The Hook
Nvidia commits up to $100 billion into OpenAI, structured around OpenAI's commitment to purchase at least 10 gigawatts of Nvidia compute capacity. OpenAI commits to a reported $300 billion five-year Oracle cloud deal from 2027, using 4.5 gigawatts of capacity. Oracle raises its 2026 capex guidance to approximately $50 billion. CoreWeave finances its GPU fleet with approximately $28 billion of combined debt and equity, including a $7.6 billion facility secured by the very Nvidia GPUs it purchases. Then: AMD signs a deal with OpenAI for 6 gigawatts of capacity, plus a warrant giving OpenAI the right to acquire up to 160 million AMD shares at $0.01 per share against purchase and price milestones.
The Question
When the same capital flows from a chip vendor to a model lab and back to the chip vendor's own revenue line, is that demand — or is it accounting?
The Paper Trail
Each transaction in the loop has a disclosure anchor: Nvidia's announced investment in OpenAI (September 2025), the Oracle cloud deal (reported by multiple outlets), Oracle's raised capex guidance (earnings call), CoreWeave's debt structure (disclosed alongside its March 2026 financing commitments), and the AMD-OpenAI warrant arrangement (AMD 8-K, October 6, 2025). The GPU-collateralized CoreWeave facility is the most structurally unusual: the collateral is the asset being purchased, which is the Nvidia product, whose demand the Nvidia investment in OpenAI was designed to generate.
The Synthesis
On a quarterly income statement, circular demand from a captive customer looks identical to organic demand from a new one. The difference only materializes if external, end-user AI revenue grows fast enough to replace the circular spending. The Sequoia framing — the current AI infrastructure spend requires approximately $600 billion in annual end-user revenue to generate a conventional return — is the test the loop has to pass. Whether it passes is not in the filings yet.