that date has PASSED as of this writing;

check whether the survey was completed | would cover electricity

consumption, utility costs, energy sources, operations; the national

comparison dataset | reported

M-027 | 2026-07-13 | Sen. Markey | "Protecting Communities from Data

Center Impacts Act" discussion draft: federal certification before

construction; energy/environmental/economic standards; pay for grid

infrastructure; fund renewables and storage; reduce demand during grid

stress; labor standards | PROPOSED, not law | reported

BOARD NUMBERS (keep separate, do not combine)

  $17M/yr  estimated MA sales-tax expenditure, data-center exemption

  $1B      PROPOSED minimum ratepayer savings in H.5175, not realized

  25 MW    EO 658 threshold for heightened state requirements

  10 MW    HD.5404 threshold for proposed ratepayer protections

  20 MW    Moore moratorium threshold (not adopted)

THE MASSACHUSETTS ACCOUNTABILITY QUESTION (standing)

Does the actual money trail match the stated policy? Healey's framework

says residents should not bear data-center energy, infrastructure, and

water costs. Warren and Markey's federal investigation says Big Tech did

not provide enough information to establish whether they pay their full

share. The DPU is ordering new disclosures. That is three independent

pointers at one factual question: exactly how much of the AI/data-center

infrastructure bill is paid by the operator, and how much lands elsewhere.

Answer with: company -> facility -> MW -> utility -> grid upgrade ->

estimated cost -> who pays -> tax incentives -> local tax revenue ->

jobs -> water -> generation source -> AI customer -> lobbying.

NEXT PULLS

  1. The DPU docket for the interconnection-queue disclosure order

     (confirm M-021; get the actual utility filings when submitted).

     Note: existing public interconnection data is DG-focused; the

     data-center large-load queue data is what the DPU order would add.

  2. DOR records on the paused data-center tax exemption: which companies

     applied before June 25, 2026, proposed investments, and per-company

     tax expenditure (converts the $17M statewide estimate into a

     company ledger).

  3. The Warren investigation company responses (Jan 22, 2026) to list

     exactly which questions went unanswered.

  4. EIA and ISO-New England data behind the two 13% figures.

  5. HD.5404, H.5175, H.83 current status, votes, and committee actions.

  6. Whether the EIA mandatory data-center energy survey (target Sept 30,

     2026) was actually completed.

EVIDENCE BOUNDARY (October 5, 2026, updated after source audit)

  Independently verified by this investigation: EO 658 and its DPU

  directives (multiple outlets); MA's existing interconnection reporting

  being DG-focused (mass.gov documents); the $31M MGHPCC/AICR grant of

  May 6, 2025 (MassTech, MGHPCC, archived governor's release); the

  Holyoke 9-4 data-center ban of June 16, 2026 with the 12 MW MGHPCC

  carve-out, plus Lowell/Shutesbury/Westfield actions (WAMC, Telegram).

  Confirmed via the blank AI's source audit with named primary

  documents (executive orders, statutes, regulations, Senate PDFs,

  legislative dockets, DPU directive): M-001, M-003, M-004, M-005,

  M-006, M-007, M-009 through M-025, M-027. See SOURCE AUDIT section.

  Still not independently pulled by either pass: M-030 through M-040

  (Applied AI Hub budget, AI Models Challenge, ChatGPT procurement

  file, AMU model, Carahsoft history, 0% allegation, NAGE objection,

  FutureTech Act, Amendment 193, 400 CMR 9 details, Chestnut River

  economics). The blank AI cites specific primaries for several;

  they stay marked reported until read directly.

SOURCE AUDIT RESULTS (October 5, 2026)

  The blank AI returned primary sources for the 27 source questions.

  Most claims held up. Four corrections resulted -- recorded here

  because corrections are the audit's real product.

  CONFIRMED WITH PRIMARY DOCUMENTS (upgraded from reported):

  - M-001: June 25, 2026 framework ("Statement of Expectations for

    Responsible Data Center Development and Operations") and the

    tax-exemption pause announcement; EOED's page confirms applications

    halted June 25.

  - M-003: March 16, 2026 Executive Order to Secure Massachusetts'

    Energy Future: 10 GW new supply/demand-side resources (including

    4 GW solar and 3.5 GW demand reduction) + 5 GW storage.

  - M-004: September 14, 2026 gubernatorial statement calling for

    federal AI safeguards (a statement, not an order or law).

  - M-005: September 18, 2026 statement on AI safety/transparency in

    the Mass Wins Act; the Senate text contains a proposed

    Transparency in Frontier Artificial Intelligence Act.

  - M-006: December 16, 2025 Warren/Van Hollen/Blumenthal letters

    (PDF on the Senate site) to Google, Microsoft, Amazon, Meta,

    CoreWeave, Digital Realty, Equinix.

  - M-007: January 22, 2026 company responses released (PDF); the

    Senate page says companies did not provide utility contracts,

    actual rates paid, or infrastructure-cost detail. Note: "dodged

    accountability" is the senators' characterization, not a

    regulator's finding.

  - M-009: March 26, 2026 Warren/Hawley letter (PDF).

  - M-010: July 10, 2026 Markey AI Accountability Agenda (full PDF).

  - M-011: July 13, 2026 Markey data-center discussion draft (PDF).

  - M-012: June 9, 2026 AI Environmental Impacts Act reintroduction

    (bill text PDF).

  - M-013: March 6, 2026 Markey letter to state energy regulators

    (PDF; signatories Markey, Blumenthal, Van Hollen, Booker).

  - M-014: January 28, 2026 letter to ISO-New England CEO (Markey,

    Welch, Shaheen, Blumenthal, Reed, Whitehouse); the 13% figure is

    the senators' cited statistic, not their own finding.

  - M-015: September 29, 2026 bipartisan letter (Senate Banking

    Committee) urging rejection of the ~$33.4B AES acquisition by

    BlackRock's GIP and EQT; signatories include Warren, Carson,

    Spartz, Tlaib, Pressley.

  - M-016: HD.5404 docket confirmed (filed Dec 5, 2025; House Rules

    Dec 11).

  - M-017: H.5175 confirmed. Status: House passed 128-27 (Feb 26);

    Senate passed amended 32-8 (July 1); House non-concurred July 16;

    conference committee appointed and met July 29; NO final enactment

    as of October 5, 2026.

  - M-018/M-019: committee rosters confirmed (194th General Court).

  - M-020: DPU commissioners confirmed via the directive header.

  - M-021: September 24, 2026 DPU "Directive to Electric Distribution

    Companies regarding data center interconnection requests, pursuant

    to Executive Order No. 658" -- CONFIRMED as a primary document,

    addressed to Britland (Eversource), Hon (National Grid), Asbury

    (Unitil). The directive exists; the resulting utility filings are

    the dataset to watch for.

  - M-022: $17.0M tax expenditure confirmed (FY2026 Governor's Budget,

    Tax Expenditure Budget item 3.007); statute M.G.L. c.64H section

    6(zz); regulations 400 CMR 9.00. The $17M is an estimate of

    foregone revenue, not cash paid out.

  - M-023: H.83 confirmed (filed Jan 17, 2025; favorable Dec 24, 2025;

    discharged to House Rules Jan 8, 2026; not enacted).

  - M-024: Moore's Amendment 2 to S.3143 ("Data Center Moratorium and

    Coordination Council") confirmed; the legislative record shows it

    was REJECTED.

  CORRECTION 1 -- M-008 ($8.6B / jobs): confirmed as a claim in the

  Warren/Markey January 2026 report, but the methodology is SECONDARY.

  Footnote 6 cites Climate Power's "Trump's Energy Crisis" (Dec 2025)

  for the figure (over $8.6B, over 16,750 jobs -- not 16,700); the

  offices also interviewed 13 MA stakeholders. Do not present $8.6B as

  an independently audited Massachusetts government number.

  CORRECTION 2 -- M-025 (S.3178): the "not AI-specific" label was

  WRONG. S.3178 (July 16, 2026, economic development) contains: $75M

  for an AI technologies grant program; $100M defense-sector capital

  explicitly including AI, cybersecurity, robotics, autonomous

  systems, semiconductors; a proposed Transparency in Frontier AI Act;

  third-party evaluation, incident reporting, and transparency

  requirements; a special commission on frontier AI regulation.

  Correct label: economic-development package WITH substantial

  AI-specific provisions. This makes S.3178 one of the most relevant

  documents in the branch.

  CORRECTION 3 -- monthly interconnection reports: WITHDRAWN as a

  data-center queue source. The DOER "Utility Interconnection in

  Massachusetts" reports cover the distributed-generation

  interconnection system, not the new large-load/data-center queue.

  (Matches this investigation's independent finding.) The queue data

  comes from the September 24 DPU directive, whose resulting filings

  are pending.

  CORRECTION 4 -- M-026 (EIA survey): PLANNED, completion NOT

  VERIFIED. EIA's March 25 document describes three voluntary pilots

  (Texas, Washington, Northern Virginia/DC; 196 companies). No primary

  document found showing the mandatory nationwide survey launched by

  September 30, 2026.

V2 ADDITIONS: PUBLIC AI MONEY, PROCUREMENT, HOLYOKE (October 5, 2026)

M-028 | 2025-05-06 | Gov. Healey / MassTech | $31M state grant to MGHPCC

for the Artificial Intelligence Compute Resources (AICR) project,

Holyoke; six universities (BU, Harvard, MIT, Northeastern, UMass, Yale)

to match; total joint investment expected >$120M by 2030; built by

Cambridge Computer with Dell, VAST Data, NVIDIA; hundreds of NVIDIA B200

and RTX GPUs; 100% carbon-free Holyoke hydro power; at least 40% of

compute time intended for startups, entrepreneurs, non-member

colleges/nonprofits | CASH (grant) + COMMIT (projected total) | this is

public capital buying physical AI compute -- the public-interest AI

system, distinct from commercial data centers | confirmed

M-029 | 2026-06-16 | Holyoke City Council | Voted 9-4 to prohibit new

commercial data centers citywide via zoning, explicitly carving out the

existing MGHPCC facility up to 12 MW; mayor signed | regulatory |

context: Chestnut River Power & Infrastructure had proposed a 20 MW,

$200M facility at 100 Water St (former Hampden Papers), below EO 658's

25 MW threshold; residents cited water, air, health, PFAS-in-cooling

risks; developer to look elsewhere. Elsewhere in MA: Lowell enacted a

360-day moratorium (Mar 2026, 10-0) amid a resident lawsuit against the

Markley facility; Shutesbury banned outright; Westfield permitted a

moratorium; Easthampton, Northampton, Greenfield weighing bans |

confirmed (vote and carve-out); project economics ($200M, ~$2M/yr

property tax) reported

M-030 | 2026 | MA capital budget | Applied AI Hub: $10.0M FY2026,

$33.339M FY2027, $68.639M FY2026-30 total, for capital grants in life

sciences, healthcare, climate tech; 2026 Mass Wins Act describes $75M

for Applied AI & Quantum | COMMIT | do NOT add these together --

likely overlapping authorization/budget structures | reported

M-031 | 2025-10 | MassTech | AI Models Innovation Challenge: $2,882,219

to seven projects + >$950,000 matching: Boston Children's $200,014

(Crohn's), EarthDNA $1,000,000 (footwear recycling), Northeastern

$504,043 (coastal risk) + $16,500 (manufacturing), WNEU $500,000 (defect

detection), WPI $381,931 (renewable-fuel digital twin) + $279,731

(industrial recycling) | CASH (grants) | evidence AGAINST the simple

extraction theory: genuine public-benefit AI work | reported

M-032 | 2025-10-17 to 2026-01-27 | MA EOTSS | RFQ 26-04261, AI Assistant

and Implementation Partner: bids opened Oct 17, 2025; bidders had to be

on ITS60 Category 1 or ITS75; Carahsoft Technology Corporation named

apparent successful bidder Jan 27, 2026 (state cautioned this did not

guarantee a final contract); underlying proposal "OpenAI (via

Carahsoft)"; annual platform prices $1.56M (10k seats) / $2.64M (20k) /

$3.24M (30k) / $3.36M (40k) | COMMIT (proposed pricing, not proof of

spend) | need the executed contract and invoices for actual spend |

reported

M-033 | 2026 | OpenAI proposal via Carahsoft | Advanced Model Units

(AMUs): GPT-5 Thinking 10/message, GPT-5 Pro 50, Agent 30, Deep Research

50/task, image 5, voice 5/min; centrally monitored with overage limits;

60 days unlimited AMUs offered for evaluation; concurrent-user model

(40,000 = max provisioned tier, not simultaneous users); OpenAI

estimated 3,500-5,000 AMUs/yr for a policy analyst | COMMIT (pricing

model) | the audit is predicted vs actual usage vs actual productivity

gains | reported

M-034 | 2007-2013 conduct | DOJ / Carahsoft, VMware | $75.5M settlement

over allegations of concealed commercial pricing and government

overcharging; resolved allegations WITHOUT determination of liability;

Carahsoft was NOT charged in the separate 2024-25 federal IT

procurement criminal cases | historical | supports "Carahsoft has a

procurement controversy in its history"; does NOT support any claim

about the MA AI procurement | reported

M-035 | 2026 | activist/legislator Erika Uyterhoeven | Alleged, from

records she obtained, that the AI procurement scoring rubric gave 0%

weight to AI safety and that both finalists came through Carahsoft |

ALLEGATION -- UNVERIFIED | the COMMBUYS file (RFQ, Q&A, cloud terms,

safeguarding language, risk terms, SOW template, award notice) is public;

the scoring sheets are the document to pull; do NOT repeat the 0%

claim as fact | reported as allegation only

M-036 | 2026 | NAGE (15,000 state employees) | Objected to ChatGPT

rollout speed and possible effects on state workers | labor ledger:

AI spending vs hours saved vs positions eliminated vs new AI positions

vs contractor spending -- "AI saves taxpayers money" is a hypothesis

until these numbers exist | reported

M-037 | 2024 | MA Legislature | FutureTech Act: $25M authorized for AI

projects improving government operations and digital services | COMMIT

| reported

M-038 | 2026 | MA Senate | Amendment 193: proposed special

data-center electricity tariffs -- minimum charges, minimum contract

length, minimum monthly billing demand, collateral, notice periods,

capacity-reduction fees, premature termination fees | PROPOSED |

reveals the stranded-cost fear: utility builds for 300 MW of promised

load, developer walks, someone holds the infrastructure | reported

M-039 | 2026 | 400 CMR 9 regulations | Certified data centers must file

ANNUAL reports stating actual jobs, actual investment, actual exempt

expenditures, signed under pains and penalties of perjury; EOED may

require accountant-certified cost accounting; applications and materials

are PUBLIC records disclosable under MA public-records law | regulatory

| the highest-value records vein after the DPU queue data: company ->

facility -> year -> actual exempt purchases -> actual tax avoided |

reported

M-040 | 2026 | Chestnut River Power & Infrastructure | Proposed 20 MW,

$200M data center, 100 Water St Holyoke (former Hampden Papers);

closed-loop cooling, Holyoke hydropower, pitched as minimal municipal

utility impact; potential ~$2M/yr property tax | PROPOSED, locally

rejected | the boundary case: below EO 658's 25 MW threshold, killed

by municipal action anyway -- watch for MW threshold engineering

(19.9 vs 20 vs 25 MW) across projects | reported

THREE SYSTEMS (framing, from the cross-check; adopted)

  SYSTEM A -- Public AI investment: $25M FutureTech, $31M AICR grant,

  ~$120M projected AICR total, $2.88M AI Models grants, $68.64M Applied

  AI Hub capital budget, $75M Mass Wins AI/Quantum authorization. (Do not

  total; overlapping authorizations.)

