AMD Buys Taalas to Etch Model Weights Into Chips

AMD agreed on August 6 to buy Taalas, a Toronto startup that etches model weights into a chip's metal layers, per AMD's investor relations announcement.

It's AMD's second inference chip deal in weeks, after Cerebras. Taalas' pitch is speed. Its HC1 chip served Llama 3.1 8B at 16,960 tokens per second. Taalas claimed that's 48 times faster than Nvidia GPUs and 8.5 times faster than Cerebras hardware. The real prize is cost. Serving a model you've already trained is the line item eating gross margins on AI products priced below their own inference bill.

Taalas builds what it calls Hardcore Models. Founded in 2023, the company finalizes only two of a chip's 100-plus metal layers per model, cutting tape-out time to about two months. AMD didn't disclose what it paid. The deal closes in the fourth quarter, and AMD plans to run Taalas silicon next to its Instinct GPUs in Helios racks under ROCm.

The trade-off is baked in, literally. A chip specialized to one model's weights only pays off if that model doesn't change. That suits a company running one workhorse model at massive scale, and suits almost nobody else. AMD is betting SaaS builders will make that trade once inference costs get painful enough.