If you run AI models in production, the bill has a shape you’ve probably noticed: a surprising amount of it traces back to memory. Every token a model generates has to be served out of fast memory sitting right next to the GPU, and that memory — High Bandwidth Memory, or HBM — is scarce,…