SEC filings — primary-source pass

Web-verified 2026-06-08. Structured + EDGAR links: spec-sec-filings-primary.json. Re-anchors the proven core to the companies' own mandatory disclosures instead of press. Filings corroborate the proofs; they don't change them. "As-reported" is itself contestable (non-GAAP add-backs, segment definitions) — the same scrutiny the data-integrity block applies to government numbers applies here.

CoreWeave 10-K (FY2025) — CIK 1769628 · EDGAR — STRONG

• Revenue ~$5.1B.
• Customer concentration: ~67% from Microsoft; backlog anchored to a few counterparties — OpenAI ~$22.4B, Meta ~$35.2B implied, Anthropic, Jane Street.
• Debt > $21B (from <$8B in 2024); large portion variable-rate ~11%; Q4-2025 interest $388M ≈ ⅓ of revenue.
• Net loss ~$1.2B; 43 sites, >850 MW; continued dependence on financing + stable power.
• Risk factors explicitly flag few-customer dependence, contract renegotiation/cancellation, and customers building in-house compute.

This is the cascade node the TLA+ trace runs through: a neocloud that is ~⅔ one customer with ~$21B of variable-rate debt eating ⅓ of revenue. The 10-K's own concentration + leverage language is the OpenAI → CoreWeave → Oracle fragility, in legalese.

NVIDIA 10-K (FY2025) — CIK 1045810 · EDGAR — STRONG

• Customer concentration disclosed: Direct Customer A ~12%, B ~11%, C ~11% (each ≥10%).
• Subsequent: top-4 ≈ 61% of Q3-FY2026 sales; concentration grew ~36% → ~61% in a year.
• Risk factor: "a significant amount of our revenue stems from a limited number of partners and distributors… revenue could be adversely affected if we lose… any of these end customers."

NVIDIA's filing confirms the demand side of the loop is a handful of buyers — the same hyperscalers/neoclouds it invests in. The 36%→61% tightening is the SCC tightening, and it corroborates the vendor-financing self-funding metric (~10% funded / ~56% headline).

Microsoft 10-Q (FY2026 Q1, qtr ended 2025-09-30) — CIK 789019 · EDGAR — STRONG

• ~27% of OpenAI (as-converted), equity method; $13B committed, $11.6B funded as of Sep 30 2025.
• OpenAI's losses flow through "other income (expense), net" — Q1-FY2026 other income showed a ~$683M loss, "primarily… net recognized losses on equity method investments, including OpenAI."
• External reporting infers Microsoft's share implies an OpenAI quarterly loss ≈ $11.5B+.

Primary confirmation of the Z3 result (OpenAI needs ≥$1.03T external; insolvent at zero inflow): Microsoft books its 27% share of OpenAI's losses, and the implied ~$11.5B/qtr burn is the cash the loop must keep raising. Reconciles the $13B-cash vs ~$135B-marked gap.

Oracle 10-Q / 8-K (FY2026, qtr ended 2025-11-30) — CIK 1341439 · EDGAR — STRONG

• RPO ~$523.3B (Nov 30 2025; +~359% YoY in a prior quarter) — a backlog dominated by OpenAI/Stargate cloud capacity.
• OCI revenue guided on a hockey-stick: $18B → $32B → $73B → $114B → $144B.
• ≥$72B of data-center partner debt for Stargate: $16.3B Michigan (largest single-facility tech debt ever), ~$38B Texas/Wisconsin, ~$18B New Mexico.
• PIMCO anchored ~$10B of the Michigan bond after US banks retreated.
• FY2026 capex ~$50B (>2× prior year); $18B of public bonds sold in a single day (Sep 2025) to fund OpenAI data centers.

The clearest primary picture of the loop's funding mechanics: a $523B backlog that is largely one circular counterparty (OpenAI/Stargate), financed by ≥$72B partner debt + $50B capex + $18B bonds — with private credit (PIMCO) the marginal lender when banks pull back. This is "solvent only while capital keeps flowing," in Oracle's own filings.

Amazon (CIK 1018724) & Alphabet (CIK 1652044) — STRONG

• Amazon: ~$200B 2025 capex (mostly AI); booked a ~$9.5B pre-tax gain in Q3-2025 from marking up its Anthropic stake after Anthropic's $13B raise at a $183B valuation; committed up to a further $25B (Apr 2026).
• Alphabet: 2026 capex guided $180–190B (raised from $175–185B), 2027 to "significantly increase"; also a fair-value Anthropic stake (billions in gains) and ~$100B SpaceX equity.

Key accounting finding. The same circular AI stakes distort hyperscaler earnings in opposite directions by method. Fair value (Amazon, Google on Anthropic): a higher private round = a gain booked to profit — Amazon's Q3-2025 profit got a $9.5B lift from revaluing a company it is simultaneously funding. Equity method (Microsoft on OpenAI): the investor books its share of losses (~$11.5B/qtr implied). So the loop flatters Amazon/Google reported profit and drags Microsoft's — and the Anthropic IPO will finally test those private marks against a public price. Treat hyperscaler "AI gains" as partly self-referential paper marks.

NVIDIA 13F-HR (Q4 2025) — CIK 1045810 — STRONG

• ~11% of CoreWeave (~47.2M shares, ~$3.66B; added ~$2B in Q4); Intel ~$7.9B (~50% of the 13F portfolio — funding a rival/foundry); Nebius ~$100M; plus Synopsys, Nokia, Coherent; exited Applied Digital, Arm, Recursion, WeRide.

The vendor-financing loop in NVIDIA's own filing: it owns ~11% of CoreWeave — a neocloud that is ~67% Microsoft and exists to buy NVIDIA GPUs. The equity edges of the SCC, disclosed by the chipmaker itself.

Oil / Dubai note

The commodities block (macro-futures-vs-physical) now carries the Brent–Dubai EFS (the sour/Asia differential): it narrowed and flipped negative (−11¢, Aug 2025), expected to widen in 2026 on OPEC+ sour supply.

Pull queue (remaining)

The eventual Anthropic / OpenAI S-1 (will test the private marks) · SpaceX S-1 (summarized in fin-spacex-spcx) · full 13F text parse for MSFT/AMZN/GOOGL institutional cross-holdings.

Takeaway

The strongest corroboration of the circular-funding thesis is the firms' own SEC risk factors: CoreWeave and NVIDIA disclose the concentration and leverage the models formalize. The caveat is symmetrical with the data-integrity block — "as reported" deserves the same zero-trust reading as the government's headline numbers.

← Research index · structured data: spec-sec-filings-primary.json · spec-sec-filings-primary.md