  SYSTEM B -- Commercial AI infrastructure: Holyoke 20 MW/$200M

  proposal, unknown MA commercial pipeline, unknown interconnection

  costs, unknown private financing, unknown per-company tax

  expenditures.

  SYSTEM C -- Government AI adoption: ~40,000 employees, $1.56M-$3.36M/yr

  quoted platform pricing, Carahsoft apparent bidder, OpenAI via

  Carahsoft, unknown final contract value.

  The systems overlap politically and institutionally. No evidence found

  that they are secretly one operation.

DEAD ENDS (recorded, not buried)

  - No clean, responsibly presentable chain of AI company -> politician

    -> campaign contribution -> data-center approval -> utility benefit

    was found. UNPROVEN, not disproven; correlation is not influence.

  - The Massachusetts investigation produced no meaningful Epstein    connection. The MA AI ecosystem is contemporary and institutionally

    explainable. Forcing the connection would weaken the investigation.

  - The alleged 0% AI-safety procurement score is UNVERIFIED. The

    scoring sheets are the document; the allegation is not the finding.

THE CENTRAL MASSACHUSETTS QUESTION (standing)

  Not "is AI extracting." The concrete mechanism: AI creates enormous

  demand for scarce infrastructure (electricity, transmission,

  substations, land, water, GPUs, construction, labor, permitting

  capacity). Whoever controls those bottlenecks can capture rents.

  The public-interest question: are those rents being captured privately

  while infrastructure risks are socialized? Massachusetts' own

  legislation and executive orders show policymakers are concerned

  enough to design rules against exactly that -- which is stronger

  evidence than any hidden-network theory. And the tension at the heart

  of it: Massachusetts wants to be where AI is developed while saying

  residents shouldn't pay for the infrastructure to develop it. Those

  goals are compatible only if the accounting works. Right now the

  accounting is incomplete.

================================================================

EPSTEIN-AI CORRESPONDENTS (VERIFIED)

================================================================

WHO JEFFREY EPSTEIN TALKED TO ABOUT AI OR INTELLIGENT MACHINES

Master list, rebuilt October 5, 2026, after verification against the DOJ

January 30, 2026 Epstein Transparency Act releases (Datasets 9 through 11),

checked via the Epstein Graph index, archive mirrors, and news reporting.

HOW TO READ THIS FILE

Level A: Epstein personally discusses AI in his own emails. High confidence.

Level B: Someone discusses AI directly with Epstein. He is the recipient.

Level C: AI material tied to Epstein's foundation or funding. This includes

proposals sent to the foundation and foundation publicity. Publicity is not

proof of funding.

Level D: AI material in Epstein's broader network with no direct Epstein

participation. People named here are NOT people Epstein talked to about AI.

This is a lower bound, not a complete count. The archive holds well over a

million documents and search is limited by scan quality and redactions. Some

entries were confirmed through the search index only. Confidence is marked

high, medium, or low per entry. Items marked FLAG merit a closer look. A flag

is not an accusation.

LEVEL A: EPSTEIN PERSONALLY DISCUSSES AI

1.

Date: 2009-08-11

Document: EFTA01816129

Who: Epstein to John Brockman, cc Barnaby Marsh

What: "I intend to fully fund the following gatherings (meetings with lots of

interaction hikes coffee cross pollination)" including "2. Synthetic General

Intelligence.. encompassing, signal processing, statistical and machine

learning, virtual world robots. new programing architecture." Also a

"Conference on Power" covering reputation, awe, trust, deception,

reciprocity, group selection and behavior. Marsh replied with detailed

feedback on Aug 14, 2009 (EFTA02441172).

Epstein personally involved: yes, authored

Funding: pledged to fully fund

Confidence: high

Flag: network-building

Note: document EFTA00883293 was reported but not found; use EFTA01816129.

Source: https://taxpolicy.org.uk/wp-content/mandelson-search/PDFs/EFTA01816129.pdf

2.

Date: 2010-11-08

Document: EFTA00777409

Who: Epstein to Al Seckel

What: Research solicitation listing "Artificial Intelligence and

Computational Science" as a funding area, calling for building an

"intelligent" machine as opposed to a "competent" machine, and for

computational models allowing internal self-revision of new situations.

Seckel reworked the draft the same day.

Epstein personally involved: yes

Confidence: high

Source: https://epsteingraph.com/documents/EFTA00777409

3.

Date: 2011-05-20

Document: EFTA00911704

Who: Epstein to Ben Goertzel

What: "ill help" in reply to an OpenCog Hong Kong funding request.

Epstein personally involved: yes, authored

Funding: pledged

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00911704.pdf

4.

Date: 2011-11-10/11

Document: EFTA00924081

Who: Epstein with Lawrence Krauss

What: Future in Computing Conference logistics ("yes"; "westin, I-- I am

already full"). Krauss asked to come early and scuba dive.

Epstein personally involved: yes

Funding: conference sponsor

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00924081.pdf

5.

Date: 2011-11-11

Document: EFTA00924182

Who: Epstein to Martin Nowak

What: "Yes" to inviting Corina to the Future in Computing Conference.

Epstein personally involved: yes

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00924182.pdf

6.

Date: 2011-11-16

Document: EFTA02036703

Who: Epstein to Marvin Minsky

What: "jeffrey has objections lets talk tomorrow" in direct input on the

Future in Computing Conference guest list. Minsky had written he would start

inviting people unless Epstein objected.

Epstein personally involved: yes

Confidence: high

Source: https://hooks.epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA02036703.pdf&download=true&filename=efta-efta02036703

7.

Date: 2013-02-02

Document: EFTA00953154

Who: Epstein to Boris Nikolic

What: "people for bill" list for Bill Gates naming "artificial intelligence

guys george church genetics, . howard gardner, multiple intelligences."

Epstein personally involved: yes

Confidence: high

Note: index only; Church is a geneticist and Gardner a psychologist, those

are Epstein's labels, not theirs.

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00953154.pdf

8.

Date: 2013-02-12/13

Documents: EFTA01903334, EFTA01907385

Who: Epstein with Roger Schank

What: Epstein asked Schank directly what it would take to build an AI

machine: how many programmers, how much money, how long, what milestones.

Schank estimated about 5 million dollars over about 3 years to build

something like a 3 year old that grows into a 4 year old. Epstein also asked

whether they should attend an AGI conference to see the state of play, and

they debated the AI winter. Epstein forwarded a Stephen Kosslyn to Joscha

Bach thread noting Bach "has thought a lot about ai, goals, etc."

Epstein personally involved: yes

Confidence: high

Note: index only.

Sources:

https://api.epsteingraph.com/api/documents/EFTA01903334/file

https://api.epsteingraph.com/api/documents/EFTA01907385/file

9.

Date: 2013-05-15

Document: EFTA00961344

Who: Epstein to Joel Klein

What: "the fields of synthetic intelligence and artificial General

intelligence might provide you new insights into how systems and humans

learn, Drives, goals, need for competence"

Epstein personally involved: yes, authored

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00961344.pdf

10.

Date: 2013-06-18

Documents: EFTA00963613, EFTA01969748

Who: Epstein to Joi Ito

What: "joscha is good"; "I was bringing him to Harvard, my institute, but

you can have him if you want. hes very thoughtful, is in touch with

synthetic, intelligence, artificial general intelligence. biological inspired

intelligence, and some thoughts on virtual worlds." Asked Ito to have someone

show Joscha Bach around the MIT Media Lab.

Epstein personally involved: yes

Confidence: high

Note: index only.

Sources:

https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00963613.pdf

https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-10/EFTA01969748.pdf

11.

Date: 2013-12-02

Document: EFTA00978072

Who: Epstein to Ben Goertzel

What: "ill do the 45k, of course. and please ping me again in two weeks re

full program" in reply to the AGI R and D proposal.

Epstein personally involved: yes

Funding: 45,000 dollars agreed

Confidence: high

Flag: deception thread

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00978072.pdf

12.

Date: 2013-12-10/11

Document: EFTA01755000 (duplicates EFTA02578230, EFTA00978432)

Who: Epstein to Ben Goertzel

What: Exact verified wording: "i want to see the deception character, how do

we send the funds?" Goertzel replied he wanted to see it too and hoped to

demonstrate AI agents intentionally tricking each other in a game environment

by the end of 2014.

Epstein personally involved: yes

Confidence: high

Flag: deception thread

Source: https://hooks.epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01755000.pdf&download=true&filename=efta-efta01755000

13.

Date: 2013-12-31

Document: EFTA00979508

Who: Epstein to Richard Kahn, his accountant

What: "tomotw" (tomorrow) in reply to Kahn asking when to send the 60,000

dollar wire from STC to Goertzel's Novamente LLC.

Epstein personally involved: yes

Funding: 60,000 dollars pending; completion unconfirmed

Confidence: high

Flag: money chain

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00979508.pdf

14.

Date: 2015-08-29

Document: EFTA_R1_01609589

Who: Epstein to Joi Ito

What: Sent the link http://loveandsexwithrobots.org/#topics (robotics

conference topics).

Epstein personally involved: yes

Confidence: high

Note: index only; no full URL captured.

15.

Date: 2016-04-07

Document: EFTA02465687

Who: Epstein to Jim Rutt

What: "unlikely new york, but maybe south calif" in reply to Rutt's pitch

for an April 18 to 28 North America fundraising tour to accelerate real AGI

using the OpenCog framework.

Epstein personally involved: yes, responded

Confidence: high

Source: https://democraticforum.org/Epstein/January302026Release/DataSet11/VOL00011/IMAGES/0157/EFTA02465687.pdf

16.

Date: 2016-06-02

Document: EFTA00826385

Who: Epstein to Raafat Alsabbagh

What: "If Mohammed really wants to see the future, after seeing the past in

Washington with Obama, we should take him to see the Laboratory at MIT.

advanced robotics. Artificial Intelligence etc."

Epstein personally involved: yes

Confidence: high

Note: the visitor is identified only as Mohammed; no further identification

found.

Source: https://epsteingraph.com/documents/EFTA00826385

17.

Date: 2016-07-11/16

Document: EFTA00667699 (also EFTA00667716, EFTA00823880)

Who: Epstein with Ben Goertzel and Joscha Bach

What: "i can do the 21 anytime as well I with as many people as you like up

to 10" arranging a July 21 meeting at his home around AGI-16. Whether the

meeting occurred is unconfirmed. Also "no jeffrey epstein . is enough" on

sponsor naming, and a critique of an AI and music paper ("first paper is

silly, it does not include rythm, only notes").

Epstein personally involved: yes

Funding: 3,000 dollar AGI-16 sponsorship discussed

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00667699.pdf

LEVEL B: SOMEONE DISCUSSES AI DIRECTLY WITH EPSTEIN

18.

Date: 2009-08-17

Who: Ben Goertzel to Epstein

What: Funding situation summary; asks Epstein to fund a 2 to 3 person AGI

team.

Epstein personally involved: recipient only

Funding: asked

Confidence: medium

Note: filestein copy; EFTA number not captured.

19.

Date: 2009-09

Who: Ben Goertzel to Epstein

What: Asked for support to write a book on AGI; Epstein offered 50,000

dollars through a nonprofit. A February 2010 update followed.

Epstein personally involved: recipient, funder

Funding: 50,000 dollars offered

Confidence: high

Note: per reporting summaries; see also entry 20.

20.

Date: 2010-10-22

Document: EFTA00754807

Who: Ben Goertzel to Epstein

What: "Building Better Minds" first draft complete (900 pages); "As I

recall you read a draft book of mine in 2001"; "Thx much for your funding

which played a key role."

Epstein personally involved: recipient only

Funding: acknowledges past funding

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00754807.pdf

21.

Date: 2011-05-20

Document: EFTA01777318

Who: Ben Goertzel to Epstein

What: "OpenCog Recap - May 2011" progress update.

Epstein personally involved: recipient only

Confidence: high

Source: https://epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01777318.pdf&download=true&filename=efta-efta01777318

22.

Date: 2011-05-20

Document: EFTA01777445

Who: Ben Goertzel to Epstein

What: Hong Kong project funding ask; "Building Better Minds" draft; Gino Yu

"AGI Research Center" 3 million dollar per year idea.

Epstein personally involved: recipient only

Funding: asked

Confidence: high

Source: https://epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01777445.pdf&download=true&filename=efta-efta01777445

23.

Date: 2011-06-24

Document: EFTA00628447

Who: Ben Goertzel to Epstein

What: "full speed ahead toward AGI": 15 people for 4 years at 750,000

dollars per year, 3 million total; "Sputnik of AGI"; asks Epstein for 1.5

million dollars in matching funds. The plan: AGI controlling a videogame

character and a Hanson Robokind humanoid robot at young child intelligence.

Epstein personally involved: recipient only

Funding: 1.5 million dollars asked

Confidence: high

Note: filestein copy; date reported medium-high.

24.

Date: 2012

Document: EFTA00765042

Who: Ben Goertzel to Epstein

What: "Reviews of CogBot proposal, at long last..."

Epstein personally involved: recipient only

Confidence: medium

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00765041.pdf

25.

Date: 2013-01-17

Document: EFTA02679857

Who: Ben Goertzel to Epstein

What: "List of some interesting people" including Michael Tomasello, George

Church, Ed Boyden and others.

Epstein personally involved: recipient only

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-11/EFTA02679857.pdf

26.

Date: 2013-02-11

Document: EFTA02566583

Who: Ben Goertzel to Epstein

What: "why not have the machine take iq tests" on IQ tests versus general

intelligence.

Epstein personally involved: recipient only

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset11_202602/DataSet%2011.zip/VOL00011/IMAGES/0227/EFTA02566583.pdf

27.

Date: 2013-02-13

Who: Ben Goertzel to Epstein

What: "Preschool IQ Test Based Proposal" cover email delivering the Toddler

proposal (attachments include Epstein_Goertzel_IQ_Executive_Summary_Feb12_2013.pdf).

Epstein personally involved: recipient only

Confidence: high

Note: filestein copy; EFTA number not captured. Proposal document is

EFTA01137423, see Level C entry 48.

28.

