AMD Acquires Taalas, Which Bakes AI Model Weights Directly Into Silicon

AMD has acquired Canadian startup Taalas, which bakes neural network weights directly into silicon — allowing the chip to run Llama 3.1-8B at over 16,000 tokens per second per user, several times faster than a conventional GPU. Financial terms were not disclosed; AMD plans to integrate the technology into its Instinct accelerator lineup. Google is pursuing the same approach in parallel with Gemini.

Each chip of this type carries a single hard-baked model — you can't swap it out — but for services with a fixed task (moderation, speech synthesis, embedded assistants) speeds of 16,000 tokens per second per user completely change the economics. This is a signal: the market is moving toward specialized chips purpose-built for a specific AI model.