Short answer
Often less than the headline, and sometimes more. A compute commitment is a contract to buy computing capacity over many years; some are binding take-or-pay contracts that must be paid even if the buyer does not use the capacity, while others can be cancelled on short notice, depend on delivery, or are only a letter of intent.
Follow-up question How do you read the next giant compute headline?
Compute is the processing power that AI models use to train and to answer. Almost no lab owns enough of it, so labs reserve it from clouds and data centre operators with contracts that can last a decade. A press release says "$100 billion." The contract behind it says who must pay what, when, and under which conditions, and those details are what count.
Until recently, outsiders rarely saw them. Anthropic's IPO filing now lays out its arrangements, which makes it a useful case study for how such deals work across the industry.
The four kinds of commitment
| Kind | What it means | Binding? |
|---|---|---|
| Take-or-pay | You pay a minimum, whether you use the capacity or not | Strongly |
| Pay on delivery | Fees start when capacity is delivered and accepted | Once delivered, with exits if the supplier fails |
| Cancellable | Either side can leave after a notice period | Only for the notice period |
| Letter of intent | A statement of plans | Generally not |
The same dollar figure means very different things in each row.
One lab's contracts, line by line
These are Anthropic's disclosed arrangements, as stated in its October 2026 filing:
| Counterparty | Amount | Period | Terms |
|---|---|---|---|
| At least about $114.5B | To July 2033 | Pays any shortfall; non-cancellable barring breach | |
| Amazon | At least $110B | To April 2036 | Pays any shortfall; non-cancellable barring breach |
| Microsoft | At least $44.9B | To May 2033 | Owed regardless of usage |
| Broadcom (leases of TPU hardware) | About $161.2B, excluding interest | Five-year leases | Non-cancellable except on default |
| SpaceX's xAI unit | Up to $120.5B | Into 2029 | Either side may cancel on 90 days' notice after early 2027 |
| Nscale | Up to about $44.6B | Into 2033 | Pay on delivery; supplier must secure financing |
| Lambda | About $41.3B | Into 2034 | Pay on delivery |
| AMD | Expected above $20B | Not fixed | No set dollar amount |
| Additional Google TPU capacity | Estimated $125–150B | Delivered 2028–2031 | Depends on capacity and conditions |
A worked comparison helps. The first three rows add up to about $269 billion in minimum commitments. Those run alongside revenue of $16.3 billion in the first half of 2026. Adding all the rows with stated amounts other than AMD and the additional TPU capacity gives about $637 billion. That is our own simple sum of the named contracts. The November 2 amendment now gives the filing's own total: about $895.4 billion of purchase commitments and leases, nearly all for compute and infrastructure, so the named contracts are only part of it. Some contracts may overlap or be amended, and some are ceilings rather than minimums. The filing says its commitments stretch as far as 2045.
The risk is plain in the document. Anthropic warns that its revenue is mainly consumption-based, while these bills are fixed and long-term, and that it may make these commitments on revenue expectations that do not come true.
The other side of every contract
One company's commitment is another's backlog. "Remaining performance obligations" (RPO) is the accounting term for revenue under contract and not yet recognized. So commitments reappear in the supplier's filings:
| Supplier | Backlog | How soon it becomes revenue |
|---|---|---|
| Microsoft (commercial) | $677B | About 30% in 12 months |
| Oracle | $664B | About 13% in 12 months |
| Google Cloud | $560.8B | Just over 55% of Alphabet's backlog in 24 months |
Amazon's filing connects the two sides directly. AWS's commitment from OpenAI grew by $100 billion over eight years, on top of $38 billion, and its commitment from Anthropic by more than $100 billion over ten. Both include obligations tied to AWS's own chips.
The chain continues upstream. Nvidia's supply commitments jumped from $119 billion to $279 billion in a single quarter, so the chip supplier is itself committing to memory makers and factories on the expectation that today's orders will hold.
This means a single contract can inflate several companies' numbers at once, and a single customer's trouble can reduce several at once. Backlog is real evidence of demand, but only as good as the buyer's ability to pay.
Why totals are slippery: the OpenAI example
OpenAI shows how far a headline can drift. Sam Altman had spoken of about $1.4 trillion in infrastructure commitments. In February 2026, CNBC reported that the company was telling investors to expect about $600 billion of compute spending through 2030. In September, TNW, summarizing Financial Times reporting on a July presentation, gave about $856 billion for compute and infrastructure.
Those numbers cannot be compared directly. They may cover different costs and periods, and none comes from an audited filing. The change also does not mean a contract was torn up; a "target" and a signed obligation are different objects. OpenAI's own contractual picture stays private until it publishes a prospectus. Nvidia's letter of intent for up to $100 billion is another example: Nvidia's chief financial officer said in December 2025 that it was still not final.
Gigawatts are not dollars
Many deals are announced in gigawatts, the amount of electrical power the data centres will draw. It tells you the physical scale, not the price. A deal for 5 gigawatts might include chips or just the buildings and power. Nvidia's guarantee for an Ohio campus covers land, power and shell, not the computers. So check what is inside the figure before converting it to money.
Try it: Look at Oracle's or Microsoft's backlog figure in its latest quarterly report, then find the share due within 12 months. Multiply the two, and compare the result with its current annual revenue.
How do you read the next giant compute headline?
Ask five questions. Is it a signed contract, a framework or a letter of intent? Is the amount a minimum, an estimate or a ceiling? How long does it run, and what is due in the next 12 months? Can either side leave, and on what notice? And can the buyer pay from revenue, or does it need its next funding round?
If an announcement answers none of them, treat the figure as a statement of ambition until a filing says otherwise. This explains how contracts work and is not a view on any company's shares.
Sources
- Form S-1 registration statement, Anthropic, PBC — Anthropic / US SEC, October 26, 2026
- Amazon Form 10-Q, quarter ended September 30, 2026 — Amazon / US SEC, October 30, 2026
- OpenAI resets spend expectations, targets around $600 billion by 2030 — CNBC, February 20, 2026
From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0: credit the guide, link to https://guidetoartificialintelligence.com/en/learn/ai-compute-commitments-explained/, no commercial use without permission, and keep the same licence.