CoreWeave, Oracle and OpenAI's $50 Billion Revenue Report: Who Holds the Risk

On October 8 the Financial Times reported that OpenAI told investors its annualized revenue was near $50 billion at the end of September. A figure of about $70 billion had been going around for a week. OpenAI didn't comment. AI stocks fell together. By the close CoreWeave had lost about 8%, Oracle about 5.5%, and Broadcom and AMD about 4% each. Nvidia lost about 3%, roughly $170 billion of market value. Micron was down about 4% at midday.

I don't think the risk is spread that evenly. It sits where three things meet: OpenAI fills a big share of a company's signed future revenue, there's little moat behind that contract, and the company is borrowing to build for it. The table below is the first of those. The matrix after it covers the second.

Oracle has the biggest OpenAI book on S&P's estimate, and Microsoft has the biggest one a company has disclosed itself. Microsoft was down only about 1% at midday. CoreWeave, which fell the most, has the smallest OpenAI share of the four companies that rent OpenAI capacity. So the share alone doesn't explain October 8. And the chip names fell about as much as each other, though their ties to OpenAI are very different. Broadcom is tied to how OpenAI's chips get financed. AMD gave OpenAI warrants that vest as it buys. Nvidia's planned $100 billion investment shrank to a $30 billion stake. Micron has no OpenAI contract at all, and at midday it was down more than Broadcom.

The moat matrix is where the framework comes in. Transaction embedding and system of record measure whether a customer's own workflow runs through the company. Bundling and scale measure whether it can sell the same capacity to someone else. I read those four columns as the answer to one question: what does the company still own if OpenAI pays late? The framework doesn't model that directly. That's my reading of it.

Among the four landlords, the matrix splits cleanly. Microsoft and Amazon are strong on all four columns. If OpenAI pays slower, Azure and AWS still run the enterprise work that carries them. CoreWeave sits at the other end. It's weakened or not applicable on every column. Its OpenAI slice is smaller than the others', but nothing much stands behind any of its tenants, and it's building with borrowed money. That's why I'd put it first even with the smallest share.

Oracle is the hard case. Its database moat is real, and the matrix shows strong transaction embedding. Nobody rips out an Oracle database because OpenAI's revenue came in light. But the framework scores Oracle as one company, and I don't think that moat covers the new cloud business, which is being built on debt. S&P cut Oracle to BBB− in July and expects a free cash flow deficit of nearly $42 billion in fiscal 2027. If the OpenAI half of the backlog converts slowly, the database keeps earning while shareholders absorb the debt and any new stock.

The three groups hold three different kinds of claim on OpenAI, and each breaks in its own way.

• Landlords hold a lease. Oracle, Microsoft, Amazon and CoreWeave have signed multi-year capacity contracts. The landlord builds first and gets paid over years. If the tenant can't pay, the landlord has capacity it has to fill again. That's easy with a deep bench of other customers and hard without one.
• Chip vendors hold an order. They're paid on delivery. If OpenAI slows, next year's orders shrink, but chips already shipped are already paid for. Broadcom is the partial exception. On its September call it said outside partners, including Apollo and Blackstone, fund the assets, that it may provide modest residual value guarantees, and that it has less clarity on OpenAI's financing than on Anthropic's. That ties Broadcom's 2028 plan to how OpenAI gets funded more than a plain order would.
• Micron holds prepayments. Customers on its long supply agreements have paid cash up front, and Micron held $12.9 billion of it at the start of September. Micron doesn't sell to OpenAI directly. The OpenAI risk reaches it through memory demand from the chipmakers OpenAI buys from.

Put the table and the matrix together and you get a ranking that doesn't quite match October 8. CoreWeave and Oracle's shareholders come first, and the market got that part right. Broadcom deserves more attention than a generic chip name, because of the financing tie. That's my reading of the call, not a framework output, since the matrix rates Broadcom strong on embedding, record and bundling. Microsoft and Amazon hold the biggest OpenAI books with the most behind them. Micron sits furthest from OpenAI of anyone here.

The strongest objection runs the other way. A lease is a contract. An order isn't. If OpenAI needs to save money, the first thing it can cut is next year's chip purchases. Its rent at Microsoft, Amazon, Oracle and CoreWeave is signed and hard to walk away from. So a slowdown at OpenAI would show up first as lower chip orders while the landlords keep billing. On that reading Nvidia, AMD and Broadcom are more exposed than this note says, and the sell-off had the order roughly right.

There's real weight in that. Frontier labs and the newer GPU clouds that serve them are the marginal buyers of Nvidia's newest systems. AMD's OpenAI deal is its proof that a second supplier can win a frontier lab, so a slip would hurt the story more than the revenue. Broadcom has said OpenAI would become its second-largest custom-chip customer if the 2028 plan deploys. And the value of Nvidia's $30 billion stake rests on OpenAI's valuation, which now has to be argued from a smaller revenue base.

Micron is the weakest part of my case, and the objection is right to push on it. Memory prices aren't set by contracts. They're set by marginal demand, and right now the margin is AI. If OpenAI's build slows and its chip suppliers order fewer accelerators, they need less high-bandwidth memory. Prices for that memory and for ordinary DRAM could soften well before any prepayment is at risk. Micron's prepayments cover part of its volume, not its price. So Micron's tie to OpenAI is indirect, but a memory stock can still fall hard on an indirect tie.

The objection also has a point about Microsoft and Amazon. A contract is only as good as the payer. Microsoft has the largest disclosed OpenAI book, and Amazon committed up to $50 billion of equity to OpenAI, $15 billion up front, on top of its cloud deals. If OpenAI's growth stalls, Azure and AWS growth slows too, and both stocks are priced for that growth.

Here's where I think the objection stops. Chipmakers lose growth when an order is cut, but they haven't spent years of capital on one tenant. Nvidia sells to every large cloud. The order risk is real, but it's a growth risk spread across many buyers. The lease risk at CoreWeave and Oracle is concentrated and funded with debt. Broadcom is the one chipmaker where the objection and this note agree. And the point about Microsoft and Amazon is about the stocks, while the matrix is about whether the businesses keep earning. I think the stocks can fall without the moats moving.

For CoreWeave, I'd watch whether its backlog keeps growing from customers other than its two biggest. For Oracle, it's whether cloud revenue converts fast enough to narrow the cash deficit before more debt or stock gets issued. For Microsoft, it's whether Azure keeps growing as OpenAI's share of the backlog falls. For Amazon, it's whether the $100 billion expansion shows up in AWS backlog growth. For Broadcom, it's whether OpenAI's financing closes and how much Broadcom guarantees. For AMD, it's whether the first gigawatt of MI450 deploys on time, which also vests the first warrant tranche. For Nvidia, it's whether other buyers keep absorbing its newest systems. For Micron, it's memory pricing, which the Micron note covers.

The two numbers measure different things. OpenAI counts revenue net, after the share that goes to cloud partners. The higher figure appears to have been built gross, to compare with Anthropic, which counts partner sales in full. So the business didn't shrink on October 8. What changed is how big OpenAI looks next to the bills it has signed. That's worth worrying about, but each company in this note carries that worry differently, and a customer that fills half your backlog is a risk you can live with only if you own something it can't take with it when it leaves.