Key points
- Nvidia, Microsoft, Oracle, Broadcom, AMD, and CoreWeave have combined for close to $1 trillion in OpenAI chip and cloud commitments since 2025.
- Nvidia's original $100 billion OpenAI plan fell apart. It put in $30 billion instead, as part of a $110 billion round that closed February 2026.
- OpenAI cut its own 2030 spending target to about $600 billion in February 2026, well below the $1.4 trillion Sam Altman had touted.
- A May 2026 snag on an $18 billion Broadcom chip order, stuck without a firm Microsoft purchase commitment, showed how one hesitant buyer can jam the loop.
Nvidia (NVDA) sells GPUs to CoreWeave (CRWV). Nvidia also owns close to 11% of CoreWeave. And in September 2025, CoreWeave turned around and sold $6.3 billion of compute back to Nvidia. Same two companies. Three different roles. All running at once.
That's a circular AI deal. The name sounds abstract. The mechanics aren't. A company invests in its own customer, sells that customer product, then buys some of that customer's output back. By mid-2026 it was the thing Wall Street couldn't stop arguing about, because versions of this loop show up almost everywhere OpenAI touches, and the dollar figures involved stopped sounding real months ago.
Nvidia and Microsoft own pieces of what they sell to
OpenAI is where the biggest version of this plays out, and Nvidia's own numbers are the best proof of how fast these headline figures go stale. In September 2025, Nvidia and OpenAI announced a plan for up to $100 billion, paid out as OpenAI deployed 10 gigawatts of Nvidia hardware, in exchange for non-controlling shares. Nvidia's own finance chief was telling analysts within months that nothing had actually been signed. By February 2026, the Wall Street Journal confirmed it outright. No contract. No money moved. The original $100 billion plan never became real.
What actually happened is smaller and built differently. OpenAI closed a $110 billion funding round on February 27, 2026, Amazon put in $50 billion, SoftBank $30 billion, and Nvidia $30 billion, valuing the company at $730 billion. Nvidia's $30 billion is a straight equity stake this time, not tied to gigawatts or deployment milestones the way the original plan was. The number most people still quote for Nvidia's stake in OpenAI, that $100 billion, is stale and about three times too big.
Microsoft (MSFT) got there a different way. It had been writing checks to OpenAI since 2019, something like $13 billion total, before the two sides tore up the old arrangement and started over on October 28, 2025. What came out the other side was a new corporate shell for OpenAI's for-profit business, and Microsoft holding a little over a quarter of it, a piece worth close to $135 billion the moment the ink dried, since grown to about $228 billion by May 2026 as OpenAI's valuation climbed. It's still paying Microsoft to run workloads on Azure too. Landlord and part-owner, same building.
AMD did it backwards
AMD (AMD) built the same kind of loop from the other direction. OpenAI is on the hook to deploy up to 6 gigawatts of AMD's Instinct chips over the next several years. What AMD handed back was paper, the right to buy 160 million AMD shares for a penny apiece. No cash, no discount, just that. Cash it in fully and OpenAI would own something close to a tenth of AMD. There's a catch, several actually. The shares only vest in pieces, tied to OpenAI actually hitting purchase targets and AMD's own stock clearing specific price levels, and the window to cash any of it in runs out in October 2030. Nobody attached a real dollar figure to the whole arrangement. Both companies just called it "billions" and let everyone else do the math off the hardware numbers. Same setup as Nvidia and Microsoft, mirrored. This time it's the buyer who ends up owning a slice of the seller.
Oracle, Broadcom, and CoreWeave skipped the equity, not the risk
Not everyone in this story owns a piece of OpenAI. Oracle (ORCL), Broadcom (AVGO), and CoreWeave don't, and each one's exposure looks a little different up close.
| Company | OpenAI commitment | Notable detail |
|---|---|---|
| Oracle (ORCL) | About $300 billion of cloud capacity, part of the Stargate buildout | Executives call the contract non-cancellable, the thing propping up Oracle's long-term valuation story |
| Broadcom (AVGO) | 10 gigawatts of custom AI chips, OpenAI's own design | First chip, nicknamed Jalapeño, unveiled June 2026 with early samples for testing; full build reported near $500 billion |
| CoreWeave (CRWV) | $11.9 billion to start in early 2025, expanded twice to $22.4 billion by that September | Nvidia holds near 11% of CoreWeave on its own, a separate stake |
None of these three hold OpenAI stock. Their risk runs straight through the contract instead. They're financing a massive buildout on the promise that OpenAI can eventually pay for it. That's arguably worse than owning equity, not better. A contract is only as good as the company that signed it. Our neocloud explainer gets into how CoreWeave and companies like it built an entire business model on exactly this kind of concentrated exposure to one customer.
