OpenAI Compute Commitments: $720 Billion in Contracts, $50 Billion in Revenue

Six compute contracts that OpenAI or its counterparties have disclosed add up to about $720 billion. Spread across their stated or assumed terms, those OpenAI compute commitments cost somewhere between $117 billion and $136 billion a year. OpenAI told investors its revenue was running at close to $50 billion a year at the end of September 2026. The bill is more than twice the revenue, before a single salary is paid.

Short answer

OpenAI has disclosed or been named in six compute contracts worth about $720 billion: Oracle, Microsoft, two with AWS, CoreWeave and Cerebras. Averaged over their terms, they imply $117 billion to $136 billion of compute spending a year. That is 2.3 to 2.7 times the roughly $50 billion revenue run rate OpenAI reported for September 2026.

What counts as an OpenAI compute commitment in this ledger

A compute commitment, for this post, is a contract under which OpenAI agrees to pay someone else for data centre capacity it will use. It is a purchase obligation for a service, priced in dollars, with a counterparty that can be named. That definition is narrower than most of the totals that circulate, and it is narrower on purpose.

The test for inclusion is simple. Either OpenAI or the counterparty has stated the dollar value in its own announcement, or a named outlet reported it from documents or people familiar with the deal. Every row in the ledger below meets one of those two bars, and the table says which one.

Cloud capacity, not chip supply deals

The chip agreements with Nvidia, AMD and Broadcom are left out. Those are supply and investment arrangements measured mostly in gigawatts, and several are framework letters rather than priced purchase orders. Adding them mixes two different kinds of obligation and double counts, because the chips OpenAI rents from Oracle or Microsoft are often the same chips those suppliers are selling. The equity side of the Nvidia relationship is traced separately in the filing-by-filing trace of Nvidia's OpenAI investment.

The exclusion makes this ledger smaller than the $1.4 trillion headline. I think that is the correct trade. A smaller number you can source line by line is more useful to a buyer than a larger one assembled from press releases that describe capacity in megawatts.

Announced value against contract value

Several of these figures are ceilings. CoreWeave describes its contracts as worth "up to" a sum. The Cerebras value comes from a person familiar with the deal, not from either company. The Oracle figure has never appeared in an Oracle filing under OpenAI's name. Each of those caveats is carried into the table rather than dropped, because they change how much weight a row can bear.

The ledger: six disclosed OpenAI compute contracts

Here is every compute contract that clears the bar, in order of size. The source column names the first document that put the number on the record, not the most popular article about it.

Values are as announced and include ceilings. Retrieved 25 October 2026. Contracts whose value has not been reported in dollars, including any Google Cloud capacity arrangement, are excluded rather than estimated.

The two largest rows carry the least disclosure

Oracle and Microsoft together account for $550 billion, or 76% of the total. Neither row comes with the kind of detail the smaller contracts have. Microsoft's announcement states that OpenAI "has contracted to purchase an incremental $250B of Azure services" and says nothing about the term, the ramp or the capacity behind it.

The Oracle figure is the opposite problem. The term and start year are reported, but the dollar value has never been put on the record by either party. Oracle's remaining performance obligations jumped 359% to $455 billion in the quarter the deal was reported, which is strong circumstantial evidence and still not a named disclosure. How much of Oracle's backlog rests on this single customer is examined in the credit analysis of Oracle's RPO backlog.

Turning OpenAI's contract totals into an annual compute bill

A $300 billion contract over five years and a $38 billion contract over seven years are not comparable until you divide by the term. The annual figure is what has to be covered by revenue, so it is the only version of the number that tells you anything about affordability.

The method is deliberately crude. Each contract value is divided by its term, assuming even spending across the life of the deal. Where no term exists, the table states the assumption used, so you can replace it with your own.

Working: 300/5 + 250/8 + 100/8 + 38/7 + 22.4/5 + 10/3 = 117.0. Replacing 250/8 with 250/5 gives 135.7. Straight-line averages, not a payment schedule.

The assumptions that move the annual number

The Microsoft term is the single largest swing factor in the whole calculation. Moving it from eight years to five adds $18.7 billion a year to the bill. I chose eight as the low case because it is the longest term attached to any other contract in the ledger, and five as the high case because it matches Oracle's.

The CoreWeave assumption matters far less. Even if its term were half what I assumed, the total would move by about $4.5 billion. The Cerebras value is a floor, so the true figure is probably higher, but at $3.3 billion a year it cannot change the conclusion either way.

Straight-line spreading also understates the peak. Oracle's capacity does not start until 2027, and Amazon says all of the original AWS capacity is targeted to be deployed before the end of 2026. When contracts overlap at full capacity, one year can carry well above the average. The steady-state figure is a floor on how much revenue OpenAI needs in its heaviest years, not a ceiling.

The revenue side: $50 billion, and a correction worth reading

OpenAI does not publish audited revenue. What exists are figures the company gives investors, which then reach the press. The most recent of those is the one that matters for this ledger, and it arrived with a correction attached.

The Financial Times reported on 8 October 2026 that OpenAI had recently told investors its revenues were approaching $50 billion on an annualised basis at the end of September. The FT described that as far short of the $70 billion it and other outlets had reported weeks earlier. Yahoo Finance, summarising the same report, noted that at $50 billion OpenAI's annualised revenue would sit below Anthropic's reported $65 billion.

