AMD buys Taalas to etch model weights into metal layers.

AMD agreed to acquire Taalas on August 6, adding chips etched for a single model's weights to its lineup, per AMD's announcement.

The move targets inference cost, not raw compute. Taalas says its HC1 chip served Llama 3.1 8B at 16,960 tokens per second. That's a claimed 48 times the throughput of Nvidia GPUs and 8.5 times Cerebras's chips on the same task. For AI-native software priced below its own inference bill, that ratio is the gap between a viable product and a subsidized one.

Taalas builds what it calls Hardcore Models. Its chips are physically wired to one model's weights, finished by completing only two of a chip's more than 100 metal layers. Tape-out takes around two months, according to AMD. The tradeoff is fixed: a Taalas chip can't switch models without a new production run.

Terms weren't disclosed. The deal is expected to close in the fourth quarter. After that, Taalas's Toronto-designed silicon runs in AMD's Helios racks alongside Instinct GPUs, under AMD's ROCm stack. It's AMD's second inference-specialized chip deal in weeks, following an earlier agreement with Cerebras.

The bet only pays off for the wider market if AMD sells that speedup by the token, beyond its own racks. Both this deal and the Cerebras one target the same line item: the cost of serving a model that's already trained. That's the COGS number squeezing every AI product priced below what it costs to run.