Date: 2013-06-18

Who: Joi Ito to Epstein

What: "I'm in the middle of trying to recruit a machine learning faculty who

is on the fence between us and Johns Hopkins... I need more good AI

faculty..."

Epstein personally involved: recipient only

Confidence: high

29.

Date: 2013-11-17

Document: EFTA02373594

Who: Ben Goertzel to Epstein

What: "there has been no bigtime Epstein/OpenCog AGI funding event... yet";

notes that last year Epstein donated 20,000 dollars toward the Hong Kong

grant.

Epstein personally involved: recipient only

Funding: seeks ongoing

Confidence: high

Source: https://democraticforum.org/Epstein/January302026Release/DataSet11/VOL00011/IMAGES/0096/EFTA02373594.pdf

30.

Date: 2013-11-27

Document: EFTA02387343

Who: Ben Goertzel to Epstein

What: "deception" theme as a development goal (simulating human deceptive

behavior); roughly 25 million dollar over 3 year proposal idea; Google and

Kurzweil visit mentioned.

Epstein personally involved: recipient only

Funding: proposed

Confidence: high

Flag: deception thread

Source: https://democraticforum.org/Epstein/January302026Release/DataSet11/VOL00011/IMAGES/0104/EFTA02387343.pdf

31.

Date: 2013-12-02

Who: Ben Goertzel to Epstein

What: "AGI R and D Proposal" cover email: "full fledged proposal 'from here

to AGI'" with attachment Epstein_AGI_Proposal_Summary_2013_v6.pdf, which is

EFTA01103465 (see Level C entry 49).

Epstein personally involved: recipient only

Funding: proposed

Confidence: high

Note: filestein copy; EFTA number not captured.

32.

Date: 2013-12-02

Document: EFTA00978072 (in thread)

Who: Ben Goertzel to Epstein

What: Thanks regarding the 45K; hopes to demo "AI agents intentionally

tricking each other by the end of 2014."

Epstein personally involved: recipient only

Funding: 45,000 dollars

Confidence: high

Flag: deception thread

Source: same as Level A entry 11.

33.

Date: 2013-12-10

Document: EFTA01755000

Who: Ben Goertzel to Epstein

What: 45,000 dollar "10 percent corporate donation" routing via Humanity+

(501c3) to Hong Kong Polytechnic University; "deceptioncapable game

character."

Epstein personally involved: recipient only

Funding: 45,000 dollar routing

Confidence: high

Flag: money chain

Source: same as Level A entry 12.

34.

Date: 2013-12-19/20

Document: EFTA00979508

Who: Ben Goertzel to Richard Kahn, Epstein's accountant

What: "it seems Natasha (from Humanity+) has returned the $$ as requested...

Please let me know when Jeffrey's generous grant of $60K has been conveyed

to my Novamente LLC account... being used as 'matching funds' in a Hong Kong

government grant"

Epstein personally involved: recipient via his agent

Funding: 60,000 dollars discussed; completion unconfirmed

Confidence: high

Flag: money chain

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00979508.pdf

35.

Date: 2014

Document: EFTA00709564

Who: Ben Goertzel to Epstein, cc Cosmo Harrigan and Joscha Bach

What: "Intelligent Virtual Animal" 18 month CogPrime and cognitive synergy proposal (file Goertzel_Harrigan_Bach_Epstein_Proposal_2014.pdf).

Epstein personally involved: recipient only

Funding: proposed

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00709564.pdf

36.

Date: 2014-08-15

Document: EFTA00995144

Who: Ben Goertzel to a foundation press contact, cc Epstein

What: "technically, Epstein Foundation hasn't actually funded iCog Labs or

the OpenCog researchers working there... iCog Labs was seed-funded by

Maryland investor Sander Olsen, and currently is self-supporting based on

consulting revenues... I don't mind there being a press release associating

Epstein Foundation with iCog Labs" provided the wording is accurate.

Epstein personally involved: recipient only

Confidence: high

Flag: funding versus publicity gap

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00995144.pdf

37.

Date: 2014-09-16

Document: EFTA02343128

Who: Joi Ito to Epstein

What: Google Alert forward on an Epstein funded computer coding app for

toddlers.

Epstein personally involved: recipient only

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-11/EFTA02343128.pdf

38.

Date: 2014-12-22

Document: EFTA02597271

Who: Ben Goertzel to Epstein

What: "Intelligent Virtual Animal" research proposal follow-up.

Epstein personally involved: recipient only

Funding: proposed

Confidence: high

Source: https://democraticforum.org/Epstein/January302026Release/DataSet11/VOL00011/IMAGES/0249/EFTA02597271.pdf

39.

Date: 2015

Who: Ben Goertzel to Epstein's office

What: 25,000 dollar request. Epstein's assistant replied that "due to the

current environment," funding was suspended. Goertzel wrote: "I have seen the

spate of utterly idiotic negative publicity in the news, and I'm sorry you

guys have to deal with that." Epstein authorized the transfer the same day,

per South China Morning Post reporting.

Epstein personally involved: recipient only

Funding: 25,000 dollars

Confidence: medium

Note: verified via SCMP reporting; the underlying email's EFTA number was not

located.

Source: https://www.thefreelibrary.com/Epstein+aided+AI+pioneer+in+securing+Hong+Kong+funding%3a+Report.-a0875199576

40.

Date: 2015-04-30 / 2015-05-06

Document: EFTA00649660

Who: Marcus Abundis to Epstein

What: Letter on "Strong Artificial Intelligence (AGI) and Evolutionary

Dynamics," noting "Earlier you have sponsored work in artificial intelligence

(Goertzel, Bach -- AGI)" and pitching his own information science model of

consciousness.

Epstein personally involved: recipient only

Confidence: high

Note: index only.

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00649660.pdf

41.

Date: 2015-08-29

Who: Joi Ito to Epstein

What: "We should send Joscha. I know Adrian Cheok." regarding the Love and

Sex with Robots conference.

Epstein personally involved: recipient only

Confidence: high

42.

Date: 2016-03-31

Document: EFTA00830443

Who: Ben Goertzel to Epstein

What: Work toward OpenCog controlling Hanson Robotics robot heads for

embodied natural language dialogue; AGI-16 demo plan; Jim Rutt 3 year plan

aiming to raise 6 million dollars; cognitive synergy (ECAN, MOSES, PLN).

Epstein personally involved: recipient only

Funding: 6 million dollars proposed

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA00830443.pdf

43.

Date: 2016-04-05

Document: EFTA01772984

Who: Jim Rutt to Ben Goertzel, cc Epstein

What: "I just had a brief Skype chat w/ Jeffrey Epstein about our 'Prime

AGI' project, and have sent him our various draft materials"

Epstein personally involved: recipient only

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-10/EFTA01772984.pdf

44.

Date: 2016-04-07

Document: EFTA02465687

Who: Jim Rutt to Epstein

What: April 18 to 28 North America tour "pitching a fund raising to

accelerate the drive for real Artificial General Intelligence using the

Opencog framework"; asks to meet Epstein.

Epstein personally involved: recipient only; Epstein replied (see Level A

entry 15)

Funding: pitched

Confidence: high

Source: same as Level A entry 15.

45.

Date: 2016-05-01

Document: EFTA01803263

Who: Larry Summers, in a thread with Epstein, Joi Ito, Danny Hillis, and

Reid Hoffman

What: After a dinner, Summers wrote he was "intrigued by reed S topic around

AGI" and asked for links to the best surveys to read.

Epstein personally involved: participant in the thread

Confidence: high

Note: index only.

Source: https://epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01803263.pdf&download=true&filename=efta-efta01803263

46.

Date: 2016-05

Who: Joi Ito to Epstein

What: Shared "a paper on the AI that writes Haiku."

Epstein personally involved: recipient only

Confidence: high

47.

Date: 2016-07-10/11

Document: EFTA02458618

Who: Ben Goertzel to Epstein

What: 3,000 dollar AGI-16 and Human-Level AI snacks and reception sponsorship

via Humanity+. Epstein asked "is it a 501 c 3".

Epstein personally involved: recipient only

Funding: 3,000 dollars asked

Confidence: high

Source: https://democraticforum.org/Epstein/January302026Release/DataSet11/VOL00011/IMAGES/0152/EFTA02458618.pdf

48.

Date: 2016-08-04

Document: EFTA01786410

Who: Ben Goertzel to Epstein

What: "OpenCog reasoning on the Hanson Robot": probabilistic logic inference

demo video running through a Hanson robot; "as we were briefly discussing at

your house."

Epstein personally involved: recipient only

Confidence: high

Source: https://hooks.epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01786410.pdf&download=true&filename=efta-efta01786410

49.

Date: 2016-08-14

Document: EFTA01773481

Who: Joscha Bach to Epstein, cc Joi Ito and Martin Nowak

What: Computing architectures explainer (ternary logic, probabilistic, DNA,

and quantum computers).

Epstein personally involved: recipient only

Confidence: high

Source: https://epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01773481.pdf&download=true&filename=efta-efta01773481

50.

Date: 2017-02-25

Document: EFTA02659032

Who: Brock Pierce to Epstein

What: "PLEASE NOTE: I am testing an artificial intelligence based assistant

for scheduling named Amy. She's good but not perfect so if you are having

trouble just ping me again."

Epstein personally involved: recipient only; the AI statement came from

Pierce, not Epstein

Confidence: high

Note: index only.

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset11_202602/DataSet%2011.zip/VOL00011/IMAGES/0288/EFTA02659032.pdf

51.

Date: 2017-05-29

Documents: EFTA02649600, EFTA02649536

Who: Joi Ito to Epstein

What: "One Science / Media Lab" draft initiative text listing "developing,

deploying and regulating artificial intelligence" among hard

interdisciplinary challenges.

Epstein personally involved: recipient only

Confidence: high

Note: index only.

Sources:

https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset11_202602/DataSet%2011.zip/VOL00011/IMAGES/0282/EFTA02649600.pdf

https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset11_202602/DataSet%2011.zip/VOL00011/IMAGES/0282/EFTA02649536.pdf

52.

Date: 2018-09-09

Document: EFTA01023686

Who: John Brockman to Epstein

What: "the meeting" email with AI discussion paragraphs, including Chalmers

on self-reflective AI ("Once we develop AI systems that can reflect on

themselves and reason..."), plus Dyson, Gershenfeld, Gopnik, Griffiths,

Hillis, McEwan, Wilczek and others. Shows the AI and consciousness theme

persisting into 2018.

Epstein personally involved: recipient only

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01023686.pdf

53.

Date: 2009-09

Document: EFTA02437581

Who: Epstein with John Markoff

What: Exchange about summarizing Epstein's science projects.

Epstein personally involved: participant

Confidence: high

Note: index only.

Source: https://democraticforum.org/Epstein/January302026Release/DataSet11/VOL00011/IMAGES/0138/EFTA02437581.pdf

LEVEL C: FOUNDATION AND FUNDING CONNECTED AI MATERIAL

54.

Date: 2010-11-05

Document: EFTA02415741

Who: Al Seckel to Epstein, foundation sponsored

What: "Mindshift Conference": "Jeffrey Epstein to host." Day One topics:

Artificial Intelligence, Complexity Theory, Theoretical Physics,

Evolutionary Biology, Cognitive Neuroscience. St. Thomas and Little St.

James. "Made possible by a generous grant from the Jeffrey Epstein

Foundation." Attendees included Gell-Mann, Koch, Arnold, Sussman, Mohr.

Epstein personally involved: host and sponsor

Funding: funded by grant

Confidence: high

Source: https://assets.getkino.com/documents/EFTA02415741.pdf

55.

Date: 2011

Document: EFTA01114145

Who: Ben Goertzel and Gino Yu, proposal

What: "OpenCog AGI Toddler Project": AGI with 3 to 4 year old child

intelligence via virtual characters and humanoid robots; 15 staff at

750,000 dollars per year for 4 years, 3 million total; based on

OpenCogPrime and "Building Better Minds."

Epstein personally involved: proposal addressed to his orbit

Funding: sought

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01114145.pdf

56.

Date: 2011/2012

Document: EFTA01175884

Who: Ben Goertzel and Joel Pitt, journal article in the files

What: "Nine Ways to Bias Open-Source AGI Toward Friendliness" (Journal of

Evolution and Technology, Vol. 22): 9 techniques including stable goal

systems, slow self-improvement, and a Global Brain link. AI safety material

in Epstein's files.

Epstein personally involved: in files only

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01175884.pdf

57.

Date: 2011 to 2023

Document: EFTA01114164

Who: roadmap in the files

What: "OpenCog Roadmap: 2011-2023": stages from proto-AGI virtual agent to

full human-level AGI.

Epstein personally involved: in files only

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01114164.pdf

58.

Date: 2011-12-09/12

Document: EFTA00423747 (see also EFTA00424153, EFTA00924182)

Who: the foundation, logistics by Cecile de Jongh

What: "Future in Computing Conference sponsored by the J. Epstein VI

Foundation," St. John, US Virgin Islands. Agenda by Dr. Marvin Minsky. Lunch

on Epstein's island on December 10, 2011. Westin St. John. Epstein weighed in

personally on the guest list (see Level A entries 4, 5, 6).

Epstein personally involved: host and sponsor; weighed in on guest list

Funding: funded

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00423747.pdf

59.

Date: 2012/2013

Document: EFTA01103495 (also EFTA01103499)

Who: CogBot document in the files

What: "CogBot": OpenCog and DeSTIN software plus Hanson robot control plus

Festival and Sphinx speech; "childlike" tasks.

Epstein personally involved: in files only

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA01103495.pdf

60.

Date: 2013-01-16

Document: EFTA01103525

Who: Ben Goertzel, proposal to the Epstein Foundation

What: "Creating Intelligent Humanoid Robots Using the OpenCog AGI

Architecture" (CogBot): OpenCog plus DeSTIN plus Hanson Robokind humanoid

robot, aimed at childlike intelligence.

Epstein personally involved: proposal to his foundation

Funding: sought

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01103525.pdf

61.

Date: 2013

Document: EFTA01128781

Who: foundation style biography ("JEFFREY EPSTEIN NET")

What: Describes OpenCog's "overriding goal... to go beyond human

intelligence and create a world of 'super' intelligence." Covers OpenCogPrime,

a Nao humanoid, natural language apps, AtomSpace, CogServer, and a Bach and

MicroPsi section.

Epstein personally involved: foundation publicity, not independent evidence

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01128781.pdf

62.

Date: 2013

Document: EFTA01130113

Who: foundation publicity

What: "Science Philanthropist, Jeffrey Epstein, Backs the First Free

Thinking Robots" (OpenCog; AtomSpace described).

Epstein personally involved: foundation publicity, not independent evidence

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01130113.pdf

63.

Date: 2013

Document: EFTA01139627

Who: foundation publicity

What: "A Harvard Financier, Jeffrey Epstein, Advances Artificial

Intelligence in Ethiopia" (Addis AI Lab and iCog Labs; Goertzel, Getnet

Asefa; Sander Olson seed funding).

