Last checked: 7 November 2026
In short: Compute is the processing power needed to train and run AI models, measured in chips, hours of use or energy. It is one of the biggest costs in AI, for suppliers and for the firms that buy from them.
Table of Contents
How it works
- Training. Building a model uses huge amounts of compute over weeks or months, often on thousands of chips.
- Inference. Each time someone asks the model a question, it uses a small amount of compute to produce the answer.
- Supply. Cloud firms build data centres full of chips and sell that capacity by the hour or on long contracts.
Where you’ll meet it
Every ChatGPT or Microsoft Copilot reply uses compute in a data centre. You see it as usage charges on AWS, Microsoft Azure or Google Cloud bills, and as limits on how many messages a plan allows.
What it means for your business
- Usage drives cost. Longer prompts, larger models and AI agents that take many steps use more compute and cost more.
- Pick the smallest model that works. A smaller, cheaper model often handles routine tasks such as sorting emails just as well.
- Watch supplier pricing. Compute costs flow into AI prices, so build some headroom into budgets for price changes.
Watch out for
- Bills that jump. Pay-as-you-go AI can cost far more than expected when staff use it heavily. Set spending limits.
- Capacity limits. At busy times suppliers may slow responses or cap usage, which can hold up work.
Example
Oracle. In July 2025 OpenAI said it and Oracle would develop an additional 4.5 gigawatts of data centre capacity for its Stargate project. In September 2025 the Wall Street Journal reported that OpenAI had signed a contract to buy about $300 billion of computing power from Oracle over roughly five years, starting in 2027. That made it one of the largest cloud contracts ever reported.
Compute vs GPU
| Compute | GPU | |
|---|---|---|
| What it is | Processing power as a resource | A chip that provides it |
| How it’s measured | Chip hours, energy or cost | Number and type of chips |
| Who thinks about it | Budget holders and suppliers | Engineers and data centre teams |
Frequently asked questions
How can a small business cut its AI compute costs?
Use the smallest model that does the job, keep prompts short, and set monthly spending caps. Review usage reports to spot staff or tools that use far more than expected.
Why do AI companies spend so much on compute?
Larger models trained with more compute have tended to perform better. Companies also need capacity to serve millions of users at once, which requires vast data centres.
Will compute keep getting cheaper?
The cost per task has fallen as chips and models improve. Total spending still rises, because firms use AI for more work and newer models take more steps per answer.