Compute

What it means

When you read that a company is "buying compute," it is paying for the computer chips, memory, and storage that AI runs on, along with the data centers that hold them. Dictionary.com defines the noun as "the processing, memory, and storage resources required for a computer, network, or program to function."¹

Think of a restaurant kitchen. The recipe is the AI model, the ingredients are the data, and compute is the stoves and the cooks' time. A bigger kitchen can cook bigger meals and serve more guests at once. When someone in your organization asks an [AI agent](/definitions/ai-agent) to draft a report, that request uses a slice of somebody's kitchen.

Where the word comes from

"Compute" is old as a verb. Dictionary.com traces it to the Late Latin computus and records it in English "in 1375–1425, for an earlier sense."¹ Its use as a noun for a resource is much newer. We could not find a dated first use, so we don't give one.

By 2018 the amount of compute had become something AI labs measured and published. That year, as MIT Technology Review later reported, "OpenAI found that the amount of computational power used to train the largest AI models had doubled every 3.4 months" since 2012.² In 2022 a group of researchers named compute as one of three drivers of progress: "Compute, data, and algorithmic advances are the three fundamental factors that guide the progress of modern Machine Learning."³ Since then the word has moved from labs into law and finance. A 2024 paper observed that "governments and companies have started to leverage compute as a means to govern AI."⁴ Companies now also report compute in their filings.⁵

How others define it

• A dictionary: "the processing, memory, and storage resources required for a computer, network, or program to function."¹
• AI governance researchers: "Computing power, or \"compute,\" is crucial for the development and deployment of artificial intelligence (AI) capabilities."⁴
• A policy primer: The Carnegie Endowment notes that "There is no widely agreed upon definition of compute" and separates three senses. As a measure, it "refers to the number of floating-point operations per second"; as hardware, it "refers to the chips that perform calculations, process data, and execute instructions"; and as a stack, it "consists of three essential components that work together."⁶
• The law: The EU AI Act measures compute without naming it. A general-purpose AI model is "presumed to have high impact capabilities" "when the cumulative amount of computation used for its training measured in floating point operations is greater than" 10^25.⁷ A floating-point operation is a single arithmetic calculation, and 10^25 is a one followed by 25 zeros.

Where they agree and differ

All of these sources treat compute as the physical capacity behind AI. They differ on what kind of thing it is: equipment you own (chips), an amount you use (calculations), or a whole system of hardware, software, and buildings. A news story about "compute" may mean any of the three.

The strongest objection is to treating compute as a measure of what an AI can do or how risky it is. Lennart Heim and Leonie Koessler support compute thresholds, yet they write that "training compute is an imperfect proxy for risk." They argue that thresholds should work "as an initial filter to trigger additional regulatory requirements and scrutiny."⁸ We think that objection is right. Compute tells you how big a system is. It does not tell you whether the system is well directed.

In the news

On October 26, 2026, Fortune reported that Anthropic's filing to go public shows the company "committed to spending between $542 billion and $567 billion on future compute and infrastructure."⁵ Fortune quotes the filing as saying that the company's ability to build and run its models "depends on our access to enormous and increasing quantities of specialized compute." The same report says Anthropic gets compute from "three main cloud service suppliers: Amazon, Google, and Microsoft."⁵

How sure we are: we are highly confident that Fortune reported these figures. We read them through Fortune and could not open the filing itself. Other outlets have reported different totals, so treat the dollar range as Fortune's reading of the filing.

What it means for your organization

Few organizations will ever buy chips. But every AI agent you use runs on compute rented from someone, and often from a small number of large cloud providers. Even Anthropic, one of the largest AI developers, names three main suppliers.⁵ That has three practical effects. Your costs rise with use, because every request and every [token](/definitions/token) draws on compute. You depend on a supply you don't control. And more compute buys capacity, not judgment: a larger system still needs [a person in the middle](/definitions/a-person-in-the-middle) who directs it and answers for it. How much capacity to buy, and for which work, is a leader's decision, the subject of Chapter 8 of The Knowledge Core: Leading a Population of AI Agents, "What Leaders Decide."⁹

One question to ask your team: If our AI providers doubled their prices or cut our capacity next quarter, which agents would we keep running, and who would decide?