An E1 engineering study of CPU update encoding, fallback-model saving and retained coordinator memory, with targeted changes and explicit limits.

A benchmark for AI compute usually arrives as one number: images per second, tokens per second, a speed-up. The number may be real, but it answers a narrower question than the one a buyer is asking. The buyer wants to know how long it takes, and what it costs, to get a usable model on the hardware they can actually get. Mixed GPUs, from different vendors, generations and places, give a single number more ways to mislead, including the people who measured it.

The Stanford AI Index 2026 tells a remarkable story about artificial intelligence. It also reveals a growing problem underneath it.