“Describe in detail the extent of circular funding in the AI and AI adjacent space - detailing low medium and high risks of economic fallout in the case of such a funding bubble collapsing”
The map below covers 40 deduplicated circular or quasi-circular funding relationships among 22 entities — model developers, hyperscalers, chipmakers, neoclouds and financiers — plus 24 quantified fallout indicators split across low, medium and high scenarios, drawn from 2025–2026 disclosures across 15 source hosts for relationships and 9 for fallout. The shortlist is illustrative, not exhaustive: it was selected from several hundred candidate disclosures. Dollar figures are mixed instruments — equity, debt, warrants, guarantees, credits and commercial commitments — and are deliberately never summed into one total.
Hover a line to focus it; click a company or a line to pin its relationships here, with sources. Click empty space to clear.
A vendor or hyperscaler invests in a customer that buys its chips or cloud: Microsoft–OpenAI ($13B in, at least $250B of Azure back), Amazon–OpenAI ($50B in, $100B of AWS back), Google–Anthropic (up to $40B in, a $200B chip-and-server pact back). The capital provider’s revenue and the customer’s spending are the same dollars seen from two sides.
Warrants, revenue shares, cloud credits, distribution rights and capacity backstops tied to purchases: AMD’s performance warrants to Meta and OpenAI, Microsoft’s guaranteed 20% cut of OpenAI revenue until 2030, Nvidia’s guarantee to purchase CoreWeave’s unsold capacity. No round-trip cash is required for the economics to be entangled.
Data-center debt, leases, private credit, GPU-backed borrowing, project finance and power commitments transmit losses without any literal circular flow: hyperscaler bond issuance topping $100B, a projected $1 trillion of new debt to finance AI investment, 12-year power and facility contracts underwriting chips whose effective useful life may be as short as one year.
Commitments support valuations and financing capacity; higher valuations enable more capex; capex becomes supplier revenue; supplier strength enables more strategic investment. Reported demand, vendor revenue, customer capacity, valuations and collateral values become mutually reinforcing — and therefore correlated on the downside.
Model valuations fall, weaker startups fail, some GPU and cloud reservations are canceled, chip and data-center equities reprice. Hyperscalers absorb write-downs from cash flow; capacity is repriced and redeployed. Losses concentrate in venture funds, employees, suppliers and a few local construction and power markets. This holds if debt stays ring-fenced, contracts are enforceable, demand keeps growing at a slower rate, and the major platforms remain liquid.
Several major model developers miss purchase or lease commitments at once; neoclouds and project vehicles refinance poorly; GPUs depreciate faster than underwriting assumed; capacity and power contracts strand; private-credit, bond, equipment-finance and utility counterparties take losses. Hyperscalers and chipmakers cut capex, hitting construction, electrical equipment, energy infrastructure and data-center regions, while repricing tightens financing across adjacent tech.
Frontier AI returns disappoint while large commitments prove uneconomic; disputes and defaults reveal that apparently diversified revenue was concentrated in a small counterparty network. Collateral falls as GPUs obsolesce and specialized facilities resist repurposing; capex drops sharply, private credit gates, tech wealth effects hit consumption, layoffs rise, power projects cancel. Recession or a financial-stability event requires leverage, maturity mismatch and bank/insurer/pension exposure substantially larger than currently disclosed.
The greatest danger is not merely double-counting investment announcements. It is correlated counterparty risk and timing mismatch: long-lived debt, leases, land, power and buildings are being justified by customers whose economics depend on rapidly depreciating chips, falling model prices, continued outside funding, and future revenue that may arrive later than contractual payments come due.
Circularity can accelerate infrastructure coordination and is not inherently improper — none of these structures are, by themselves, evidence of fraud. The issues are transparency, additionality of end demand, leverage, concentration and loss allocation. The shock absorbers are real: cash-rich hyperscaler balance sheets, phased and contingent commitments, continued secular demand for compute, and the possibility of repurposing some power and data-center assets.
Announced commitments, funded cash and realized revenue are different things; instruments in these columns are not comparable and are never summed. One relationship (SoftBank–OpenAI / Stargate) carries no row-level URL and is shown unlinked.
| From | To | Financing / investment | Return commitment | Instruments | Date | Source |
|---|
Method: 64 report-ready rows extracted 2025–2026 — 40 deduplicated circular or quasi-circular relationship clusters (15 distinct source hosts) and 24 quantified fallout indicators across low/medium/high tiers (9 hosts); selected from several hundred candidate disclosures, so the set is illustrative rather than exhaustive. Figures are announced amounts as reported by each source; equity, debt, warrants, guarantees, credits and commercial commitments are distinct instruments and no grand total is computed. Some clusters rest primarily on one party’s announcement. Long findings and caveats are truncated in the table for space; follow each source link for full terms.