Further reading - Impala AI Documentation

Research — The engine and its measured results.

Cost and utilization

Run It Hot — Why utilization, not electricity price, is the lever on cost per token.

Workload shape

Nobody’s Waiting: Common Use Cases for Async AI — The share of inference where no human is reading the output, and what that changes. Inference for Agentic Workloads Is Different — Why agent traces break the assumptions interactive serving is built on.

Engine internals

DBO, Kernel Crossover, and the Real Hardware Cliff — Where kernel choice stops scaling and what replaces it. Wide-EP failure modes, load balancing, and portability — Expert parallelism at fleet scale, and how it fails.

Building the platform

IaC You Can Reason About — Part 1 of the DevOps series. What Full BYOC Actually Takes — Part 2, on deploying into your own cloud.