Utilization: The Denominator in Every AI Revenue Claim

Every “revenue per gigawatt” claim is a division with an invisible denominator. The numerator is easy: revenue is disclosed, or estimated from run-rates. The denominator is capacity, and capacity is not the same thing as capacity doing paid work.

Two layers, both lossy

The first is physical. Operators plan around 65–75% schedulable utilization, because peaks, maintenance, cooling limits, network constraints and load shaping take the rest. A gigawatt on the interconnection agreement is not a gigawatt of billable compute.

The second is commercial. Of what is schedulable, 20–30% of a frontier fleet typically runs training, which burns capacity and earns nothing directly — its return arrives later, as a better model. What remains runs paid inference: the only part of the machine that shows up as revenue this quarter.

What it does to the arithmetic

Take a gigawatt, apply 70% schedulable and a 25% training share: 525 megawatts of paid inference, not 1,000. Any revenue-per-gigawatt figure divided by nameplate capacity is therefore overstated by a factor of roughly two, before pricing is even considered. That is the gap between a $20–40 billion theoretical band and the roughly $10 billion labs realized in 2025.

The question to ask

When someone quotes revenue per gigawatt, ask which gigawatt: contracted, built, schedulable, or paid. The difference between the first and the last is where most of the optimism lives, and it is almost never printed on the slide.