AMD bought a chip where the model IS the silicon

Every GPU in your inference stack does the same thing on every request: stream the model weights out of high-bandwidth memory, compute, repeat. Taalas, a 24-person Toronto startup founded in 2023, asked what happens if you skip that step entirely. Their HC1 chip etches model weights permanently into silicon — no HBM, no memory bottleneck, no fetching. The model is the hardware.

The result: 16,960 tokens per second on Llama 3.1 8B, which AMD claims is 48x faster than Nvidia GPUs. AMD acquired them on August 6th.

The obvious catch: once a model is cast in silicon, updating it means a re-spin. Taalas softens that — only two metal layers need changing, and they say a new chip takes about two months. Still, buyers need to commit to a model. That narrows the customer list to big inference providers and the model labs themselves.

The architecture question underneath all of this: if inference becomes the dominant AI compute cost in 2026, the GPU's flexibility starts looking like overhead you're paying for on every token.