What is the compute clause for an AI vendor contract?

The compute clause is the contract language that protects you when the compute supplier behind your AI vendor reprices. It has three parts: a disclosure right that names the primary compute supplier serving more than 25 percent of your workloads, a duration disclosure that puts the supplier contract into 12-month bands, and a pass-through trigger that caps compute-driven price increases at 10 percent year on year. Together they convert an opaque vendor relationship into a visible one without redrafting the whole contract.

How does the compute clause differ from a standard AI vendor SLA?

An SLA covers performance, availability, and incident response. The compute clause covers the supply chain that sits underneath those numbers. An SLA tells you what the vendor owes you when something breaks. The compute clause tells you what the vendor owes you when the supplier behind the vendor reprices, runs out of capacity, or signs a new fixed contract. The two are complementary, not substitutes. Most enterprise contracts in 2026 have a good SLA and no compute clause.

Why does AI vendor pricing keep moving without a renewal event?

Pricing moves between renewals when the vendor changes how it routes workloads across compute suppliers. A workload that ran on supplier A in March can run on supplier B in November, and supplier B may be on a different unit-economics curve. Without a disclosure right or a pass-through trigger in the contract, the buyer has no way to see the change and no way to challenge the resulting per-seat or per-token movement. The fix is contractual, not technical.

When should a Chief Information Officer apply the compute clause framework?

Apply the framework before signing any new AI vendor contract above 250,000 dollars annualised value, and at every renewal of an existing one. It also applies any time a vendor announces a new compute partnership, opens a new data center region, or appears in a public filing of a compute supplier. Use it as a one-page amendment template that your procurement function reaches for by default. The wrong time to apply it is after a repricing event has already moved your unit economics.

What is the first step to adding a compute clause to your AI vendor contracts?

The first step is an inventory. Pull every active AI vendor contract over 250,000 dollars annualised value and read each one for language touching compute supply, repricing, or cost-component disclosure. Most will have none. Mark those for amendment at next renewal. The inventory itself is the deliverable for the first week. Drafting the clause and negotiating it come after you know which contracts need it and which already have partial coverage.

The Compute Clause Most AI Vendor Contracts Are Missing in 2026

A compute clause is the contract language that protects you when the compute supplier behind your AI vendor reprices. Most enterprise AI contracts signed in the last 18 months do not have one. The model vendor is on the paper. The compute supplier sitting under the model vendor is not. That gap is now visible, measurable, and about to start moving budgets.

This is not a hypothetical. SpaceX filed its S-1 last week. One line in the filing said Anthropic is paying SpaceX 1.25 billion dollars a month for compute capacity through 2029. That is a single, named, multi-year fixed contract between a model lab you probably already pay for and a compute supplier whose data centers sit underneath it. Three years of compute, locked, on a number large enough to set a floor under Anthropic's pricing for the rest of the decade.

I sat with that filing on Sunday morning. By lunchtime I had rewritten the vendor review template I use with executive teams. The change was not philosophical. It was a single new section called the compute clause. This week I want to walk you through why it belongs in your template too, what the three parts of the clause look like, and what to do with the answers when you get them.

I have been on the call where this conversation should have happened and did not. A Chief Operating Officer at a large consumer brand asked me last quarter why his per-seat AI cost had moved 18 percent in nine months without a renewal event. We dug. The vendor had not changed price. The vendor had changed which compute partner it routed the largest workloads through, and the new partner was on a different unit-economics curve. He did not have the right to ask the question because the right was not in the contract. We added a sentence to the next renewal that gave him an annual cost-component breakdown. It took 90 minutes of legal time. It would have saved a board conversation if it had been in the original paper.

A second story from the same quarter. A Chief Financial Officer at a Fortune 500 manufacturing company asked me to sit in on a vendor renewal for an AI assistant rolled out to 12,000 employees. The vendor came in with a 22 percent price increase, cited inflation and infrastructure costs, and offered a five-year deal at the new rate. The CFO wanted to push back but had nothing to push back with. The original contract was silent on every variable except per-seat price. We spent two hours building a counter that asked for cost-component visibility before accepting any increase above 8 percent. The vendor agreed to disclose the compute supplier and the duration band. The price moved from 22 percent to 11 percent. The CFO did not get the contract he wanted, but he got the artefact he needed: a documented set of compute-side facts that he could take to the audit committee in October. The conversation went from "trust us, costs are up" to "here is what is driving 7 of the 11 points of the increase, and here is the band on the supply contract that produced it." That is what disclosure rights buy you when you put them in the paper.

That is the gap I want to close in this essay.

The Anthropic-SpaceX disclosure matters because it is the first time the largest single component of an AI vendor's cost base is sitting in a public filing in plain sight. SpaceX's Q1 AI unit, the one that contains X and xAI, generated 818 million dollars in revenue. Anthropic is paying SpaceX more in a single month than xAI earned in a quarter. That is the supply-chain story. The compute supplier has visibility into the buyer's economics that the buyer's own customers do not have, and the buyer is now structurally exposed to one supplier for the next three years. Public filings will keep making this asymmetry uncomfortable for procurement teams that have not modelled it.

