Big tech loves a big number. Often it is the same dollars doing a lap.
NVIDIA announces funding for OpenAI. OpenAI spends on cloud — Oracle, Microsoft Azure, AWS, and the rest. Those providers turn around and buy newer GPUs. A lot of that spend lands back at NVIDIA.
Call it investment. Call it partnership. Call it demand. On the circuit, it is also a financing loop: equity and credit go out, chip and cloud revenue come back, and the headline still reads like fresh growth.
None of that proves the technology is fake. It does mean "NVIDIA is funding OpenAI" is not the same story as "new money entered the system."
How we got here
In September 2025, NVIDIA and OpenAI announced a strategic partnership: an intent to deploy at least ten gigawatts of NVIDIA systems for OpenAI, with NVIDIA prepared to invest up to $100 billion as each gigawatt came online. Company posts framed it as infrastructure for the next phase of AI. Markets heard a demand signal.
By March 2026, a cleaner number was in the books. OpenAI closed a funding round with $122 billion in committed capital at an $852 billion post-money valuation. NVIDIA's share, according to OpenAI and wire coverage, was $30 billion — alongside SoftBank and Amazon. The earlier letter of intent did not need to survive as structured to matter. Equity did.
Then came Ohio. In July, CNBC reported talks of a backstop that could reach $250 billion. In August, NVIDIA filed an 8-K describing something more precise and still enormous: residual-value guaranties with SB Energy for a campus where an OpenAI affiliate would be the tenant, with NVIDIA's aggregate payment obligation cumulatively capped at $105 billion. NVIDIA also put $1.5 billion into SB Energy and remained the exclusive AI compute supplier at the site.
Jensen Huang's public answer was simple. This was not circular financing, he said; OpenAI would pay the lease. The filing is less slogan than credit structure. The chipmaker was helping underwrite the long-lived hall that would house the chips it sells.
Secondary readings of NVIDIA's subsequent quarterly disclosures put maximum gross guarantee exposure around $108.5 billion — including the Ohio cap and smaller AI-cloud partners — and future commitments in the hundreds of billions. Those figures need the 10-Q footnotes for a copy desk's final cite. The pattern does not.
What the record shows
While capital was arriving from the sellers of compute, OpenAI was locking offtake with the landlords of it.
In January 2025, SoftBank, OpenAI, Oracle, and MGX announced Stargate: an intended $500 billion of AI infrastructure investment over four years, with $100 billion to start. NVIDIA was named among the technology partners. Reuters carried the White House-adjacent unveil.
By mid-2025, the Oracle leg had hard edges in the press. The Financial Times, via Reuters, reported Oracle would buy about $40 billion of NVIDIA chips — on the order of 400,000 GB200s — for an OpenAI data center in Abilene. TechCrunch and others later described an Oracle cloud arrangement on the order of $30 billion a year in revenue, with OpenAI as the counterparty. Oracle's own SEC language should be pulled for the definitive contract shape; the offtake story does not wait on that polish.
In October 2025, Microsoft and OpenAI announced the next chapter of their partnership. OpenAI committed to an incremental $250 billion of Azure services. Microsoft kept a large equity stake. The press kit called it partnership. The cash path called it a customer of the cloud buying more cloud from its largest shareholder-adjacent supplier.
In November, Amazon and OpenAI announced a $38 billion agreement for AWS capacity — hundreds of thousands of NVIDIA GPUs, in the companies' telling. Bloomberg's headline did not hedge: Amazon would supply NVIDIA chips. In February 2026, Amazon went further: up to $50 billion of equity in OpenAI, and an expansion of the AWS relationship by another $100 billion over eight years, with a large Trainium component. Equity in; offtake out; silicon mix shifting, supplier still in the room.
Cloud providers do not invent accelerators from press releases. When OpenAI contracts for tens of billions a year of capacity, someone orders GPUs, networking, power, and land. For most of this cycle, that someone has defaulted to NVIDIA.