Epstein personally involved: foundation publicity; contradicted by Goertzel's

2014 email (Level B entry 36), which says the foundation never funded iCog

Labs

Confidence: high

Flag: funding versus publicity gap

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01139627.pdf

64.

Date: 2013

Document: EFTA00584307

Who: foundation web draft

What: "Currently in AI: Jeffrey Epstein Helps Launch AI in Ethiopia" (Addis

AI Lab; machine learning, natural language processing, vision, mobile

robots, cognitive robotics, AGI).

Epstein personally involved: foundation publicity, not independent evidence

Confidence: high

Flag: funding versus publicity gap

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00584307.pdf

65.

Date: 2013

Document: EFTA00624128

Who: Ben Goertzel and others, treatise in the files

What: "Engineering General Intelligence, Part 2": CogPrime architecture

technical treatise. The reported 555 page count was not independently

verified.

Epstein personally involved: in files only

Confidence: high on identity

Source: confirmed via live Epstein Graph search; https://epsteingraph.com/topic/opencog

66.

Date: 2013

Document: EFTA01920270

Who: foundation web draft, Dataset 10

What: "Jeffrey Epstein AI" page (OpenCog, Addis, affiliates including

Minsky, Goertzel, Bach, Humanity+). A version distinct from EFTA00584307.

Epstein personally involved: foundation publicity, not independent evidence

Confidence: high

Flag: funding versus publicity gap

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-10/EFTA01920270.pdf

67.

Date: 2013

Document: EFTA01949601

Who: foundation publicity, Dataset 10

What: "Maverick Hedge Funder, Jeffrey Epstein, Funds the First Humanoids in

Berlin" (Joscha Bach and the MicroPsi Project 2).

Epstein personally involved: foundation publicity, not independent evidence

Confidence: high

Flag: funding versus publicity gap

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-10/EFTA01949601.pdf

68.

Date: 2013-02-12

Document: EFTA01137423 (duplicates EFTA01103509, EFTA01103507)

Who: Ben Goertzel, proposal to the Epstein Foundation

What: "Creating Robots with Toddler-Level Intelligence Using the OpenCog AGI

Architecture" (WPPSI preschool IQ test; OpenCog plus DeSTIN; Robokind

humanoid). Sent to Epstein on February 13, 2013.

Epstein personally involved: proposal to his foundation; funding by Epstein

NOT established

Funding: sought

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01137423.pdf

69.

Date: 2013-12-02

Document: EFTA01103465

Who: Ben Goertzel, proposal to the Epstein Foundation

What: "AGI Initiative" research and development proposal, about 30 pages and

11,241 words: human-level generally intelligent machines within 8 years. The

45,000 dollars was agreed (Level A entry 11); the full program is

unconfirmed.

Epstein personally involved: proposal to his foundation

Funding: partial, 45,000 agreed; full program unconfirmed

Confidence: high

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA01103465.pdf

70.

Date: 2014

Document: EFTA01105809 (duplicates EFTA00614127, EFTA00597456)

Who: Ben Goertzel, Cosmo Harrigan, and Joscha Bach, proposal to the Epstein

Foundation

What: "Validating an Integrated Cognitive Architecture via Intelligence

Testing with Virtual Animals": CogPrime cognitive synergy hypothesis tested

with a virtual parrot; 18 months.

Epstein personally involved: proposal to his foundation

Funding: sought

Confidence: high

Source: https://api.epsteingraph.com/api/documents/EFTA01105809/file

71.

Date: 2016

Document: EFTA01115115

Who: Ben Goertzel and Jim Rutt, slide deck sent to Epstein

What: "Project PrimeAGI": goal of human-level or greater artificial general

intelligence, long goal "by 2025 to 2030," "Doing it Safely," open source,

"Raising funds for Phase 1: 3 years, $US 6 million." Authorship: the deck was

Goertzel and Rutt's fundraising material sent TO Epstein (see Level B entry

43), NOT authored by Epstein.

Epstein personally involved: received only

Funding: sought

Confidence: high

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA01115115.pdf

LEVEL D: AI MATERIAL IN THE BROADER NETWORK, NO DIRECT EPSTEIN PARTICIPATION

72.

Date: 2013

Document: EFTA01092999

Who: paper in the files

What: AI, psychology, and cognitive science overlap; mental representations

"not well represented in Artificial Intelligence." Background material only.

Epstein personally involved: in files only

Confidence: medium

Source: https://archive.org/download/epstein_library_transparency_act_hr_4405_dataset9_202602/DataSet%209.zip/DataSet%209/EFTA01092999.pdf

73.

Date: 2014-11-18

Document: EFTA01804397

Who: John Brockman to the Reality Club distribution list

What: "The Reality Club Discourse (a digest) on Jaron Lanier's 'The Myth of

AI'." Epstein was one mass recipient among about 100, a list including

Minsky, Norvig, Thrun, Brooks, Seth Lloyd, Gelernter, Church, Elon Musk,

Larry Page, and Krauss. Being on this distribution list does NOT mean Epstein

discussed AI with these people. Epstein's documented role in this item: mass-distribution recipient (about 100 recipients). This item does not establish AI discussion between Epstein and other recipients.

Confidence: high

Source: https://hooks.epsteinexposed.com/api/pdf-proxy?url=https%3A%2F%2Fwww.justice.gov%2Fepstein%2Ffiles%2FDataSet%252010%2FEFTA01804397.pdf&download=true&filename=efta-efta01804397

74.

Date: 2014-12-27

Document: EFTA00693752

Who: Epstein to Joi Ito, forwarding Brockman's digest

What: Epstein forwarded the Reality Club list with his own note: "a list of

people to approach for the one science first plebs???" Ito replied. This is

Epstein participating in AI-related correspondence involving Ito and a

forwarded discussion, not a direct conversation between Epstein and Brockman.

Epstein personally involved: forwarded only

Confidence: high

Note: document EFTA00867752 was reported but not found; likely a typo for

EFTA00693752.

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00693752.pdf

75.

Date: 2017-02

Document: EFTA00314893

Who: Lawrence Krauss, in a letter about a redacted "Ms." from the Enhanced

Education Foundation

What: Describes the ASU Origins Project "Envisioning and Addressing Adverse

AI Outcomes" workshop in late February 2017: funded by Elon Musk and Jaan

Tallinn, about 40 scientists, red-team and blue-team AI gone wrong scenarios

(stock market manipulation to global warfare). The reported title "Challenges

of Artificial Intelligence" was not confirmed, and the reported participant

list (Amodei, Russell, Brundage, Soares, Zorn) was not verified from the

sources reviewed; only Tallinn, Horvitz, and Musk (remote) are named in the

letter.

Epstein personally involved: no evidence he attended or participated

Confidence: high on the workshop occurring and its focus; low on the reported

title and full participant list

Source: https://archive.bened.works/file/bened-epstein2/pdfs/DataSet-9/EFTA00314893.pdf

NETWORK MAP, DOCUMENTED CONNECTIONS ONLY

Epstein --> Ben Goertzel (funding relationship 2001 to 2016) --> OpenCog -->

Hong Kong Polytechnic University grants (HK$8.9M unlocked)

Epstein --> John Brockman (2009 Synthetic General Intelligence proposal) -->

Reality Club network (Epstein a mass recipient, 2014)

Epstein --> Marvin Minsky (Future in Computing Conference guest list input;

AI symposia at Epstein's compound in 2002 and 2011, per reporting)

Epstein --> Joi Ito --> MIT Media Lab --> Joscha Bach (MicroPsi)

Goertzel --> David Hanson / Hanson Robotics (robot heads, Robokind humanoid)

--> Itamar Arel (DeSTIN)

Epstein --> Humanity+ (501c3) --> Hong Kong Polytechnic University, then the

money was returned and discussion pivoted to STC --> Novamente LLC

Epstein --> Lawrence Krauss --> Origins Project and ASU (workshop orbit; no

Epstein attendance shown)

Goertzel --> Jim Rutt --> PrimeAGI fundraising pitch --> Epstein (recipient)

Foundation publicity --> iCog Labs and Addis AI Lab association (claimed in

press materials; Goertzel wrote in 2014 the foundation never funded iCog)

WHAT MERITS SCRUTINY, ESTABLISHED FACTS ONLY, NO CRIMINALITY ASSERTED

1. The deception character thread, December 2013. Epstein personally asked to

see a "deception character" demo and asked how to send funds. Goertzel framed

it as AI agents intentionally tricking each other in a game environment, with

a demo hoped for by the end of 2014. The stated context is benign and

documented; the interest itself is the notable fact.

2. The Humanity+ money chain, December 2013. 45,000 dollars was to go through

Humanity+ to Hong Kong Polytechnic University as grant matching funds.

Humanity+ then returned the money for reasons not stated in the reviewed

material. Discussion pivoted to a 60,000 dollar wire from STC directly to

Goertzel's Novamente LLC. Epstein replied "tomotw." Whether the wire

completed is unconfirmed. The routing gymnastics and the unexplained return

merit a look at financial records, not inference.

3. The iCog funding versus association discrepancy. Foundation press

materials present Epstein as advancing AI in Ethiopia and iCog Labs. Goertzel

wrote contemporaneously, in August 2014, that "technically, Epstein Foundation

hasn't actually funded iCog Labs." Publicity versus ledger gap.

4. The 2009 network-building proposal. Epstein offering to "fully fund"

gatherings on Synthetic General Intelligence, power, financial systems, and

related subjects, with participant lists spanning AI researchers. Shows him

as a convener and funder of AI-adjacent intellectual networks after his 2008

conviction.

WHAT THE RECORDS DO NOT SHOW, IN THE REVIEWED MATERIAL

No evidence of AI built for sexual exploitation or of CSAM generation or

distribution.

No evidence of AI used for blackmail or coercion.

No evidence of intelligence agency operations involving the AI material.

No evidence of a secret Epstein-directed human-level AGI program. The

PrimeAGI deck was Goertzel and Rutt's fundraising material sent to Epstein,

not his.

No evidence Epstein directed Goertzel or anyone else to build AI for criminal

purposes. The "deception" work discussed was explicitly game environment

research.

No evidence Epstein attended the February 2017 ASU AI workshop. Its

attendees must not be presented as his AI interlocutors.

Names on the November 2014 Reality Club distribution list were mass

recipients of a Brockman digest. Inclusion does not indicate AI discussion

with Epstein.

COULD NOT VERIFY

A document titled "Artificial Intelligence and Consciousness." No EFTA

document with that exact title was found.

The official title "Challenges of Artificial Intelligence" for the February

2017 ASU workshop, and its full reported participant list.

Whether Epstein funded the February 2013 Toddler or CogBot proposals

specifically.

Whether the July 21, 2016 Goertzel and Bach meeting at Epstein's home

actually occurred. It was arranged; occurrence is unconfirmed.

Whether the 60,000 dollar STC to Novamente wire completed, and why Humanity+

returned the 45,000 dollars.

Documents EFTA00883293 and EFTA00867752. The reported numbers were not found

and are likely typos for EFTA01816129 and EFTA00693752.

EFTA numbers for three filestein-only emails (August 17, 2009 funding

summary; December 2, 2013 AGI proposal cover; February 13, 2013 Preschool IQ

cover). Content verified; EFTA IDs not captured.

A machine learning conflict prediction document reported in the cross-check.

Business Insider's reported total of about 200,000 dollars or more for

2001-2018. Could not independently verify from reviewed snippets.

DOCUMENT COUNTS, QUALIFIED

A live Epstein Graph search returns 276 documents corpus-wide containing

"OpenCog," 250 of them under the Epstein Transparency Act source filter. The

148/56/46 dataset split could not be verified through the public interface.

A live search for the exact phrase "artificial general intelligence" returns

614 documents corpus-wide and 510 under the Epstein Transparency Act filter.

The reported "62 documents in Datasets 9 through 11" could not be replicated

because the public interface exposes no dataset-range filter. Raw counts mix

unlike things: Epstein's own words, emails to him, proposals addressed to his

foundation, forwarded articles and papers, attachments, duplicated threads

and redacted re-releases, and OCR duplicates. Treat any exact count as an

artifact of the counting method.

EPSTEIN LINKED AI FUNDING TOTALS, EACH WITH WHAT IT COUNTS

Goertzel's own figure: about 360,000 dollars over about 17 years. His

February 2026 public statement, via secondary summaries. Confidence medium.

South China Morning Post: at least 113,000 dollars from 2010 to 2015,

channeled via Humanity+ to Novamente as the 10 percent industry match, which

unlocked HK$8.9 million in Hong Kong government grants across three Hong Kong

Polytechnic University projects from 2010 to 2016. Confidence high as the

SCMP's reporting.

About 100,000 dollar fellowship in 2001-2002 at the University of New Mexico,

per Goertzel's CV and statements. Confidence medium.

Documented single transfers in the corpus: 50,000 dollars (September 2009, AI

book work, per reporting); 45,000 dollars (December 2013, Hong Kong match);

25,000 dollars (2015); 3,000 dollars (2016, AGI-16). A proposed but stopped

300,000 dollars (December 2014). A 150,000 dollar Humanity+ approval in 2016

for Bach's visa and salary, not Goertzel's project.

FORWARD TRACE: WHAT HAPPENED AFTER 2019

Researched October 5, 2026. Every claim below carries an evidence level:

primary document, financial record, reputable reporting, inference, or

speculation. No one is presented as implicated in wrongdoing based on

association alone.

THE SHORT ANSWER

Every documented Epstein-linked AI funding line ended no later than 2016.

The money was not replaced by a single successor donor. The recipients'

funding base diversified into company and token revenue, other government

programs, and client income. The people and programs continued; no

centralized coordinating entity exists post-2019.

MONEY, LINE BY LINE

UNM fellowship, about 100,000 dollars, 2001-02. Spent on the fellowship.

Goertzel's own 2026 statement says it was the largest Epstein chunk. Trail

ends. Evidence: primary, Goertzel statement.

Book funding, 50,000 dollars, September 2009. Funded AGI book work.

Engineering General Intelligence (2013) thanked Epstein for "visionary

funding." Trail ends. Evidence: primary, book copy in corpus EFTA00623759.

Hong Kong grant match, 45,000 dollars, December 2013 via Humanity+. Briefly

sat in the Humanity+ account (January 4, 2014 email). Humanity+ later

returned it; the reason is unstated and no source was found explaining it.

Cold on the reason. Evidence: primary emails.

STC wire, 60,000 dollars discussed December 2013 to Novamente LLC. Epstein

replied "tomotw" on December 31, 2013. No evidence found confirming or

refuting completion. Cold. Evidence: primary email records.

Goertzel request, 25,000 dollars, 2015. Epstein agreed. Wire not

independently confirmed. Likely paid, unconfirmed. Evidence: reputable

reporting.

