AI data center power: what a megawatt buys you

Data center news is full of megawatts. A new campus has hundreds of them. A deal adds dozens more. But a megawatt is a measure of power, and nobody buys power for its own sake. What you want is the AI work it can do.

So what does one megawatt actually buy you? The short answer: it depends on a handful of things you can understand without knowing anything about chips. For the background on why AI draws so much power in the first place, see why AI uses so much electricity.

First, what a megawatt is

A megawatt is a rate of power: a thousand kilowatts, flowing continuously. Keep one megawatt running for an hour and you have used one megawatt-hour of energy. Keep it running all year and you have a steady supply that a site can plan around.

In AI, a site's megawatts describe how much electrical load it can support at once. That sets an upper limit on how many machines can run there. It does not tell you how much useful work comes out the other end.

Where a megawatt goes

Picture the power arriving at a site and follow it.

Some of it never reaches a chip. Power has to be converted and conditioned before equipment can use it, and a little is lost at each step. Then the heat the machines produce has to be removed, which takes fans, pumps and cooling equipment that draw power of their own. A well-designed site keeps these overheads small. A poorly designed one spends a large share of its megawatt just keeping itself running.

Location plays a part here. A site in a cool climate, or one that can reuse its waste heat, usually spends less of its power on cooling than a similar site somewhere hot. The same chips can deliver more work per megawatt simply because of where they sit.

Some of it runs the supporting cast. Networking, storage and control systems all need power. They are essential, but they are not doing the AI work themselves.

The rest feeds the accelerators, the chips that actually run models. This is the part that turns into answers.

So the first question about any megawatt is how much of it ends up doing AI work, and how much is spent getting it there.

Three things that decide the output

Once power reaches the chips, three things decide how much useful work comes out.

How busy the machines are. A chip that is switched on but waiting for requests still draws power and produces nothing. A site whose machines sit half-idle gets half the value from the same megawatt. Keeping capacity busy, by sending it a steady flow of the right work, is a major lever.

How well the work fits the hardware. Running a small, simple task on a large model wastes power. Running a big task on hardware that is not suited to it wastes time. Matching each request to the right model and the right machine gets more answers out of every unit of energy. The same split between training and answering jobs is explained in AI training vs inference.

How good the software is. The same hardware can serve very different amounts of work depending on how requests are batched, scheduled and processed. Improvements here are low-cost gains, because they need no new power at all.

Put together, two sites with exactly the same megawatts can deliver very different amounts of AI work. The megawatt is the ceiling. Utilisation, fit and software decide how close you get to it.

A better question than "how many megawatts?"

A useful measure is how much useful AI output, and sustainable revenue, each deployed watt supports.

That is a more useful question for buyers too. Instead of asking a provider how big their site is, ask:

• How much of the site's power reaches the machines doing the work?
• How busy do those machines run, on average?
• Will my workload be matched to suitable hardware and models, or put wherever there is room?
• Can you show me what was delivered for my requests, so I can judge the output myself?

A provider who can answer these clearly understands what a megawatt is for.

Why small megawatts can matter as much as big ones

The headline deals are measured in hundreds of megawatts. Most real places are not. A business park, a utility substation with spare room, an industrial site with unused power: each might have a few megawatts available today, without waiting years for a new grid connection.

On its own, one such site is modest. But if each one keeps its machines busy, matches work well and runs close to the people using it, the useful output per megawatt can be high. Add many of them together and run them as one system, and small megawatts start to add up to serious capacity.

What we are building

That is the approach behind Starfield Fabric. We're building it to deploy standardized compute close to power, connectivity and demand, then connect it through one programmable network for distributed AI infrastructure. Workloads are designed to be placed by policy, on latency, data residency, model and load, so capacity stays busy with work that fits it. Every workload can generate an auditable record of commitment, placement and delivery, so the output of each megawatt can be checked, not just claimed.

The next time you read about megawatts, ask what they buy. The answer is the work it does.