AMD acquires AI chip startup Taalas to boost inference performance by etching models into silicon

What's happening

AMD is acquiring AI chip startup Taalas to build model-specific inference chips that etch or hardwire models into silicon. Early demos of these model-specific integrated circuits reportedly achieve up to 17,000 tokens per second. Multiple outlets describe the deal as AMD buying Taalas to advance compute solutions for the growing AI inference market and to hardwire AI models into its silicon stack.

Why it's trending

Model-specific, hardwired inference promises much higher token throughput, and AMD is moving quickly to capture that performance edge.

Signal2.0× its usual volume, confirmed across 2 independent source types.