Multi-cloud capacity and reliability

A single region on a single provider is where every inference fleet starts, and for a while it is the right choice. Then three things arrive together. The provider cannot give you more GPUs of the type you need on the morning you need them. Hardware starts failing often enough that it shows up in your weekly review. And a customer, or a regulator, asks what happens when the region goes down, or where their code is processed.

No faster kernel solves these. Deciding in advance how capacity is bought, how much is spare, and where the spare sits does. The numbers here use Kite's chat fleet a year from now, with several times the 12 requests per second peak used elsewhere, when it needs 8 H100s around the clock and 16 during weekday working hours.