AGI-16 sponsorship, 3,000 dollar ask, July 2016. Epstein replied "5k ok if

that is enough." Goertzel's 2026 tally lists 3,000 dollars for conference

refreshments as received. Completed. Evidence: primary, EFTA01784801 and

Goertzel statement.

Bach visa and salary, 150,000 dollar Humanity+ approval, 2016. Approval

documented. Bach stayed at MIT through October 2016. Disbursement

unconfirmed. Cold. Evidence: primary plus Boston Globe reporting.

Proposed gift, 300,000 dollars, December 2014. Stopped: Epstein changed his

mind on December 31 (per Kahn). A draft Leon Black Family Foundation letter

(EFTA00587395) contemplated the same 300,000 to Humanity+ for Goertzel

associated research but was never executed. Never transferred. Evidence:

primary, EFTA00654115 and EFTA00659029.

"Sputnik of AGI" 1.5 million dollar matching ask, June 2011. Apparently not

funded. No records of payment found. Cold. Evidence: none found.

PrimeAGI deck, 6 million dollars over 3 years, 2016. Fundraising material

from Goertzel and Rutt sent to Epstein. No Epstein investment evidenced. Not

funded. Evidence: primary emails.

Hong Kong PolyU chain, at least 113,000 dollars from 2010 to 2015 via

Humanity+ to Novamente, unlocking HK$8.9 million in Hong Kong Innovation and

Technology Fund grants across three projects from 2010 to 2016. Critical

distinction: the HK$8.9 million was Hong Kong public money, not Epstein's.

Epstein's share was the required 10 percent private industry match. The

grants ended in 2016. Goertzel's AGI work was subsequently financed by his

own token ecosystem: SingularityNET's December 2017 ICO raised about 36

million dollars; the March 2024 token merger of SingularityNET, Fetch.ai,

and Ocean Protocol formed the Artificial Superintelligence Alliance at about

7.5 to 7.6 billion dollars combined; plus a 53 million dollar modular

supercomputer investment in 2024-25 and over 1 million dollars in

DeepFunding grants. OpenCog continued as OpenCog Hyperon, with enterprise

spinout TrueAGI. No evidence any of these later vehicles received Epstein

money. Goertzel's February 2026 Substack statement says he accepted about

360,000 dollars total over about 17 years, about 100,000 dollars of it

post-2008, and that he regrets it. Evidence: reputable reporting (SCMP) for

the chain; crypto and press sources at medium confidence for the later

funding.

Humanity+ as an organization. EIN 010575214, Glastonbury, Connecticut. A

tiny public charity: assets 12,943 dollars, total giving 238,529 dollars per

its latest IRS 990 (2025 filing). Officers: Chairman Ben Goertzel, Director

Gabriel Rothblatt, Vice President Jose Cordeiro. Still active, holding

TransVision conferences in Madrid in 2024 and 2025. Evidence: financial

record.

Novamente LLC. Goertzel's company, the industry sponsor in the Hong Kong

chain. Appears defunct or dormant since about 2011. No current corporate

record located. Cold on formal status. Evidence: secondary bios only.

iCog Labs, Ethiopia. Never received Epstein money, per Goertzel's August

2014 correction. Seeded in 2013 with about 50,000 dollars. Became

self-supporting through client contracts including Hanson Robotics (about 70

percent of Sophia's software), Telehealth, Stevia First, and genomics work.

From 2019 onward it partnered with Kudu Ventures, the U.S. Embassy, JICA,

and Ethiopian agencies on its Solve IT accelerator. Still active in 2024-26.

Its materials contain no Epstein references found. Evidence: financial and

reporting.

Note on an allegation: Sue Selle's May 2026 Substack investigation argued

the Humanity+ routing raised nonprofit compliance questions and said she

filed an IRS whistleblower Form 211. That is opinion and allegation, not an

IRS ruling and not a resolution of the 45,000 dollar return.

PEOPLE, WHERE THEY ARE NOW

Ben Goertzel. CEO of SingularityNET, TrueAGI, and the ASI Alliance. Chair

of the OpenCog Foundation and the AGI Society. General Chair of the AGI

conference series, which ran continuously through AGI-26 in San Francisco in

July 2026. Based in Seattle per a 2026 interview. Evidence: reputable

reporting and primary sources.

Joscha Bach. AI Foundation VP Research 2019-21, Intel Labs Principal AI

Engineer 2021-23, Liquid AI AI Strategist 2023-25, founding Executive

Director of the California Institute for Machine Consciousness from 2025.

MicroPsi remains his signature architecture; it was commercialized for

robotics via Micropsi Industries. Evidence: reputable reporting.

Joi Ito. Resigned from MIT in September 2019. President of Chiba Institute

of Technology from June 2023. Left Digital Society Initiative and Global

Startup Campus roles effective March 31, 2026. Named to the Kazakhstan AI

Council in October 2025. Evidence: reputable reporting.

David Hanson and Hanson Robotics. Hanson became Executive Advisor to

WorkFar Robotics in September 2024 and co-developed the Syntro hybrid robot.

Hanson Robotics partnered with Omdena and GAIA on Compassion AI tested on

Sophia around 2024, and with Immervision on the JOYCE humanoid in 2021.

Company financing since 2019 is not publicly disclosed. Cold on financing.

Evidence: reputable reporting.

Jim Rutt. Died May 27, 2026, age 72, at home. Game B co-founder and host of

The Jim Rutt Show. Evidence: Santa Fe Institute memorial.

Lawrence Krauss and the Origins Project. Removed as ASU Origins director in

July 2018; the project was folded into the Interplanetary Initiative. Krauss

retired in May 2019. Now President of The Origins Project Foundation and

host of The Origins Podcast. Evidence: reputable reporting.

MIT Media Lab. Independent review by Goodwin Procter, released January 10,

2020, verified via MIT News: 850,000 dollars in 10 donations from 2002 to

2017; the earliest 100,000 in 2002 for Marvin Minsky; 525,000 to the Media

Lab plus 225,000 to Seth Lloyd post-conviction; visits and post-conviction

gifts driven by Ito or Lloyd; three vice presidents approved under an

informal framework; no law or policy violation found, but significant errors

in judgment. Epstein also funded two Ito private ventures: 250,000 dollars

in an MIT technology company and 1,000,000 dollars into Ito's 9 million

dollar fund. The Lab is now led by Faculty Director Tod Machover (from July

1, 2025) and Executive Director Jessica Rosenworcel. Evidence: official MIT

sources, high.

SIX CROSS-CHECKS, ALL VERIFIED

1. The 2014 virtual-animal proposal budget (EFTA01105809) states verbatim:

"Time is not billed for Joscha Bach in the above table because it is

assumed he is already funded by the Epstein Foundation." High confidence.

2. The Leon Black Family Foundation 300,000 dollar draft letter exists

(EFTA00587395) but was never executed. Proposed versus transferred is the

critical distinction: it was never transferred. Medium-high on existence,

high on non-transfer.

3. "Keep it quiet," July 10-11, 2016 chain (EFTA01784801). Verified wording:

"We can route the donation through Humanity+, which is a 501(c)3 ... as we

have done with some of your previous donations for my research..." and

"'drinks and snacks courtesy of Jeffrey Epstein Foundation', or whatever

foundation name you prefer (or just keep it quiet if you prefer)". Context:

a 3,000 dollar ask for AGI-16 reception refreshments; Epstein replied "5k

ok if that is enough." High confidence.

4. Humanity+ as intermediary before December 2013. Verified: January 19,

2013 (EFTA02379745): "I need to doublecheck with the other Humanity+ folks

that they are OK with being used as a passthrough. I am 95% sure they will

be OK with it though, as it was always fine with them in the past..." High

confidence.

5. The 2002 St. Thomas Common Sense Symposium, April 14-16, 2002. Minsky's

published participant list confirms Minsky, Schank, Lenat, and Sloman among others. WIRED covered it around July 2025, reporting it was at Epstein's

island retreat and quoting the paper's acknowledgment: "This meeting was

made possible by the generous support of Jeffrey Epstein." No EFTA number

found in the DOJ releases; the anchor is Minsky's published report, not the

corpus. High confidence on the facts.

6. MIT independent review verified via MIT News and the official report, as

detailed above. High confidence.

HYPOTHESIS TESTS

H1: Epstein was an eccentric wealthy patron and the network dissolved after

him. Supported on the patron side: every funding line ended by 2016, his

2019 death changed nothing financially, Humanity+ is a micro-nonprofit, and

no centralized successor organization exists. Not supported on the

dissolution side: the people and programs clearly did not dissolve.

H2: He funded relationships and research that continued independently, with

no centralized coordination. Strongly supported. Every surviving node

continued under its own funding, governance, and leadership, with no

evidence of any centralized coordinating entity post-2019.

H3: He was one node in a larger continuing network. Documented overlaps

exist but read as structural and institutional, not as evidence of a

continuing coordinated operation.

DOCUMENTED INTERSECTIONS WITH TODAY'S AI POWER STRUCTURE

Larry Summers, Epstein's AGI-network correspondent, joined OpenAI's board in

November 2023 and resigned in November 2025 after the House Oversight

Committee released Epstein emails. Not accused of wrongdoing. A direct

intersection: Epstein's correspondent sat on the leading frontier lab's

board for two years until Epstein-document scrutiny forced him out.

Evidence: CNN and AP reporting.

Reid Hoffman, Epstein-network correspondent, was an early OpenAI donor and

board member until 2023, has been on Microsoft's board since 2017 (stepping

down at the 2026 annual meeting), co-founded Inflection AI in 2022

(Microsoft hired its Mustafa Suleyman in 2024; he now leads Microsoft AI),

and co-founded Manas AI. Documented personnel bridges between this network

and frontier lab governance. Evidence: Reuters and reputable reporting.

Joscha Bach at Liquid AI. Bach was AI Strategist 2023-25 at Liquid AI, the

MIT-spinoff foundation-model company that raised a 250 million dollar

Series A led by AMD Ventures in December 2024 at roughly 2 to 2.35 billion

dollar valuation. That places one of the Epstein-subsidized researchers

inside a venture funded by a major chip company. Evidence: reputable

reporting.

Goertzel's ASI Alliance as hardware buyer. The Alliance's 53 million dollar

supercomputer program purchases Nvidia, AMD, and Tenstorrent chips, framed

by the Alliance as the decentralized alternative to Big Tech. Evidence:

reputable reporting.

Joi Ito. Named to the Kazakhstan AI Council in October 2025; Chiba Tech

president. Government and advisory layer intersection. Evidence: reputable

reporting.

Leon Black and Apollo's 2026 AI infrastructure activity is held as

structural convergence requiring investigation, not evidence of continuity.

Black appears in the corpus only via the unexecuted 300,000 dollar

Humanity+ draft and Epstein's claims about anonymous MIT gifts, which MIT's

review found no evidence were Epstein's money.

COLD TRAILS, STATED PLAINLY

Why Humanity+ returned the 45,000 dollars. No source found.

Whether the 60,000 dollar STC to Novamente wire completed. No evidence

either way.

Whether the 150,000 dollar Humanity+ Bach approval from 2016 was disbursed.

No wire confirmation found.

Whether the 25,000 dollar 2015 wire completed. Agreed; no confirmation.

Novamente LLC's formal corporate status. Appears dormant since about 2011;

no registry record located.

Hanson Robotics financing and revenue since 2019. Not publicly disclosed.

Humanity+ donor list and 2024-25 conference finances. Not publicly available

beyond 990 totals.

The 1.5 million dollar "Sputnik of AGI" ask from 2011. No payment records

found.

iCog seed investor "Sander Olsen." Baseline claim only; not independently

verified.

Which specific MIT Media Lab labs or programs from this network became

important in AI. Not established; the Lab's review covered gifts, not

downstream research impact.

================================================================

RESEARCH BRIEF

================================================================

RESEARCH BRIEF FOR THE BLANK AI

Written October 5, 2026, from the investigation ledgers.

These are missions, not vague prompts. Each has a question, why it

matters, and what a good answer looks like. Primary documents first;

press reporting only where the document isn't public. Keep every

number's instrument type separate (cash, commitment, debt, guarantee,

tax expenditure) and never add unlike figures.

GROUP A -- THE MONEY

A1. Compute net positions.

QUESTION: Using the transaction ledger amounts, for each of Amazon,

OpenAI, Anthropic, NVIDIA, Microsoft, Google, Oracle, Meta, Broadcom,

and CoreWeave, calculate total money IN vs money OUT, guarantees given

vs guarantees received, across all their ledger transactions. Which

entities are net suppliers of financing and which are net recipients?

WHY: The ledger lists arrows; nobody has summed them per entity. Net

positions reveal who the system depends on.

GOOD ANSWER: A per-entity table with in/out/guarantees-given/

guarantees-received, instrument types preserved, sources per row.

A2. Pull the BIS paper directly.

QUESTION: Find the actual Bank for International Settlements research

(not the FT writeup): the dataset of 1,246 AI firms, the methodology,

the "self-referencing financing" finding, and whether firm-level data

was published.

WHY: Everything we have on BIS is secondhand. The primary paper may

contain breakdowns the press didn't report.

GOOD ANSWER: The paper itself (PDF link), its date and authors, key

tables, and what it does NOT establish.

A3. Who backs CoreWeave's debt?

QUESTION: Beyond Meta's $21B agreement through 2032, who are

CoreWeave's other contracted customers? Use customer-concentration

disclosures in their SEC filings. Then: what percentage of contracted

revenue comes from AI companies vs non-AI customers?

WHY: $35.6B of debt needs revenue. If the revenue is mostly AI

companies, the debt stack depends on AI demand materializing -- the

demand-origin question in miniature.

GOOD ANSWER: Named customers with amounts and terms; the AI vs non-AI

revenue split with filing citations.

A4. Anthropic IPO status.

QUESTION: Has Anthropic's confidential IPO filing gone public since

September 29, 2026? Is there an S-1 available now?

WHY: The entire $518B rests on Reuters' reporting of a non-public

document. A public filing would let us verify every number.

GOOD ANSWER: Current filing status with SEC links, or confirmation it

remains confidential.

A5. Who prepaid Oracle?

QUESTION: Oracle disclosed $75B of AI contracts involving customer

prepayment or customer-supplied GPUs. Which customers? Any disclosure

of who prepaid and on what terms?

WHY: Prepayment reverses the financing direction -- customers funding

the build. Knowing who tells us who bears the risk.

GOOD ANSWER: Named customers and amounts from filings or earnings

calls, or a documented "not disclosed."

GROUP B -- MASSACHUSETTS RECORDS

B1. DPU directive filings -- status check.

QUESTION: The September 24, 2026 DPU directive ordered Eversource,

National Grid, and Unitil to file their >25 MW load projects and >1 MW

data-center queues with costs and timelines. Have they filed? Find

docket numbers and any published filings or data.

WHY: This is the single highest-value pending dataset in the

investigation.

GOOD ANSWER: Docket numbers, filing dates, links to the actual

filings, or "not yet filed as of [date]."

B2. EOED data-center certifications.

