
Docs · System
Compute
Compute is bought by the chip-hour, and the same chip delivers different work depending on the system around it. Until there is a common measure, every price is a guess about performance.
What is sold
| Instrument | Unit | Who pays | Rules that decide it |
|---|---|---|---|
| Chip-hour rental | One hour of a named chip | Model developers and companies | The provider's own terms and specification |
| Long-term capacity contract | Dollars and megawatts over years | AI developers and cloud providers | The contract |
| Price index and futures | Index price per chip-hour | Traders, lenders and hedgers | Exchanges announced contracts in 2026. Check whether they trade. |
| Benchmark result | Measured work on a set task | Nobody yet. Used to compare. | MLPerf and similar published benchmarks |
| Facility rating | Energy and water per facility | Regulators and customers who ask | EU data centre reporting; PUE and WUE |
Reading 14
14 of 14
- Standard
MLPerf benchmarks
MLCommons
Published results for the same workloads across different systems. The closest thing compute has to a common measure of work done.
Compute
- Standard
Open Compute Project
Open Compute Project Foundation
Open hardware specifications for servers, racks and cooling. How data centre equipment is described and built to a shared design.
ComputeEnergy
- Standard
Software Carbon Intensity specification
Green Software Foundation
A method for stating the carbon emitted per unit of software work, such as per user or per request.
ComputeCarbon
- Standard
SPEC benchmarks
Standard Performance Evaluation Corporation
Long-running industry benchmarks for servers, including performance per unit of power.
ComputeEnergy
- Standard
The Green Grid
The Green Grid
Source of the data centre measures for power and water use (PUE and WUE).
EnergyWaterCompute
- Regulation
Energy performance of data centres
European Commission
EU rules requiring data centres to report energy and water use, and the planned rating scheme.
ComputeEnergyWater
- Guidance
Center of Expertise for Energy Efficiency in Data Centers
Lawrence Berkeley National Laboratory
Tools, guides and studies on data centre energy and water use from the US national laboratory that tracks it.
ComputeEnergyWater
- Data
Epoch AI
Epoch AI
Open datasets on chips, models and data centres, with trends in compute over time.
Compute
- Data
Green500
TOP500
The same machines ranked by work done per watt.
ComputeEnergy
- Data
TOP500
TOP500
The ranked list of the largest computers, on one benchmark, twice a year since 1993.
Compute
- Data
Uptime Institute
Uptime Institute
Yearly surveys of data centre operators: efficiency, outages and capacity.
ComputeEnergy
- Report
AI Index Report
Stanford HAI
Yearly figures on the cost of training and running models, hardware and investment.
Compute
- Report
Energy and AI
International Energy Agency
Data centres used about 1.5% of the world's electricity in 2024, and the report expects that to more than double by 2030.
EnergyCompute
- Report
Rethinking Load Growth
Nicholas Institute, Duke University
How much new load the US grid could take if large users cut back for a few hours a year. The case for flexible data centres.
ComputeEnergy
Nothing matches. Try fewer filters.