Nvidia's circular AI deals show compute turning into a commodity
The essay opens by naming the pattern critics keep flagging: OpenAI raises money from Microsoft and spends it on Microsoft servers; Nvidia backstops CoreWeave's debt while CoreWeave buys Nvidia GPUs. Critics compare this to the circular investments of the 1999 dot-com era and predict a bubble. The author argues they are wrong, and that circular deals instead show AI compute maturing into a fungible commodity, priced and financed the way oil, copper or electricity are.
To ground the analogy, the piece walks through how commodity infrastructure gets built. A development project too large for a bank to fund alone, such as wells and a pipeline, gets a trading firm to guarantee it will buy all the output at a set price. That guaranteed revenue lets banks lend with confidence, and the trading firm sometimes takes an equity stake as well. This exact structure, the author notes, has been used by Japanese commodities traders and development banks since the 1960s and by Jamaica in the 1980s. A recent US example: MP Materials signed a 10-year deal with the Department of War committing the department to buy all the company's neodymium-praseodymium magnet output at a minimum of $110 per kilogram; MP Materials later struck a similar off-take agreement with General Motors, a pattern the essay says is common across the electric-vehicle supply chain.
Applied to AI, Nvidia GPUs have become a fungible commodity thanks to standardized infrastructure across the industry, so much so that a company that builds a data center and cannot use it can resell the capacity, sometimes at a premium. SpaceX and Meta already do this: both lease out data-center capacity built for their own frontier-model ambitions, with SpaceX leasing its Colossus 1 and 2 data centers for over $2 billion a month.
For startups and neoclouds that cannot yet afford their own infrastructure, Nvidia plays the role of the oil trading firm: it provides capital, prioritizes GPU access, and, per SemiAnalysis estimates cited in the piece, effectively guarantees to buy back compute the startup cannot use, sometimes alongside an equity stake. Nvidia is sitting on a $1 trillion order backlog. The essay names CoreWeave's $6.3 billion deal with Nvidia and smaller operator Firmus's $505 million deal as examples, and traces a chain of related moves: OpenAI cancels a deal with Oracle and Meta steps in; SpaceX leases servers to Google; Google invests in Anthropic.
The piece flags where this healthy pattern can tip into risk. Nvidia's cumulative commitments, reportedly up to $750 billion against the $6.3 billion CoreWeave case alone, raise the question of overextension. The whole model also depends on GPUs and data centers staying fungible and resellable; a shift in chip standards, or technology that makes small local clusters viable, could undercut the resale logic that backs Nvidia's guarantees.
The essay draws a sharper line around deals built to move obligations off a company's balance sheet. It cites a Financial Times breakdown of a TeraWulf bond in which Google backstops any lease-payment failures that cascade through Fluidstack and, ultimately, Anthropic, letting Google support the debt without it appearing on Google's own books. Meta's $27.3 billion Hyperion data-center bond sale and its separate $12.3 billion deal marketed by BlackRock are described as following the same off-balance-sheet structure. Bulls argue these bonds are ultimately safe because hyperscaler profits can cover the obligations if called, but the author raises open questions: whether that holds if hyperscaler revenue growth slows, whether backstop obligations creep from hyperscalers to merely "decent" companies, whether backstops go unenforced, and whether these bonds eventually get bundled and resold the way mortgage-backed securities were before 2007.
The conclusion is not that circular deals are bad. The essay compares them favorably to the financing patterns long used in critical minerals, oil and electric vehicles, and argues the danger is specific: overextended lenders supporting circular deals, or trying to move them off their balance sheets. It closes by quoting a SemiAnalysis projection that AI debt financing will grow into a multi-trillion-dollar credit market, with over $7 trillion outstanding by 2029, making it the second-largest asset-backed debt market after the roughly $13 trillion US mortgage-backed market.
Key facts
- Nvidia strikes circular deals such as CoreWeave's $6.3 billion agreement and neocloud Firmus's $505 million deal, providing capital and effectively guaranteeing to buy back compute capacity a startup cannot use.
- SpaceX leases its Colossus 1 and 2 data centers for over $2 billion a month, and both SpaceX and Meta profitably rent out capacity built for their own frontier-model efforts.
- A TeraWulf bond is guaranteed through Google backstopping lease-payment failures that pass through Fluidstack and Anthropic, letting Google support the debt without it appearing on Google's balance sheet, per FT reporting; Meta's $27.3 billion Hyperion bond and its $12.3 billion BlackRock-marketed deal follow the same off-balance-sheet pattern.
- SemiAnalysis estimates AI debt financing will exceed $7 trillion outstanding by 2029, making it the second-largest asset-backed debt market after the roughly $13 trillion US mortgage-backed market.
- MP Materials committed to a 10-year deal with the Department of War to sell all its neodymium-praseodymium output at a minimum of $110 per kilogram, an example the essay uses to show circular financing has decades of precedent in oil, mining and rare-earth industries.
Why it matters
Circular deals get cited as evidence the AI industry is a bubble about to repeat 1999. This analysis argues the opposite reading: the deals show compute becoming a fungible commodity, priced and financed the way oil, copper or electricity are, using financing patterns with a decades-long track record in capital-intensive commodity industries rather than a dot-com-style illusion.
Who it affects
Nvidia, OpenAI, Microsoft, CoreWeave, Meta, SpaceX, Google, Anthropic, TeraWulf, Fluidstack and Firmus all appear as parties to the specific deals discussed, alongside MP Materials and General Motors as a non-AI comparison. More broadly it affects any startup or neocloud trying to secure GPU access, and the banks and bondholders financing AI data-center buildout.
How to use it
The essay offers a way to tell healthy circularity from risky circularity. A deal that gives a startup a guaranteed customer and gives lenders confidence to fund infrastructure, like Nvidia's arrangement with CoreWeave, follows a well-tested commodity-financing pattern. The warning sign is a deal built to move debt off a company's balance sheet through a chain of backstops, as in the Google-TeraWulf-Fluidstack-Anthropic bond or Meta's Hyperion and BlackRock-marketed bonds, since that is where the essay says risk gets hidden rather than shared transparently.
How solid is it
The piece is a self-published analysis on emergingtrajectories.com, submitted to Hacker News by user cl42 (who is not necessarily its author) and discussed there with 72 points and 38 comments. It leans on named third-party sources, including SemiAnalysis estimates of Nvidia's compute off-take agreements and projected AI debt-market size, and Financial Times reporting on the TeraWulf bond structure. It is an opinion and analysis essay, not a regulatory filing, financial disclosure or peer-reviewed study, and its dollar figures for Nvidia's total commitments and off-take deals are themselves drawn from third-party estimates rather than disclosed contracts.
Risks and caveats
The essay flags two distinct risks in its own argument. First, the resale logic behind Nvidia's guarantees depends on GPUs and data centers staying fungible; a shift in chip standards, or a move toward smaller local clusters, could undercut it. Second, circularity turns dangerous when overextended lenders use backstop chains to keep debt off balance sheets, a pattern the author worries could expand to weaker guarantors, go unenforced, or eventually get bundled and resold the way mortgage-backed securities were before the 2007 crisis.
“AI Debt Financing will become a multi-trillion-dollar credit market, with over $7T of debt outstanding by 2029 driven both by AI IT Capex and AI Datacenter Capex needs [...] This will make it the second largest asset backed debt market after the US mortgage-backed financing market at just over $13T.”
— SemiAnalysis, quoted in the analysis
Компания Meta Platforms признана экстремистской организацией, её деятельность на территории РФ запрещена.