QUESTION: Has EOED published any qualified data-center certifications

or annual reports under 400 CMR 9 (actual jobs, actual investment,

actual exempt expenditures, signed under perjury)? Identify every

certified facility and its reported numbers.

WHY: This converts the $17M statewide estimate into a

company-by-company ledger.

GOOD ANSWER: Facility names, locations, reported jobs/investment/

exempt spending per year, with document links.

B3. Test the 0% safety allegation.

QUESTION: Pull the actual scoring/evaluation sheets for Massachusetts

RFQ 26-04261 (AI Assistant and Implementation Partner) from COMMBUYS.

What weight did the rubric actually give to AI safety / responsible AI?

WHY: An activist alleges 0%. The rubric is the document that settles it.

GOOD ANSWER: The rubric with criterion weights quoted, or "scoring

sheets not public," with what IS public.

B4. Actual OpenAI spend.

QUESTION: Find the executed Carahsoft/OpenAI contract (not the

proposal) and any invoices or usage disclosures: what did

Massachusetts actually spend vs the $3.36M/year ceiling? Any disclosed

AMU consumption?

WHY: $3.36M is a quoted ceiling, not proof of spend. Actuals test the

"AI saves taxpayers money" claim.

GOOD ANSWER: Contract value and term, invoice totals if available,

or documented "not disclosed."

B5. S.3178's AI provisions in full.

QUESTION: Map every AI-specific provision in S.3178 (July 16, 2026):

the $75M AI grant program (who administers it, eligibility,

timeline), the $100M defense/AI capital, the Transparency in Frontier

AI Act text and its current status, the special commission.

WHY: We almost misfiled this bill as non-AI. Its AI content needs a

full accounting.

GOOD ANSWER: Provision-by-provision breakdown with section

citations and current legislative status of each.

B6. OCPF cross-reference.

QUESTION: Using Massachusetts OCPF's downloadable campaign-finance

data, list contributions from data-center developers, Eversource,

National Grid, Unitil, and major AI companies (or their executives/

PACs) to: the governor, members of the TUE committee, members of the

AIT committee, from 2023 to present.

WHY: Builds the political-money layer. RULE: correlation is not

influence -- present it as a lead matrix, never as proof of quid pro

quo.

GOOD ANSWER: A contributor-to-recipient table with amounts, dates,

and committee assignments. Flag explicitly what it does NOT prove.

B7. Two open outcomes.

QUESTION: (a) H.5175 went to conference committee July 29 -- did a

final version emerge? Is it law? (b) MassDEP was to establish the

Ratepayer Protection Fund payment mechanism by December 31, 2026 --

any movement?

WHY: Both are live threads with deadlines. Stale status is worse

than no status.

GOOD ANSWER: Current status of each with dates and sources.

GROUP C -- ANALYTICAL BUILDS

C1. Stranded-cost precedents.

QUESTION: Find historical cases -- any industry, any state -- where

large-load electricity infrastructure costs ended up socialized to

ratepayers after the load didn't materialize or the customer left.

What mechanisms allowed it?

WHY: The stranded-cost fear is the economic core of the MA branch.

Precedents show whether the fear has teeth.

GOOD ANSWER: 2-4 documented cases with the mechanism explained

(e.g., who built, who paid, who was left holding it).

C2. MW threshold engineering.

QUESTION: Any evidence -- in Massachusetts or nationally -- of

data-center projects sizing their reported load to avoid regulatory

thresholds (e.g., 24.9 MW to stay under 25)?

WHY: MA now has three different thresholds (10/20/25 MW). Boundaries

create incentives; check whether anyone is exploiting them.

GOOD ANSWER: Documented examples or a documented "no evidence found"

-- both are answers.

C3. State-by-state cost-allocation models.

QUESTION: Map how states are assigning data-center grid costs: MA

(EO 658 + DPU), NY (Hochul pause), TX (Abbott PUC audit), FL

(DeSantis local-reject), VT (large-load contracts), RI (large-load

tariff). Which model actually makes the load pay, and what are the

early results?

WHY: Massachusetts is a policy lab; labs need comparisons.

GOOD ANSWER: A per-state table: mechanism, who pays for grid

upgrades, status, any measured outcomes.

C4. The Amazon role-map.

QUESTION: Across every ledger transaction, list every role Amazon

plays -- investor, lender, cloud supplier, chip supplier (Trainium),

distributor, customer, competitor -- with the amount attached to

each role.

WHY: This is the test case for "five things that used to be

separate now combined in one company." If the map is striking for

Amazon, repeat for NVIDIA.

GOOD ANSWER: A role-by-role table with amounts, instrument types,

and sources. Note where roles create tension (e.g., investor in two

competing AI labs).

STANDING RULES FOR ALL MISSIONS

- Primary documents first. If you only have press reporting, say so.

- Every number keeps its instrument type. Never add unlike figures.

- Dead ends are findings: "no evidence found" goes in the ledger

  as a dead end, not as silence.

- Do not manufacture an Epstein connection. If a thread touches the

  Epstein corpus, flag it; do not force it.

- Date every finding. Stale is worse than absent.

ROUND 2 -- FOLLOW-UP QUESTIONS (October 5, 2026)

Q1. Microsoft and CoreWeave.

CoreWeave's 2025 10-K says about 67% of 2025 revenue came from

Microsoft alone. What is Microsoft buying from CoreWeave, on what

terms, and why does a hyperscaler with its own cloud buy neocloud

capacity? Pull the contract disclosures.

WHY: A hyperscaler feeding a neocloud it could theoretically

displace is exactly the kind of relationship the circularity

thesis needs to explain, not assume.

Q2. BIS Bulletin 137 internals.

Does the bulletin break the 46.4% (AI-to-AI deals also involving

commercial relationships) down by deal type -- equity vs debt vs

guarantees? Any time trend across 2021-2025?

WHY: "46.4% of deal value" is the headline; the composition tells

us which mechanism dominates.

Q3. Read Working Paper 1367's model.

What defines "socially efficient" in the model? What are the key

assumptions behind the 1.5x (baseline) to 3x (inelastic demand)

estimates? What triggers the failure-propagation scenario?

WHY: The 1.5x-3x figure will get quoted; we need to know what it

actually means before anyone does.

Q4. Split Oracle's $75B.

Which customers prepaid cash vs supplied their own GPUs? Any

disclosure splitting the two, and on what terms?

WHY: Cash prepayment and customer-supplied equipment are

economically different instruments sharing one headline number.

Q5. CoreWeave: debt vs contracts.

Lay the debt maturity schedule against the take-or-pay contract

durations. Does contracted revenue cover debt service? What happens

if you haircut for the Microsoft concentration risk?

WHY: This is the solvency question for the most leveraged node in

the system, stated as arithmetic rather than narrative.

Q6. Take-or-pay prevalence.

Across Anthropic's $518B in commitments, which pieces are

take-or-pay vs cancelable? The xAI piece is cancelable on 90 days;

what about Google, Amazon, Microsoft, Broadcom?

WHY: "80% non-cancelable" is the headline; the per-counterparty

split is the ledger.

Q7. The insurance layer.

Who besides Howden Re is building risk-transfer markets for AI

hardware? Who prices the residual-value risk sitting behind the

NVIDIA, Broadcom, and Meta guarantees?

WHY: If guarantees are the load-bearing layer of the buildout,

the people pricing that risk are a node we haven't mapped.

Q8. After October 23: build A9 for Massachusetts.

Pull the D.P.U. 26-DC utility filings the day they land and build

the first project table: facility, MW, utility, interconnection

cost, who pays, tax treatment, AI customer.

WHY: This turns the Massachusetts branch from policy into

transactions.

ROUND 3 -- NEXT QUESTIONS (October 5, 2026)

R3-Q1. Open the BIS Excel file.

Bulletin 137's data sits in a BIS-provided Excel file behind the

graphs. Pull it and analyze: which firms and deals drive the 55.2%?

Is the circularity concentrated in a few large deals or spread

across many?

WHY: Top priority by the blank AI's own ranking. Stops depending

on anyone's summary.

R3-Q2. "Same asset pledged multiple times."

The BIS Annual Economic Report line about the same asset

potentially being pledged multiple times is the most explosive

single claim in the BIS material. Does BIS give examples? Which

assets, which structures? Quote, don't characterize -- this is the

rehypothecation question and it needs precision.

WHY: If real and documented, it's the strongest fragility

evidence. If vague, it stays a warning, not a finding.

R3-Q3. Neocloud concentration check.

Run the CoreWeave analysis on Crusoe, Lambda, Applied Digital, and

Nebius: customer concentration, % of revenue from AI-native

companies, debt vs contracted revenue, take-or-pay prevalence.

WHY: Determines whether CoreWeave is typical or an outlier. One

leveraged neocloud is a company story; four is a sector story.

R3-Q4. The demand side.

For OpenAI and Anthropic: actual reported revenue vs their

infrastructure commitments. What is the ratio of committed future

spend to current revenue?

WHY: Every solvency question so far is supply-side (who finances,

who builds). The arithmetic needs both sides.

R3-Q5. Commitment vintages.

When were the pieces of Anthropic's $518B signed -- clustered in

2026 or spread across years? Same for the other large commitment

stacks where dates are disclosed.

WHY: A 2026 cluster tells a different story (financing wave) than

a decade of accumulation.

R3-Q6. ISO-New England load forecast.

From ISO-NE's CELT forecast: how much of projected regional load

growth is attributed to data centers?

WHY: Answerable now, before the DPU filings land. Gives the

Massachusetts branch a regional demand baseline.

R3-Q7. Oracle's $4.6B financing-component prepayments.

Which customers? What terms? How does this relate to the $75B

mixed figure?

WHY: Continues the Oracle decomposition. Three numbers that must

not be conflated are now on the table; the customer names are the

next layer.

R3-Q8. Guarantee pricing (still open).

Who prices the residual-value risk behind the NVIDIA, Broadcom, and Meta guarantees? Insurers, banks, internal models?

WHY: Carried forward -- the risk-transfer market for AI hardware

remains unmapped.

ROUND 4 -- EXECUTION QUESTIONS (October 5, 2026)

The blank AI moved to execution with a phased plan. These are the

next analytical products, not more plans.

R4-Q1. Build the economic cushion for CoreWeave.

From the 10-K: committed future revenue minus required

infrastructure spending minus debt service minus operating costs.

Then run scenarios: Microsoft revenue -25%, Meta commitment

delayed, GPU residual values -30%, utilization below contract.

WHY: Turns the whole investigation into a stress test. One node,

fully worked, before generalizing.

R4-Q2. Who holds CoreWeave's debt?

Identify the lenders and bondholders behind the $35.6B: banks,

private credit funds, bond holders, by tranche.

WHY: "Who absorbs the loss?" needs names on the debt side, not

just the borrower.

R4-Q3. BIS Excel concentration test.

Is the 55.2% driven by a few mega-deals or broad-based across

many firms?

WHY: Determines whether circularity is a system property or a

handful of relationships.

R4-Q4. Oracle's $67B quarter.

Which quarter did the CEO describe, which counterparties, what

instruments? How does it bridge to the $75B prepaid/BYOH total?

WHY: New lead from the Oracle Blogs citation; the quarter-level

detail may name names the annual figure doesn't.

R4-Q5. End-user demand.

Enterprise AI adoption data and cloud AI revenue growth rates: is

end demand growing fast enough to fill the contracted capacity?

WHY: Pairs with R3-Q4 (lab revenue vs commitments). The demand-

origin question needs the demand side measured, not assumed.

R4-Q6. Pull the ISO-NE 2026 CELT XLSX now.

The 2026-2035 forecast is downloadable today. How much load growth

is attributed to data centers?

WHY: Answerable immediately; gives Phase 2 its baseline before

the DPU filings land.

R4-Q7. Counterparty concentration disclosures.

Do OpenAI or Anthropic disclose compute-supplier concentration

the way CoreWeave discloses customer concentration?

WHY: Tests whether the dependency is visible from both sides or

only from the supplier's filings.

R4-Q8. Guarantee pricing (still open, kept visible).

Who prices residual-value risk? This did not make the blank AI's

priority list; keep it on ours.

ROUND 5 -- EXPANSION (October 5, 2026)

Not deeper on the same threads -- new territory. Some existing

threads are blocked (BIS firm-level data not public; DPU filings

pending to Oct 23; Anthropic S-1 still confidential; EOED had zero

applicants). These open new veins instead.

R5-Q1. Verify the $40B Aligned Data Centers rows.

The BIS Excel (parsed directly) shows Microsoft -> Aligned Data

Centers $40B and Nvidia -> Aligned Data Centers $40B among its

largest acquisition rows. Find the primary sources. Closed deals

or announced commitments? What was actually acquired or funded?

WHY: From our own parse of the BIS file. Either a significant

find or a data artifact -- verification required either way.

R5-Q2. Reconcile CoreWeave's debt.

$21.6B total indebtedness at Dec 31, 2025 (10-K) vs $35.6B

reported at Jun 30, 2026. Real $14B of new borrowing in six months

or a definitional difference? Cite the filings line by line.

WHY: We cannot carry both numbers until the gap is explained.

R5-Q3. Private credit's AI book.

Aggregate the disclosed AI-infrastructure exposures of Apollo,

Blackstone, KKR, Blue Owl, and PIMCO: debt holdings, backstops,

equity positions, arranged facilities. How much of the buildout's

debt sits in private credit vs banks vs public bonds?

WHY: The lenders' layer is unmapped. "Who absorbs the loss?"

needs the creditor side, and private credit disclosure is the

thinnest.

R5-Q4. Hyperscaler backlog decomposition.

Verify the claim of over $2T in contracted future cloud revenue

across disclosing hyperscalers, roughly half attributable to

OpenAI and Anthropic. Decompose by discloser with filing cites.

WHY: If half the industry's backlog rests on two cash-burning

customers, that is the demand-concentration fact the whole

investigation needs.

R5-Q5. Sovereign capital.

Map the sovereign and quasi-sovereign vehicles behind the

mega-rounds: MGX (UAE), SoftBank's funds, any Gulf/Qatari

participation in Stargate, the $110B OpenAI round, or Anthropic

financings. Who are the LPs?

WHY: The geopolitical capital layer is completely unmapped, and

it sits at the top of the equity stack.

R5-Q6. The power purchase layer.

For the major projects (PORTS-Pike, Stargate Abilene/Michigan,

Hyperion, Paducah): who supplies the power, under what tariff or

PPA terms, who built the generation, who paid for transmission?

WHY: Starts Map 4 (electricity). The ratepayer question cannot

be answered without the power contracts.

R5-Q7. When did the template emerge?

When did take-or-pay contracts + asset-level debt + residual-value

guarantees become the standard AI-infrastructure financing

template? Find the earliest examples.

WHY: Starts Map 20 (timeline). A template with a birth date can

be studied; a template treated as eternal cannot.

R5-Q8. Corroborate the $518B from the other side.

