Why compute is bought in years and used in months

A GPU cluster is sold the way a building is sold. You sign for three to five years, you pay whether you use it or not, and a large part of the money goes in up front. In 2026 most contracts longer than three months ask for at least a quarter of the value in advance, and the biggest providers report that most of their deals now prepay half or more of the hardware cost.

A GPU cluster is used the way electricity is used. A training run needs everything for six weeks. A product launch needs double for a month. Inference spikes on Monday morning and sleeps on Sunday. Microsoft's own research clusters averaged about half of their capacity in use across the year.

What the mismatch creates

Two things, on opposite sides of the same market. On one side, months that are paid for and never used. On the other, buyers who need a block for a quarter and cannot get one, because every provider prefers to sell years.

The short block is the scarce thing, so it carries the price. Nebius told its shareholders it prices a three-to-six-month block at three to four times its base rate. Hourly capacity at a neocloud runs about a third above a one-year contract on a B200, and a hyperscaler charges roughly three times the contract rate.

What the open market does

It moves the months. A company with spare compute, usually an AI company, keeps the contract and sells the window it will not use. The buyer gets a block for the quarter it needs, under the hourly rate, without signing for three years. The data centre agrees once, the machines are set up for the buyer, and nothing is transferred on paper.

That is the whole idea. Everything else on this site is how to do it safely.