There is a second tell in the same week's news. Andrej Karpathy joined Anthropic's pretraining team the same day the filing surfaced. The pitch is using Claude to accelerate Claude. The recursive improvement loop is no longer a slide. It is a hire backed by a 15 billion dollar annualised compute envelope. The pace at which model behaviour, inference cost, and serving infrastructure all move together is about to step up. If your contracts only protect you on the model variable, you are about to find out that the other variables move faster.

So what is the compute clause. Three parts. Each one is a single sentence in a contract amendment. Each one gives a procurement team the right to ask a question that today most of them cannot ask without renegotiating.

Part one. The disclosure right. A line in the contract that says the vendor will name, on request, the primary compute supplier serving more than 25 percent of your workloads on an annualised basis. Not the API logo. Not the cloud region. The actual entity providing the silicon and the data center. The threshold matters. Twenty-five percent is high enough that small auxiliary routing is exempt, and low enough that any material concentration risk gets surfaced. You are not asking for trade secrets. You are asking the same question your audit committee asks about your own concentration risk on revenue, supply, and customer. The default answer most vendors will give you on day one is "we use multiple compute partners." That answer is no longer enough. The Anthropic-SpaceX disclosure showed the gap between marketing language about diversified compute and the reality of a single multi-billion-dollar fixed commitment. The disclosure right closes that gap inside your contract, not outside it.

Part two. The duration disclosure. A line that says the vendor will disclose, in bands, the remaining contractual duration of its primary compute commitment. Bands are enough. You do not need the exact dates. You need to know if the supply contract is a 12-month renewal in a tight market or a four-year fixed deal. The disclosure should be in 12-month buckets and updated annually. Most vendors will accept this because the bands protect their commercial position while giving you the planning horizon you need. The reason this part matters more than people expect is timing. A model vendor with a three-year fixed compute deal has different incentives in year one than in year three. In year one they will compete on price to fill capacity. In year three they will compete to renew the underlying supply contract and you become the funding source for that renewal. If your contract gives you no visibility into where the supply cycle is, you cannot time your own renewal to land on the right side of it.

Part three. The pass-through trigger. A line that defines a price-change ceiling tied to compute-cost movement. The structure I have been recommending is a 10 percent year-on-year cap on price increases attributable to compute, with anything above that triggering a 90-day renegotiation window. The cap is not aggressive. It just removes the option of a one-sided repricing if the underlying supply contract gets repriced. Without this clause, your unit economics are derivative of someone else's contract. With it, you have a known maximum exposure. The 90-day window is the part that does the real work. Most vendor repricings get sprung on buyers with 30 days of notice, often in budget season when there is no time to source an alternative. A 90-day window gives your procurement function the space to run a parallel evaluation, pull a second vendor into the conversation, and protect your bargaining position. Vendors who refuse the trigger are telling you something useful about how they plan to behave in a tight compute market.

Read the three parts together and you see the shape. Disclosure, duration, pass-through. Each one is small. Together they convert your AI vendor relationship from a black box that absorbs whatever the compute layer does into a glass box where the procurement function can see the same risks the board is being asked to underwrite.

A diagram helps.

                    Model Vendor (you pay)
                            |
                            v
              +-------------+-------------+
              |   Compute Clause Surface  |
              |                           |
              |  1. Disclosure right      |
              |  2. Duration disclosure   |
              |  3. Pass-through trigger  |
              +-------------+-------------+
                            |
                            v
                Compute Supplier (you do not pay)
                            |
                            v
                  Data center, silicon, energy

Above the dotted line, you have a single counterparty and a recurring invoice. Below the line, you have an opaque set of supply commitments that price your invoice. The compute clause is the surface that makes the below-the-line decisions visible.

If you do nothing else with this essay, do three things this week.

Pull every active AI vendor contract over 250,000 dollars annualised value. Have your procurement lead read each one for any language that touches compute supply, repricing, or cost-component disclosure. If a contract has zero language on any of the three, flag it for amendment at next renewal. You are not redlining today. You are building the inventory of contracts that need the clause.
Schedule one 30-minute call with each of your top three AI vendors. Ask them, on a recorded call with your procurement lead present, the three compute clause questions in plain language. Most will answer. The ones that refuse have given you information that should change your concentration risk. Document the answers. The audit trail is the deliverable.
Add a single line to your next board AI update. Name the compute supplier sitting behind each of your top three AI vendors and the duration band of their primary compute commitment. If you cannot fill in the line, that is the work for the next 60 days. The board does not need the clause itself. The board needs the executive to demonstrate that the question is being tracked.

These three actions cost nothing beyond procurement time and one conversation per vendor. They are doable inside a single week. They are also the exact preparation work that separates the AI vendor review of 2026 from the AI vendor review of 2025.

One last thought before you close this tab. The temptation right now is to read every AI infrastructure headline as someone else's problem. SpaceX, Anthropic, Nvidia, OpenAI. These are the names in the filings. Your name is not in the filing. It feels like watching weather on a screen. The reality is that every one of those filings flows downstream into the contract you signed for an AI assistant for your sales team, the contract you are about to sign for a co-pilot for your engineers, and the contract your CFO is about to be asked to underwrite for a multi-year platform deal. The compute clause is the simplest mechanism I know to make sure the weather upstream becomes information for the people downstream who actually have to plan for it. It is also the single change to your vendor template that costs the least and protects the most. Start there this week. The harder conversations get easier once the disclosure rights are in the paper.

While you are here, the back catalogue has more on this.

Common Questions