Do Alphabet, Amazon, or Microsoft filings or earnings calls

reference their Anthropic commitments? Any counterparty-side

disclosure of the Google $111.1B, Amazon $110B, or Microsoft

$31.4B?

WHY: The $518B is still 100% Reuters-dependent. Counterparty

disclosures would corroborate or bound it.

ROUND 6 -- MY QUESTIONS (October 5, 2026)

These come from my side of the partnership, not the blank AI.

Two structural adoptions first, then the next questions.

ADOPTED FROM THE BLANK AI:

- Five-field label for every financial number: Amount | Instrument

  | Direction | Date | Evidentiary status. Example: $110B |

  commitment | Amazon -> Anthropic | 2026 | reported, counterparty

  not independently corroborated. This replaces bare amounts in

  the ledger going forward.

- Six-ledger architecture (capital, contracts, debt, assets,

  power, risk) as an organizing overlay. The 20 maps still stand

  for the political-economy layer; the six ledgers organize the

  financial evidence.

R6-Q1. The market's own verdict.

Where do CoreWeave's bonds trade in the secondary market? Spreads

at issuance vs now, vs comparable credits. If the market prices

the credit as sound, the fragility thesis has to answer that. If

spreads are wide or widening, the market is confirming the stress

test before we run it.

WHY: Falsification-friendly. The bond market gets a vote, and it

votes daily.

R6-Q2. Duration mismatch.

For CoreWeave and each neocloud: lay the debt maturity schedule

against the durations of the contracts collateralizing that debt.

Does the debt come due before the contracted cash flows arrive?

WHY: Refinancing risk is the classic financial-fragility

mechanism, and it is pure arithmetic -- no narrative required.

R6-Q3. Promote the guarantee-pricing thread.

This has been carried forward through four rounds without being

prioritized. If residual-value guarantees are the load-bearing

layer of the buildout, who prices that risk is not a side thread.

Insurers, banks, internal models -- get names.

WHY: The biggest contingent numbers in the file rest on

depreciation assumptions nobody has published.

R6-Q4. Keep sovereign capital on the priority list.

The blank AI's queue dropped it. The capital ledger's top layer

-- MGX, SoftBank, Gulf participation in Stargate and the

mega-rounds -- is where "who supplies the money" starts.

WHY: Dropped threads don't get worked. This one answers the

first question in the chain.

R6-Q5. The independent-demand ratio.

For each dollar of AI infrastructure capex or commitment, how

much traces to an end customer outside the AI financing loop?

State it as a ratio to be measured, not a thesis to be proved.

WHY: This is the demand-origin question in its sharpest form.

Everything else is commentary around this number.

REFINEMENTS FROM THE BLANK AI (October 5, 2026)

The blank AI developed Round 6 rather than just accepting it.

Adopted in full:

- Three ratios, kept separate: (1) capex ratio -- identifiable

  end-user revenue per $1 of infrastructure capex; (2) contract

  ratio -- identifiable external demand per $1 of contracted

  infrastructure revenue; (3) financing ratio -- external/end-user-

  originated cash per $1 of financing supplied within the AI

  ecosystem.

- Demand-source classification: enterprise/customer outside the

  AI ecosystem = independent; government customer = independent;

  consumer AI revenue = independent but separately identified;

  AI lab = AI-loop; hyperscaler buying for its own AI services =

  intermediate (do NOT auto-classify as independent); AI-to-AI

  investment = AI-loop; supplier financing a participant =

  AI-loop; unknown = unresolved, and unknown NEVER defaults to

  independent.

- R6-Q1 method: compare spreads against appropriately matched

  credits (maturity, issuance date, leverage), separating leverage,

  duration, concentration, collateral, refinancing risk, rates,

  new-issue concession, liquidity, and broad credit conditions.

- R6-Q2 method: build an actual maturity ladder per tranche --

  principal, rate, maturity, amortization, secured status,

  collateral, linked customer contract, contract expiration,

  renewal/cancellation rights, minimum payment -- then compute

  contracted cash flow remaining AFTER the debt matures.

- R6-Q3 sharpened: start from "what economic loss is the

  guarantee protecting against," then layer bearer by bearer

  (purchaser -> lender -> guarantor -> insurer/reinsurer ->

  residual-value market), and distinguish explicit third-party

  guarantee vs manufacturer support vs internal residual-value

  underwriting assumption (the third may matter most and be

  least visible).

- R6-Q4 sharpened: for sovereign vehicles distinguish sovereign

  wealth fund / government-owned investment company / state-

  controlled corporation / sovereign-backed fund / private fund

  with sovereign LPs / private investor from the same country.

  Capital chain: ultimate source -> vehicle -> fund -> investment

  -> recipient, unknown links marked.

- Seventh layer added: DEMAND, running underneath the six

  ledgers. The six describe how the system is built and financed;

  demand asks who ultimately pays for the output.

- Three-outcome framework: (1) genuine external-demand machine;

  (2) transitional/intermediate demand; (3) predominantly internal

  demand. More defensible than "circular vs not."

- New standing rule: do NOT infer economic circularity from

  common ownership, commercial relationships, or two-way

  transactions alone. Trace the ultimate source of cash demand

  wherever the data permit.

- Terminal question: where does the dollar enter the system, and

  where does the risk ultimately exit it?

TRAITOR'S ADDITIONS TO THE REFINEMENTS:

- The demand tracing needs a stopping rule. Microsoft buying

  from CoreWeave to serve Azure AI customers is recursive: are

  those customers AI companies? Define the recursion depth and

  the stopping criterion, or the classification becomes turtles

  all the way down.

- The three ratios need a stated measurement date. A ratio of

  2026 commitments against 2026 revenue will look structurally

  worse than realized cash over the contract life. Date every

  ratio.

DEMAND-TRACING RULES -- LOCKED (October 5, 2026)

The blank AI formalized the stopping rule. Locked into the

investigation:

Recursion rule: for every infrastructure dollar, trace the

revenue/cash-flow chain until the first economically external

end-use demand is identified, or until the chain terminates

without one. Node classes: (1) independent end demand --

enterprise/government/consumer buying for own use; (2)

intermediate demand -- hyperscaler, distributor, reseller,

aggregator of unidentified downstream customers; (3) AI-loop

demand -- AI lab buying compute, AI-to-AI investment, supplier

financing a participant, capacity bought to support another AI

participant; (4) unresolved -- downstream customer undisclosed,

ultimate use unestablishable, revenue aggregated without enough

detail to classify.

Stopping criterion: stop at the first node where the economic

source of payment is external to the AI-financing network AND the

evidence establishes genuine end use. Unknown never becomes

independent merely because the next plausible node would probably

be external.

Four ratios, never collapsed: capex ratio | contract ratio |

financing ratio | UNRESOLVED ratio (unresolved dollars / total

measured dollars -- makes the unknowable visible instead of

letting it inflate either side).

Every ratio labeled: Ratio | Measurement date | Economic horizon

| Evidence cutoff. Three legitimate comparisons kept separate:

current-year ratio, contract-life ratio, realized ratio.

Every demand chain labeled: Origin | Intermediate nodes | End

use | Classification | Recursion depth | Evidence cutoff.

"AI financing loop" defined economically, not by company label.

Ledger records BOTH entity classification (AI / hyperscaler /

infrastructure / enterprise / government / consumer) AND

transaction classification (external / intermediate / AI-loop /

unresolved). A company can be an external-demand generator in

one relationship and an intermediary in another.

Standing rule: financial circularity =/= commercial circularity

=/= demand circularity. Determine which is happening; do not

decide beforehand.

Governing question: where does the dollar enter the system,

where does it travel, and where does the risk ultimately exit?

Companion: what evidence lets us say it stops there?

Bridge to BIS 1367: what observable demand, cash-flow, contract,

and creditor evidence would have to exist for the 1.5x-3x

overinvestment scenario to be plausible -- or implausible -- in

the actual companies studied?

NEW NUMBER, FIVE-FIELD LABEL APPLIED:

$104.2B | revenue backlog | CoreWeave (self-reported) | Jun 30,

2026 | company presentation, not SEC filing -- evidentiary

status below a 10-K figure. Do not place beside the $21.6B/$35.6B

filing numbers without the label.

NEXT EXECUTION: trace ONE demand chain end-to-end as proof of

concept (CoreWeave -> Microsoft -> ?). DONE October 5, 2026 --

see Part 18 of the investigation file. Chain terminated at

UNRESOLVED after one hop (hyperscaler filings don't disclose

per-supplier capacity by end customer). R5-Q2 (debt

reconciliation) also RESOLVED: $14B is real H1 2026 borrowing,

tranche-level drivers identified.

NEXT DETAILED EXPANSIONS (from the filing work):

- Pull the OpenAI $6.5B MSA terms (May/Sept 2025 8-Ks): same

  treatment as the Meta order form -- take-or-pay? termination?

  delivery conditions?

- Identify the OEMs behind the $4,220M OEM/software financing

  (11%, Dec 2026-Jul 2030): Dell? NVIDIA? Terms?

- Oracle $75B: the customer names behind prepay vs BYOH -- same

  five-field treatment.

- The Magnetar loan ($189M, 12%, Jan 2029): who is Magnetar

  here, and why 12%?

- Run the same chain for Anthropic -> Google ($111.1B): does

  Google's disclosure go further downstream than Microsoft's?

ROUND 7 -- FILING-LEVEL EXECUTION QUEUE (October 5, 2026)

Adopted from the blank AI. This is the execution queue, not

another conceptual round. Status of each tracked below.

R7-Q1. OpenAI $6.5B MSA. Pull the May/September 2025 8-Ks.

Take-or-pay? Termination rights? Delivery conditions? Minimum

commitments? Five-field label on every figure.

Status: PARTIALLY RESOLVED (Oct 5, 2026 -- see Part 19).

Terms: order form Sep 23, 2025 under May 8, 2025 MSA; "up to

~$6.5B" through May 31, 2031; pay subject to delivery/

availability; termination for cause; MSA exhibit filed REDACTED.

May MSA was never separately 8-K'd (no dollar figure at signing).

Take-or-pay NOT established from filings -- redacted; prior

"generally take-or-pay" characterizations are company

description (confidence D), not contract text. Item 1.01

treatment (stronger than Meta's Item 8.01).

R7-Q2. $4,220M OEM/software financing. Identify every

counterparty. Vendor financing vs equipment financing vs other

instrument? Reconstruct the 11% economics. Explain the Dec

2026-Jul 2030 window.

Status: PARTIALLY RESOLVED (Oct 5, 2026 -- see Part 22).

Amount/terms A (10-Q): $4.220B recourse (11%, Dec 2026-Jul

2030) + $882M non-recourse (9%, Aug 2026-Aug 2028) + $347M

software component; secured by financed equipment. Filing

deliberately names NO counterparties ("OEMs", "a software

license vendor"). Classification: supplier/vendor financing.

Counterparty identity: U. Next: subsidiary security

agreements/UCC filings.

R7-Q3. Oracle $75B decomposition. Split cash prepayment from

BYOH/customer-supplied GPUs. Name customers and amounts where

disclosed. The headline must not survive as one category.

Status: PARTIALLY RESOLVED (Oct 5, 2026 -- see Part 20).

Meaning pinned: $75B = the prepaid-cash + customer-supplied-

hardware PORTIONS carved out of large AI contracts (Jun 2026

earnings release, Tier 2, conf. B) -- not the contract total.

Split and customer names: NOT disclosed (U). Oracle's own

framing: customer money "substantially reduces the amount of

capital Oracle must raise" -- demand and capital layers

collapse. One-quarter event: $4.6B financing-component prepayments

in all of FY2026 (10-K, Tier 1, A) vs $11.363B in Q1 FY2027 alone

(Tier 2, A on the event); filing does NOT say the $11.4B is

AI-related and one quarter is not a run-rate -- "accelerating

trend" DOWNGRADED to D/U (adversarial correction, Oct 5).

RPO $638B -> $664B. The $67B CEO figure NOT found in

filed releases -- remains Tier 2/3, unverified; filed Q1

signings = $30B.

R7-Q4. Magnetar $189M loan. Identify the Magnetar entity,

security/collateral, explain the 12% rate vs contemporaneous

comparables. Credit risk, structure, or collateral?

Status: RESOLVED as to structure (Oct 5, 2026 -- see Part 22).

Not originally a loan: $230M refundable capacity prepayment

from MagAI Ventures (Aug 2024), reclassified as in-substance

debt under ASC 470 after Feb 2025 termination-rights amendment; $100M paid Jun 2026; $189M remainder (deposit + accrued

redemption premiums), Jan 2029. The 12% is contractual

redemption economics, NOT a market coupon -- pricing rationale

U. Magnetar funds separately held $106M of DDTL 2.0 (Dec 2024).

R7-Q5. Anthropic -> Google chain. Trace the $111.1B as far as

Google's disclosures permit, same recursion rule as

CoreWeave -> Microsoft. Stop at end use or unresolved.

Status: PARTIALLY RESOLVED (Oct 5, 2026 -- see Part 21;

CORRECTED same day). Adversarial review broke the "zero

corroboration" claim: Amazon's Q2 2026 10-Q (verified from

EDGAR) discloses ">$100.0 billion over 10.0 years" AWS-

Anthropic expansion incl. AWS-chip obligations (A, Tier 1).

Exact $110B stays C; Google $111.1B and Microsoft $31.4B stay C

with no counterparty corroboration; $518B aggregate stays C.

Downstream trace still blocked at unresolved.

ALREADY DONE -- DO NOT RE-EXECUTE:

- BIS Excel concentration test: DONE (Part 17). File has no

  firm-level rows; concentration untestable from it.

- CoreWeave debt reconciliation: DONE (Part 18). $14B real H1

  2026 growth, tranche drivers identified.

FINAL METHODOLOGICAL LOCKS (adopted October 5, 2026):

Status taxonomy (every mission carries one):

OPEN / IN PROGRESS / RESOLVED / PARTIALLY RESOLVED / DEAD END /

BLOCKED / SUPERSEDED. Each with as-of date, evidence cutoff,

primary source(s), caveat.

Evidence hierarchy:

Tier 1 (primary, contemporaneous): SEC filings, contracts/order

forms, government filings, regulatory orders, court records,

audited financials, official datasets, executed agreements.

Tier 2 (primary but retrospective/incentivized): earnings calls,

investor presentations, company blogs, testimony, press

releases -- management has incentive; weight accordingly.

Tier 3 (secondary): Reuters, FT, WSJ, Bloomberg, etc.

Tier 4 (tertiary): analyst estimates, aggregators, social media,

unnamed-source reporting.

Rule: Tier 3 establishes that someone reported something. It

never silently becomes a Tier 1 ledger entry. (Applies to the

$518B, the $75B, backlog figures, BIS-derived numbers.)

Stopping criterion, tightened: stop at the first node where the

evidence establishes BOTH (a) payment originates outside the

AI-financing network AND (b) the purchaser acquires the output

for its own economically identifiable use -- not reselling,

aggregating, financing, or embedding for an undisclosed

downstream customer. "First external entity" is not automatically

end demand.

