Compute Agreement

Programs that generate text or images need a great deal of computing work. This computing work is not done by ordinary office computers, but by huge halls full of specialized machines. Hardly any company builds such halls itself, because that costs billions and takes years. Instead, companies rent the machines from someone who already owns them. A compute agreement is the contract covering exactly this rental: who provides how many machines, for how long, and at what price? Such contracts often run for five years or longer and fix the total sum in advance.

Why these contracts move balance sheets

For AI companies, compute is the single largest cost item. It is more expensive than salaries, offices, and advertising combined. Anyone without access to enough machines simply cannot train their model. That’s why these contracts are not an accounting detail, but a matter of existence.

For providers, the calculation runs the other way. A data center costs money before it has a single customer. A firm commitment over several years makes this investment predictable. That is precisely why providers like to announce such deals publicly: they prove that the expensive halls are being put to use.

For investors, these contracts are at the same time a warning sign. Ten billion in revenue is only real money if the customer can actually pay in the end. If that customer is itself still running at a loss and gets its money from investor funding rounds, the commitment is left hanging in the air. Critics then speak of circular deals, because money circulates among a small number of connected companies.

What such a contract regulates

The core is the promised capacity. It is rarely measured in individual machines, but in power demand, for instance in gigawatts. One gigawatt roughly corresponds to the consumption of a major city. This unusual unit has a practical reason: today the limiting factor is not the chip, but the electricity available at the site.

Then there is the timeline. Capacity is usually delivered in stages, because the halls still have to be built. A contract for 2030 thus says something about buildings that don’t yet exist. Delays in power lines or chip deliveries can therefore shift entire contract volumes.

A third point is the customer’s obligation. Often a minimum purchase applies: the customer pays for the capacity even if it doesn’t use it. This works similarly to a phone contract with a base fee, just with a lot more zeros. That protects the provider, but can tie the customer down for years, even if newer technology becomes cheaper.

Well-known deals and what they signal

Such agreements have appeared in the news almost monthly since 2023. OpenAI has signed contracts with Microsoft, Oracle, Amazon, and other providers, each in the tens or hundreds of billions of dollars. Anthropic, in turn, works together with Google and Amazon. Providers' stock prices often react immediately and sharply to such announcements.

You can recognize these deals by certain phrasing. There is talk of a “multi-year partnership,” “strategic collaboration,” or “access to capacity.” What matters here is the distinction between a binding order and a letter of intent. The latter sounds almost the same in the press release, but commits to nothing.

A common mistake is to read the contract sum as annual revenue. Ten billion over seven years is about 1.4 billion per year, and most of it only flows at the end. Anyone wanting to assess reports about these contracts should therefore watch three things: duration, bindingness, and the question of where the customer’s money comes from.