AMD buys Taalas to etch AI models straight into silicon

AMD is acquiring Taalas, a three-year-old Toronto startup doing what GPUs deliberately don't: burning a model's weights directly into transistors. No HBM, no shuttling parameters in and out — just mask-ROM recall fabric for the weights plus SRAM for KV cache and adapters. Its HC1 test chip hit 16,960 tokens/sec on Llama 3.1 8B back in February, a claimed 48x an Nvidia GPU and 8.5x Cerebras. HC2, targeting 20B parameters, is due this summer.

The catch is the whole design: a model-specific IC is welded to one model. New weights mean a respin, though Taalas says only two metal layers change, so iterations aren't full restarts. Terms undisclosed — Taalas raised $169M in February on ~$219M total, and the deal should close in Q4 2026. It's AMD's third AI buy in nine months, after MK1 and Mext.

Read it against December: Nvidia paid roughly $20B for Groq's people and core assets — SRAM-based LPUs, model-agnostic, fast at anything — in a structure so unusual that Senators Warren and Blumenthal are probing whether it dodged merger review. Two opposite bets on the same thesis: inference is where the money moves next. Nvidia bought flexibility. AMD bought commitment.

Why it matters: the chip war just stopped being about training FLOPs and started being about who serves tokens cheapest.