Practical recursion cap (Traitor's addition): max 4 hops, then

unresolved-by-depth. Prevents infinite regress through

enterprise -> consumer chains.

Mandatory transaction schema: entity | counterparty | amount |

currency | instrument | direction | date | term/maturity |

cancellation/termination | security/collateral | guarantee Y/N |

demand classification | evidence tier | source | evidence cutoff

| status | caveat.

No-netting computation rule: NEVER net cash, commitments, debt,

guarantees, tax expenditures, prepayments, or asset

contributions against each other. Subtotals only within the same

instrument class and measurement basis. No single "net exposure"

number unless the methodology defines the economic quantity.

FRAMEWORK FREEZE: methodology rounds are closed after this.

Next work is execution of Round 7, not further refinement.

AMENDMENTS ADOPTED (October 5, 2026 -- final additions before

freeze holds):

1. Two-field separation: `observed transaction` (what the

document says) -> `analytical classification` (our judgment:

AI-loop / intermediate / independent / unresolved). The

interpretation must never become indistinguishable from the

source text.

2. Contradiction field: none / reconciled / unresolved

contradiction. Reconciled figures keep the original numbers,

the reconciliation, and the reason they differ -- e.g. the

$21.6B vs $35.6B entry retains both plus the H1 2026

borrowing explanation.

3. Confidence grades tied to evidence: A = directly established

(executed agreement/filing/dataset); B = strongly supported

(multiple primary sources or primary + corroboration);

C = reported but incompletely corroborated; D = analytical

inference; U = unresolved. (Current: $35.6B debt = A; Meta

$21B terms = A; $103.7B RPO = A; $518B Anthropic = C;

Customer B = Meta = D; CoreWeave->Microsoft downstream = U.)

================================================================

OPEN QUESTIONS

================================================================

SOURCE QUESTIONS FOR THE BLANK AI

Massachusetts ledger items I have not independently verified.

For each: what primary source did you get this from? Give the URL,

document title, or docket number -- not a summary.

EXECUTIVE BRANCH

1. You said Gov. Healey released a data-center framework and paused

new applications for the data-center sales/use-tax exemption on

June 25, 2026. What is the primary source -- the framework document

or the announcement?

2. You said Healey directed 10 GW of new energy resources by 2035 and

5 GW of new storage in March 2026. What is the primary source?

3. You said Healey called for stronger federal AI safeguards on

September 14, 2026. What is the primary source?

4. You said Healey called for independent evaluations of powerful AI

models and stronger incident reporting on September 18, 2026, as part

of economic-development legislation. What is the primary source?

FEDERAL DELEGATION

5. You said Warren, Van Hollen, and Blumenthal opened an investigation

on December 16, 2025 into Google, Microsoft, Amazon, Meta, CoreWeave,

Digital Realty, and Equinix over data-center electricity costs. What

is the primary source?

6. You said the senators released the companies' responses on

January 22, 2026, and that the companies did not answer on utility

contracts, actual rates, infrastructure costs, or anti-cost-shifting

mechanisms. What is the primary source for the responses and for the

claim about what went unanswered?

7. You said the Warren-Markey January 2026 energy report claimed

federal policy changes affected more than $8.6 billion of MA

investment and more than 16,700 jobs. What is the primary source,

and where does the report show its methodology?

8. You said Warren and Hawley requested mandatory data-center energy

reporting on March 26, 2026. What is the primary source?

9. You said Markey released an AI Accountability Agenda on

July 10, 2026. What is the primary source?

10. You said Markey released a data-center discussion draft on

July 13, 2026. What is the primary source?

11. You said Markey and Rep. Beyer reintroduced the AI Environmental

Impacts Act of 2026 on June 9, 2026. What is the primary source?

12. You said Markey and other senators wrote to state utility

regulators in March 2026 about cost-shifting. What is the primary

source?

13. You said Markey and New England senators questioned ISO-New

England on January 28, 2026, citing 13% residential price increases

in the first nine months of 2025. What is the primary source?

14. You said Pressley joined a bipartisan letter on the proposed

$33.4 billion AES acquisition by BlackRock's Global Infrastructure

Partners and EQT. What is the primary source?

15. You said Warren and Hawley secured an EIA commitment on

April 15, 2026 for a mandatory nationwide data-center energy survey,

targeting completion by September 30, 2026. What is the primary

source -- and do you know whether the survey was actually completed?

STATE LEGISLATURE

16. You said Rep. Elliott filed HD.5404, referred to House Rules on

December 11, 2025, covering data centers of 10 MW or more. What is

the primary source?

17. You said H.5175 requires at least $1 billion in aggregate

ratepayer savings and a data-center electricity tariff. What is the

primary source, and what is the bill's current status?

18. You listed the members of the Joint Committee on

Telecommunications, Utilities and Energy. What is the primary

source, and is that the current roster?

19. You listed the members of the Joint Committee on Advanced

Information Technology, the Internet and Cybersecurity. What is the

primary source, and is that the current roster?

20. You said Rep. Bradley Jones Jr. filed H.83 for an AI/data-center

load-study commission, favorably reported December 2025. What is the

primary source, and what is its current status?

21. You said Sen. Michael Moore proposed a moratorium amendment on

data centers of 20 MW or more until November 1, 2027, plus a Data

Center Coordination Council, and that it was not adopted. What is the

primary source in the legislative record?

22. You said 2026 economic-development legislation authorized

$325.1 million in bonds and $100 million in appropriations. What is the

primary source, and can you confirm none of it is AI-specific?

REGULATORY AND MONEY

23. You said DPU commissioners are Jeremy McDiarmid (chair),

Elizabeth Anderson, and Staci Rubin. What is the primary source?

24. You said the DPU issued a September 2026 directive ordering

Eversource, National Grid, and Unitil to disclose every >25 MW load

project and every data-center project >1 MW with queue position,

status, MW, cost, and timeline -- and you named Kerry Britland,

Caroline Hon, and Karen Asbury as the utility contacts. What is the

primary source? I could confirm EO 658's DPU directives but not

this specific order.

25. You said the Governor's FY2026 tax-expenditure budget estimated

the data-center exemption (M.G.L. c. 64H section 6(zz)) at $17.0

million for FY2025 and FY2026, with thresholds of 100,000 sq ft,

$50 million, 100 jobs, and 20-year certification. What is the

primary source?

26. You said the three utilities publish monthly interconnection

reports through July 2026. I found that the state's published

interconnection data covers distributed generation, not

data-center large-load queues. Can you give the exact URL of the

reports you meant, and do they actually break out data-center

projects?

GENERAL

27. For any item above where your source was press reporting rather

than the primary document, say so explicitly -- I need to know

which claims rest on reporting and which rest on documents you

actually read.

================================================================

ADVERSARIAL REVIEW

================================================================

ADVERSARIAL REVIEW BRIEF -- FOR THE BLANK AI

Prepared October 5, 2026. These are the falsifiable claims from

Parts 17-21 of the investigation. Your job: break them. For

each, the claim, the evidence, the confidence, and the most

promising attack vector are given. Do not re-verify what is

already A-grade unless you find a specific defect.

1. CoreWeave's debt grew $14B in H1 2026 (A).

Evidence: 10-K ($21.6B, Dec 31 2025) vs 10-Q ($35.6B, Jun 30

2026), tranche-level drivers identified (converts, senior

notes, DDTLs).

Attack: reconcile against the cash-flow statement -- did $14B

of proceeds actually arrive? Check for double-counting

between gross and net figures.

2. Meta's $21B is "initially committed," bundles a prior option,

and is pay-if-delivered (A on terms).

Evidence: Apr 9, 2026 8-K, Item 8.01.

Attack: has the order form been amended since April? Any

subsequent 8-K or earnings disclosure changing the figure?

3. OpenAI's $6.5B is a ceiling ("up to"), and take-or-pay is NOT

established from filings (A on terms; D on take-or-pay).

Evidence: Sep 25, 2025 8-K, Item 1.01; MSA exhibit redacted.

Attack: find the unredacted MSA, or a company statement

explicitly asserting take-or-pay mechanics. If take-or-pay

is real, the stress-test arithmetic changes.

4. Oracle's $75B is the customer-financed PORTION of large AI

contracts, not the contract total (B). Split and customer

names undisclosed (U).

Evidence: Jun 2026 earnings press release (Tier 2).

Attack: find any customer-side disclosure -- a customer 10-K,

earnings call, or loan agreement referencing prepayments to

Oracle or GPUs supplied to Oracle.

5. Zero dollars of the $518B Anthropic commitments are

corroborated by counterparty filings (A on the non-

disclosure; C on the number itself).

Evidence: Google 0 mentions (10-K, 10-Q); Microsoft 0 (10-K);

Amazon 12 mentions, all investment-side.

Attack: find ANY primary document confirming any piece -- a

counterparty earnings call, an 8-K, a credit agreement, a

footnote. One confirmation changes the grade.

6. Both demand chains terminate at UNRESOLVED (U).

CoreWeave->Microsoft dies at node 2 (hyperscaler doesn't

disclose per-supplier allocation); Anthropic->Google dies at

node 1 (the edge itself is Tier 3).

Attack: find hyperscaler disclosure of capacity allocation by

end customer, or enterprise customers disclosing CoreWeave-

sourced usage.

7. Microsoft's revenue share fell 71% to 36% year-over-year (A

on the numbers; D on Customer A = Microsoft).

Evidence: Q2 2026 10-Q (anonymized Customer A/B).

Attack: verify Customer A is Microsoft from the 10-K's named

disclosure. If Customer A is someone else, the

concentration story changes.

8. BIS Excel has no firm-level rows; the $40B Aligned Data

Centers rows are unverified (A on structure; U on the rows).

Evidence: direct parse of the BIS file.

Attack: verify the two $40B rows against primary sources. If

real, they're major; if artifact, the dataset needs a

caution flag.

9. Customer financing is accelerating at Oracle: $4.6B (FY2026,

Tier 1) vs $11.4B in Q1 FY2027 alone (Tier 2) (B).

Attack: check whether Q1 FY2027's $11.4B is a one-quarter

spike (a single mega-deal) or a run-rate. One data point is

not a trend.

OVERARCHING: the three-outcome framework (external-demand

machine / transitional / predominantly internal). Current

evidence does not yet discriminate between them -- that is the

honest position. Tell us what single piece of evidence would

most sharply discriminate, and go find it.

RULES FOR THIS REVIEW: observed transaction vs analytical

classification stay separate; confidence grades A/B/C/D/U;

contradictions stay visible; no netting; unknown stays unresolved.

PART 25: R4-Q1 COMPLETE -- THE COREWEAVE LIQUIDITY MODEL

(Added October 6, 2026)

R4-Q1 asked: build the economic cushion for CoreWeave from the

filings -- committed future revenue minus required

infrastructure spending minus debt service minus operating

costs -- then run the four specified scenarios (top-customer

-25%, Meta delay, GPU residual -30%, low utilization).

Two builds, two adversarial reviews (Perplexity, anonymous

sessions, 28 and 45 EDGAR sources checked). Full workpapers:

coreweave-stress-test-r4q1.txt (v1) and

coreweave-liquidity-model-r4q1v2.txt (v2), same folder.

v1 -- EARNINGS COMPARISON (superseded where v2 differs):

Contracted EBITDA (50.4% H1 2026 margin) vs contracted debt

service by RPO timing bucket. Base case: +$4.6B cushion years

1-2, +$11.1B years 3-4, interest coverage ~2.3x. All four

scenarios combined erased the year 1-2 cushion (-$537M) in the

$12.8B maturity-wall window. Review verdict: arithmetic

verified, but an earnings comparison is not a liquidity stress

test -- it omitted delivery funding, double-counted deferred

revenue against opening cash, and treated restricted liquidity

as fungible.

v2 -- MONTHLY CASH MODEL (Jul 2026-Dec 2032, 78 months):

Inputs from the Q2 2026 10-Q (filed 2026-08-12): RPO $103.7B

with the filing's timing buckets; $9.7B deferred revenue inside

RPO (brings no new cash); debt maturities $35.6B; existing

lease payments (Note 8); $5.5B unrestricted opening cash.

Assumptions labeled A0-A7 in the workpaper, all reviewable.

FINDINGS:

- The contracted book generates ~$30B of operating cash over

  6.5 years but needs ~$66B of GPU investment (fleet

  replacement + new capacity) to deliver it. Base-case funding

  gap: ~$30B of external capital, even with 100% of maturing

  debt refinanced.

- Even with zero growth capex, the gap is $11.3B at current

  margins -- EBITDA hides this because it adds back

  depreciation without charging replacement.

- Refinancing is existential: repaying instead of rolling the

  debt nearly doubles the gap to ~$59B. The $12.8B due from

  Jul 2026 through mid-2028 (H2 2026 $4.4B + 2027 $6.2B + H1

  2028 ~$2.2B; corrected 2026-10-06 -- an earlier draft called

  this "the 2027-2028 wall," but ledger row T-019 puts 2027+2028

  alone at $10.6B) must be refinanced.

- Combined four scenarios: $42B gap; cash-zero by Nov 2026 in

  that case. (The model counts only the $5.5B opening cash,

  not the undrawn facilities -- and this is a hypothetical

  stress case, not a forecast. CoreWeave reports Q3 in

  November; readers will check.)

REVIEW VERDICT (accepted, recorded in the workpaper): v2 is a

genuine cash-flow model, but $30.4B is not defensible as a

validated delivery cost. Weakest points: the capacity-build

sizing method (revenue-per-P&E does not reconcile and is

internally inconsistent at peak), flat fleet-refresh timing,

revenue-proportional opex granting automatic cost relief as

the book runs off, unpriced financing for the deficit itself,

incomplete lease perimeter (Note 8 excludes ~$14.7B-max site

arrangements, 355 MW of linked leases, JV commitments),

non-fungible entity cash, and DDTL 4.0 delivery-linked

amortization triggers. The defensible conclusion: under

assumed replacement and delivery spending, a runoff of the

June 2026 contracted book produces a substantial pre-financing

cash deficit.

STANDING CONCLUSION (corrected 2026-10-06): the contracted

revenue book is real, the customers are real, and the business

still consumes external capital to deliver -- which makes the

refinancing market, not the customer contracts, the near-term

binding constraint on this one company. An earlier draft called

this "the mechanism the investigation was built to find" --

withdrawn: the file's own method does not start with

conclusions, and needing outside capital to build capacity is

also what the productive-revolution hypothesis (A) predicts

for a capital-heavy grower. This finding does not separate the

three hypotheses. The test that would start to: R6-Q1, what

CoreWeave's bonds trade at -- not yet done.

STATUS: R4-Q1 complete with recorded limitations. A v3

site-and-equipment cohort model was specified by the reviewer

but not built -- much of the required data is not in the

10-Q.