The 1999 comparison, and where it falls apart
The comparison everyone reaches for is Cisco lending money to dot-com-era internet providers so they'd buy its routers, until the providers ran out of real customers and the stock went down with them. It's a fair comparison. It's also become a cliché, showing up in nearly every AI-bubble take written this year, ours included now.
Behind the cliché is a real number. OpenAI is projected to lose around $14 billion in 2026, nearly triple last year's loss, while committing to hundreds of billions more in infrastructure. Bloomberg mapped the loop in an entire graphics package, JPMorgan's asset management arm published a note asking outright whether the circularity itself is a warning sign, and Michael Burry, who's made his own bet against AI-linked chip stocks, keeps making the same point. When a chipmaker's revenue depends on a customer losing money on every dollar spent, and the chipmaker is also funding that customer's ability to keep spending, you're watching the same dollar get counted twice, dressed up as new demand.
Scarcity is where the comparison breaks down. Cisco had warehouses of unsold routers by 2000, waiting on ISPs that couldn't scrounge up enough paying subscribers to justify what they'd already bought. Nvidia's chips have stayed backlogged for years running, and actual demand for compute keeps outpacing what the industry can physically build. Funding a customer that can't get enough of your product is a different bet than propping up a customer nobody wants. OpenAI isn't one fragile ISP either. Its spending is spread across Microsoft, Amazon, Google, and a long list of businesses building on top of its models, backed by some of the best-capitalized companies alive, not venture money running on fumes. None of that erases the risk. It's a much bigger, better-funded base than what came apart in 2001. We've gone deeper on that side of the argument in our bubble-or-dip breakdown and the hyperscaler capex numbers behind it.
Where it actually broke
The clearest sign that circular financing has a ceiling showed up in Broadcom's chip deal. Broadcom would only finance the first phase, roughly 1.3 gigawatts and $18 billion, if Microsoft agreed to buy around 40% of the output. Microsoft hadn't signed as of May 2026, reportedly because it wants standardized data centers it can reuse for other customers, while OpenAI needs facilities built specifically around its own chip. Broadcom's stock dropped on the news. A press release and an actual financed, revenue-generating contract are two different things, and this is exactly where that gap shows up.
$1.4 trillion was the number Sam Altman kept citing for most of 2025. Then, in February 2026, OpenAI walked investors down to something smaller, somewhere near $600 billion in compute spending through 2030. The whole structure isn't collapsing. But OpenAI just admitted, quietly, that the bigger number wasn't one it wanted to keep defending.
What actually matters if you're holding these names
Forget the press release. The filing has the real answer. A 10-K or 10-Q shows whether a company has actually booked revenue from one of these deals, not just announced one, and whether its stake in a customer still sits on the books at cost or is already carrying paper gains that could vanish just as fast as they showed up. Pay attention to the language too. "Non-cancellable," how Oracle describes its Stargate contract, means something very different from "no assurance of a definitive agreement," how Nvidia described its own OpenAI deal. And watch which buyers actually have other options. Microsoft's hesitation on the Broadcom order mattered more than it would from a smaller player, because Microsoft, unlike OpenAI, doesn't need any single vendor to keep running.
Every ticker above links to its live SEC filings. Run any of these names through our stock score tool too. It grades profitability, dilution, and balance sheet strength in about ten seconds. Faster than trusting a headline.
Sources
- Nvidia and OpenAI, strategic partnership announcement (September 2025)
- CNBC, Nvidia, OpenAI appear stalled on their mega deal (February 2026)
- CNBC, OpenAI announces $110 billion funding round with Amazon, Nvidia, SoftBank (February 2026)
- AMD and OpenAI, strategic partnership announcement (October 2025)
- CNBC, OpenAI completes restructure, Microsoft takes 27% stake (October 2025)
- SiliconANGLE, OpenAI and Oracle strike $300B cloud computing deal (September 2025)
- OpenAI and Broadcom, strategic collaboration announcement (October 2025); CNBC, first Jalapeño chip unveiled (June 2026)
- Yahoo Finance, Broadcom shares slump as OpenAI's $18B chip deal hits financing snag (May 2026)
- CNBC, Nvidia says "no assurance" of deal with OpenAI (November 2025)
- CoreWeave, agreement expansions with OpenAI (2025); CNBC, CoreWeave discloses $6.3 billion Nvidia order (September 2025)
- CNBC, OpenAI resets spending expectations to about $600 billion by 2030 (February 2026)
- Bloomberg, AI circular deals: how Microsoft, OpenAI and Nvidia keep paying each other
- J.P. Morgan Asset Management, Does circularity in AI deals warn of a bubble?
- Yahoo Finance, OpenAI's own forecast predicts $14 billion loss in 2026, citing The Information
This is general market commentary and opinion, not investment advice. Markets can go down as well as up, and you can lose money. Always do your own research and consider speaking with a licensed financial professional before making any investment decision.