A run rate takes a recent period and multiplies it to a year. It is a snapshot, not a result, and it moves with every month of growth or slowdown. For a ledger of multi-year contracts, a snapshot is the right comparison only if you also ask what revenue has to become.

Why the $20 billion gap matters for the ledger

The $20 billion correction is not an accounting curiosity. Against the low-case compute bill of $117 billion, a $70 billion run rate covers 60% of compute. A $50 billion run rate covers 43%. The revenue figure the market believed for two weeks in September made the ledger look meaningfully more affordable than it is. I would treat any coverage claim made in that window as stale.

That gap is also why the valuation debate and the compute debate are the same debate. A price set on a run rate is a price set on the ability to keep paying these contracts. That link is worked through in the analysis of OpenAI's valuation as a multiple of run rate.

The coverage ratio: OpenAI's compute bill against its revenue

Divide the annualised compute bill by the revenue run rate and you get a coverage ratio. A ratio above 1 means compute alone costs more than the company earns. It says nothing about staff, sales, research or the free users who consume inference without paying.

At $50 billion, the ratio runs from 2.3 to 2.7. In monthly terms, the run rate implies about $4.2 billion of revenue a month against $9.75 billion to $11.3 billion of averaged compute obligations. If the run rate reaches $70 billion, the figure that had been reported in September, the ratio falls to 1.7 to 1.9. It is still above 1.

My reading is that the ledger is not evidence of recklessness on its own. It is evidence that OpenAI has pre-sold its own future growth to suppliers. Revenue has to reach somewhere between $117 billion and $136 billion just to cover these six contracts, which is roughly 2.3 to 2.7 times today's level, before any margin exists.

Set against the $1.4 trillion headline

Sam Altman has said OpenAI is committed to spending about $1.4 trillion to develop 30 gigawatts of computing, as Reuters recalled when reporting the Cerebras deal. The six contracts here total $720.4 billion, or about 51% of that figure. The other half sits in chip agreements, owned data centre projects and capacity that has no priced public contract yet.

The $1.4 trillion is 28 times the $50 billion run rate. That ratio gets quoted often, and I think it misleads, because it compares a multi-year total with one year of revenue. The annualised ratio in this post is less dramatic and much harder to dismiss.

Who carries the risk on the other side of OpenAI's contracts

Every row in the ledger is an asset for someone else. A commitment OpenAI cannot meet becomes a backlog a supplier cannot convert. The exposure is not spread evenly, and the smaller names carry more of it relative to their size.

Oracle and CoreWeave: concentration on a single customer

Oracle's case is concentration at scale. A contract reported at $300 billion against a backlog of $455 billion in the quarter it landed means one customer could account for a large majority of the reported obligation. Oracle has to build the data centres before OpenAI pays for them, which turns a revenue question at OpenAI into a debt question at Oracle.

CoreWeave's OpenAI total is much smaller in dollars at $22.4 billion. The company built it in three steps during 2025: $11.9 billion, then $4 billion, then $6.5 billion. It is also a business funded heavily with debt secured on contracts like these, which is why the stress points are mapped in the breakdown of CoreWeave's debt stack. CoreWeave and Oracle shares both fell when the $50 billion revenue figure was reported, according to several market reports that week.

Microsoft and Amazon: customers that are also shareholders

Microsoft and Amazon sit in a different position because they own part of OpenAI. Microsoft's announcement valued its stake at about $135 billion. Amazon committed $50 billion of equity in February 2026, in the same announcement that expanded the AWS agreement by up to $100 billion over eight years.

That structure softens the risk for both, and it also complicates the reading of the contracts. When an investor's cash comes back as cloud revenue, the commitment is partly funded by the supplier itself. The mechanics of that loop are covered in the explainer on how the AI circular deals work. I would not treat the Microsoft and Amazon rows as equivalent to Oracle's in credit terms, even though all three are priced in the same dollars.

Where this argument is weakest

The ledger is built from public disclosures, and public disclosures are the weakest possible base for a calculation this consequential. Three problems are serious enough that you should hold the conclusion loosely.

The terms are partly invented by me, and the values are partly ceilings

Two of the six terms are my assumptions, and one of them covers the second largest contract. CoreWeave's figure is a maximum, Cerebras's value is a single anonymous source, and Oracle's has never appeared under OpenAI's name in a filing. Contracts like these usually include ramps, termination rights and capacity conditions that no outsider has seen. Any of them could make the real annual obligation materially lower than a straight-line average. A ledger is not a payment schedule, and I have not claimed it is one.

The revenue side is weaker still. A run rate shared with investors is unaudited, and this one was corrected by $20 billion within a month of the earlier figure circulating. OpenAI's revenue grew from what reports put at about $40 billion in August to close to $50 billion in September. If that pace held, the coverage ratio would fall faster than any snapshot suggests, which is the strongest argument against this post. Growth of that kind is exactly what the contracts are betting on, and a static ratio cannot capture it.

Frequently asked questions

Where to start if OpenAI sits in your stack

If you are a founder or finance lead buying from OpenAI, or from a vendor built on it, this ledger is a supplier-risk input. It is not a market opinion. Do one thing this week. List which of your products depend on a single model provider, and write down what it would cost you to switch if pricing or rate limits moved against you next year.

Then put a date in your calendar for OpenAI's next investor update leak and Oracle's next quarterly filing. Recompute one ratio: annualised compute against run rate. If it falls below 2 within a year, the bet is working. If it does not, expect the pressure to show up in what you pay.

Related reading