Renewable power and AI: what a buyer can ask

A food label that says "natural" tells you little. A label that lists ingredients, origin and a date tells you a lot more. You can compare it, and someone could check it.

Energy claims for AI work the same way. "Powered by renewables" can describe very different arrangements. This piece explains the common ones in plain words and suggests questions that lead to answers you can verify.

Why the question matters

AI work uses a lot of electricity. (We explain why in Why AI uses so much electricity, in plain words.) Some organisations now report the emissions linked to the services they buy. That can include the computing behind their AI.

So a buyer may need more than a reassuring sentence. They may need something they can put in a report and defend to an auditor.

What "renewable" claims usually mean

Most claims fall into a few broad types. Each is legitimate, but they say different things.

Certificates. In many markets, each unit of renewable electricity can create a certificate. A buyer can purchase certificates to match the electricity it uses. The electricity it actually draws may still come from a mixed grid.

Long-term contracts. A company may sign a contract with a renewable generator, often for many years. This can help new projects get built. The generator may be far from the data center, even in another part of the grid.

Local supply. Some sites draw power from renewable generation that is directly connected or very close. (See "On-site power for data centers, explained".) This is the most physical link, but it rarely covers every hour without storage or grid backup.

Grid mix. Some regions have grids with a high share of renewable or low-carbon generation. A site there uses that mix by default.

Matching: annual or hourly

One detail makes a big difference. It is the time period over which use and renewable supply are matched.

With annual matching, a provider adds up a year of electricity use and buys enough renewable energy to cover the total. The AI work may run at night, when solar output is zero, and still count as covered.

With hourly matching, use and renewable supply are compared hour by hour. This is harder to achieve and harder to claim. It gives a clearer picture of what was actually powering the work at the time.

Neither is dishonest. But a buyer should know which one a claim rests on.

Questions a buyer can ask

These questions are designed to produce answers with dates, places and documents behind them:

1. Which sites will run my workload, and in which grid regions?
2. For each site, what is the main source of electricity: grid, on-site generation, or both?
3. Is your renewable claim based on certificates, long-term contracts, local supply, or grid mix?
4. Is matching done annually, monthly or hourly?
5. Can you give me figures for my own usage, not only a company-wide average?
6. Which reporting standard do your figures follow?
7. Who checks the figures, and can I see the method?

Signs of a useful answer

It names places and periods. "Site X in region Y, for the quarter ending Z" is checkable. "Our network is green" is not.

It separates sources. A good answer says how much comes from each type of claim, rather than blending them.

It is tied to your work. If your workload runs at certain sites, the energy claim should relate to those sites. That requires knowing where your work ran. (We discuss location records in A receipt for AI work.)

It admits limits. A provider that says "we can report annual matching today and are working on hourly" is giving you something you can rely on.

Why location records help

Energy claims are only as good as the link between the work and the power. If a provider cannot say where your AI work ran, it cannot say much about what powered it.

This is where records of placement matter. A record that shows which site ran a workload, and when, lets you match it to that site's energy data. Location in such records is attested by the operator, not cryptographically proven. That is still more useful than a network-wide average.

What we are building

We think energy answers should be as specific as a dated price. (See A price without a date is a rumour: why we publish dated prices.)

Starfield deploys standardized compute close to power, connectivity and demand, then connects it through one programmable network for AI inference. Workloads are placed by policy, on latency, data residency, model and load, and every workload can generate an auditable record of commitment, placement and delivery.

Those records are designed to give a buyer the where and when that any energy claim depends on.

If your organisation reports on the energy behind its AI use, use Request capacity below.