A GPU is a general-purpose processor: it can run any AI model, render a 3D scene, or power a video game. That flexibility has a cost — the chip constantly moves model weights between memory and compute cores on every inference request.

Taalas, the Toronto startup AMD acquired this week, took the opposite approach. Its chips hardwire a specific model's weights directly into silicon, eliminating that memory bottleneck entirely. The tradeoff is stark: the chip runs only one model, but runs it extraordinarily fast — reportedly serving Meta's Llama 3.1 8B at nearly 17,000 tokens per second, roughly 48× faster than a comparable Nvidia GPU.

This matters because AI is shifting from training (done once, needs flexibility) to inference (done billions of times daily, needs speed and efficiency). When you're serving millions of users the same model around the clock, a chip purpose-built for exactly that task becomes very attractive — even if it can never be reprogrammed.