Circular AI flows · Fraudtracker

Circular AI capital flows — the Russian-doll problem

A chip vendor invests in a model lab, the lab commits to buy compute from a cloud, the cloud buys chips from the vendor. Revenue at one node is funded by capital from another node it just invested in. This page maps the public money flows so the recursion is visible.

Drag to pan, scroll to zoom, click a node for details. Edge thickness ≈ approximate annual $ flow. Edge colors: blue = equity investment, green = compute / chip revenue, amber = multi-year contractual commitment.

The most striking recursive loops

Nvidia ↔ OpenAI ↔ Microsoft ↔ Nvidia

Microsoft has invested ~$13B+ in OpenAI. OpenAI spends most of that on Azure compute, which Microsoft books as cloud revenue. Microsoft uses that revenue to buy Nvidia GPUs (~$10B+/year). Nvidia, in turn, has committed up to $100B to OpenAI partly as equity, partly as preferential pricing on H100/B200/Blackwell — meaning Nvidia is funding a customer whose spending it will then book as revenue. Each dollar can theoretically be counted multiple times across the three companies' GAAP P&Ls depending on when it changes form.

Nvidia → CoreWeave → Nvidia

Nvidia owns ~6% of CoreWeave (post-IPO, dilution adjusted). CoreWeave's primary operating expense is buying Nvidia GPUs — they had over $15B of Nvidia commitments on the books at IPO. CoreWeave's revenue comes from Microsoft (~$10B contract), OpenAI ($11.9B contract), and others — who are themselves funded by Nvidia's partner ecosystem. The H100 → CoreWeave → Microsoft → OpenAI → Nvidia pipeline is roughly closed.

AMD ↔ OpenAI

In Oct 2024, AMD and OpenAI announced a compute deal: OpenAI commits to buy ~6 GW of AMD MI300/MI400 chips (~$10B+), and in exchange AMD issues OpenAI warrants for up to 10% of AMD itself, vesting on stock-price milestones. A vendor giving its customer 10% of the company contingent on its own stock rising — there's no analogue in the previous tech cycle.

Oracle → OpenAI ($300B over 5 years)

Oracle signed a $300B compute commitment with OpenAI — multiple times Oracle's total annual capex history. Oracle has to build the data centers, which means buying $50B+ of Nvidia GPUs, which Oracle finances through long-dated debt. OpenAI doesn't generate $60B/year in cash — the deal is bridged by future OpenAI fundraising rounds, themselves underwritten partly by Microsoft and Nvidia.

Amazon ↔ Anthropic, Google ↔ Anthropic

Amazon has invested $8B+ in Anthropic; the agreement designates AWS as Anthropic's primary cloud. Anthropic's compute spend on AWS in 2025 is reported in the multi-billions. Google separately invested ~$2B in Anthropic and supplies TPU compute. Anthropic, Microsoft-OpenAI, and the cloud landscape all share this circular shape.

Why this matters for fraud / earnings-quality screening

Revenue concentration that doesn't show up as concentration. Nvidia's 10-K segments revenue by customer geography and product. The fact that ~10–15% of Nvidia's data-center revenue depends on entities Nvidia itself funds is not disclosed cleanly in any single line. Standard customer-concentration screens miss it.
Round-trip transactions look like organic demand. If Nvidia gives OpenAI $10B of credit and OpenAI spends $10B on H100s, Nvidia books $10B revenue and a $10B receivable / equity stake. Cash never moved net, but GAAP shows revenue growth. This is the same mechanism that triggered the 2000–2002 telecom-fraud wave (Lucent, Nortel, Cisco vendor financing). DSRI and CFO − NI gap are the classic signatures.
Counterparty risk is correlated. If OpenAI's commercial trajectory disappoints, the $300B Oracle deal repricing flows back through Microsoft (write-down on the equity stake), through Nvidia (revenue cliff), through CoreWeave (utilization collapse), through Anthropic (peer multiple compression). The graph is fragile because it's nearly fully connected.
Goodwill and intangibles are bloating. Microsoft's OpenAI stake is carried at fair value with substantial unrealized gain. Nvidia's CoreWeave stake similarly. AMD's OpenAI warrants are derivatives. AQI and Goodwill/TA on these balance sheets are climbing for reasons that may not survive a downturn.
The screener can quantify but not interpret. Beneish would likely score Nvidia FY26 around −2.0 (still clean) on standard ratios — because the lie, if there is one, is in the *attribution* of revenue growth, not its mechanical reconciliation to cash. This is structurally similar to Wirecard: the numbers tie out, but the underlying economics may not.

Historical analogues

Era	Pattern	How it ended
1999–2001	Telecom vendor financing (Nortel, Lucent, JDSU lent to CLECs that bought their gear).	$2T equity wipe; multiple bankruptcies; SEC fraud cases.
2006–2008	CDS/CDO daisy chains (banks insuring each other's structured products).	Lehman; AIG bailout; $700B TARP.
2017–2022	Crypto exchange / token treasury reflexivity (FTX/Alameda, Celsius, Three Arrows).	$2T+ market-cap collapse; FTX, Celsius, BlockFi bankruptcies.
2023–?	AI infrastructure capital recycling (this page).	Open question.

None of this proves wrongdoing. Many of these flows are economically rational on a per-deal basis. The pattern matters because past examples of similar shapes ended badly when the marginal participant lost confidence.