Oracle's reported Abilene buy is the blunt instrument: tens of billions of NVIDIA silicon earmarked for an OpenAI workload. AWS's $38 billion deal was sold, in part, as NVIDIA GPU capacity. Microsoft's CapEx has been dominated by short-lived compute assets — GPUs and CPUs — even as the company builds custom silicon. The exact NVIDIA invoice inside any hyperscaler's CapEx line is almost never disclosed. The direction of travel is not mysterious.
So the lap runs: NVIDIA (and sometimes Amazon, Microsoft, SoftBank) put equity or credit behind OpenAI; OpenAI signs multi-year clouds; the clouds buy accelerators; NVIDIA books data-center revenue and calls it demand. A reporter who stops at the first press release will write a growth story. A reporter who keeps walking will write a financing story that happens to produce growth metrics.
Follow any one of those announcements alone and it looks like demand. Follow them in sequence and the same economic relationship keeps getting renamed.
What changed once it was public
The OpenAI circuit is the loudest version of a wider pattern.
CoreWeave raised and spent as a neocloud on NVIDIA gear while NVIDIA held equity and, in market commentary, capacity backstops. When critics called that circular, Huang called the label ridiculous: NVIDIA's stake was a small slice of the capital customers still had to raise. In January 2026, Business Insider recorded him defending another CoreWeave check on the same grounds.
In August 2026, Anthropic's reported $35 billion cloud deal with NVIDIA-backed Lambda sat on a Texas campus where, wire reporting said, NVIDIA held the lease. Investor, supplier, landlord — three nameplates, one stack.
AMD's October 2025 pact with OpenAI flipped the direction and made the mechanism explicit. OpenAI would take up to six gigawatts of AMD GPUs. AMD issued warrants for up to 160 million shares at a penny, vesting as purchases happened. The customer's buying created the supplier's equity dilution. That is not a secret handshake. It is in the 8-K.
Noah Smith, Semianalysis, Fortune, Bloomberg's circular-deals graphics, and a run of specialist newsletters have already named the thing: vendor financing, demand enablement, circular deals. Huang's later line, as Axios paraphrased from a Goldman gathering, did not so much deny the loop as rebrand it. NVIDIA puts a little money in; a lot of money comes back. If that is circular, he said, do more of it.
What the rule requires
When the next partnership drops, the useful questions are smaller than the headline:
Who wrote the check? Who signed the offtake? Who buys the GPUs? Who guarantees the lease if the tenant stumbles? Are those four answers four companies — or two, wearing different hats?
Investment can be real. Cloud spend can be real. GPU shipments can be real. The error is treating each press release as independent evidence of new money entering the system when the same balance sheets are waving from both ends of the deal.
Related-party offtake is still offtake. Residual-value credit is still credit. The duty, for a reader of filings, is to say which hat each nameplate is wearing — not to flatten the stack into "demand."
What we are not treating as proof
The public record does not show that ChatGPT is a mirage, or that token demand is invented, or that NVIDIA's revenue is fake in the Global Crossing sense of swap-for-swap. Filings and company posts support a different claim: suppliers and hyperscalers are using their balance sheets to midwife the offtake that becomes their own top line.
What is still missing
The record does not show, dollar for dollar, how much of NVIDIA's $30 billion equity check was spent on Oracle racks that then bought NVIDIA GPUs. Uses of proceeds at that granularity wait for an S-1 related-party table, a leak, or a filing nobody has forced into the open. The percentage of NVIDIA data-center revenue that depends economically on NVIDIA's own customer credit is likewise UNKNOWN. Google Cloud's OpenAI offtake, in primary detail comparable to Microsoft and Amazon, is thin in the public record.
Whether OpenAI's committed capital is cash received, a multi-year draw, or a mix remains UNKNOWN beyond the company posts. Oracle's definitive SEC contract shape for the Abilene offtake is UNKNOWN. Secondary readings of NVIDIA's maximum gross guarantee exposure around $108.5 billion still need the 10-Q footnotes for a copy-desk cite.
Those gaps matter. They are not a reason to treat the press kit as the cash path.
Big tech loves a big number. Follow the cash path, not the press kit.