Most teams know roughly how much compute they need for the next year. Almost nobody knows the exact number for each month. Reserved blocks used to make you guess anyway.
How it works
You commit to an average number of GPUs over a term, from one month to a year. Each month you can move the block up or down by up to 25%, and the discount stays the same.
Grow a block when a big run starts, shrink it during evaluation weeks
Changes take effect at the start of the next billing day
Anything above the block is billed at the on-demand price, by the second
An example
A team commits to 128 Forge 80 GPUs for a year at 35% off. In March they run at 160, in April at 96. Their average stays at 128, so the whole year is billed at the committed rate.
Who it's for
Resizable blocks suit teams with a steady base and a few big runs a year. If your use is very spiky, on demand is still the better deal, and the cost calculator on our pricing page shows where the line is.




