AMD Buys Taalas to Etch AI Models Directly Into Silicon

AMD is acquiring Toronto-based Taalas, a startup that hardwires AI models into specialized inference chips, to strengthen its challenge to Nvidia. Financial terms were not disclosed; Taalas had raised about $219 million.

Key Takeaways

AMD is acquiring Toronto-based Taalas, a startup that hardwires AI models into specialized inference chips, to strengthen its challenge to Nvidia.
Financial terms were not disclosed; Taalas had raised about $219 million.

AI synthesis3 sources6

The AI chip war is moving from the training floor to the inference engine — and AMD is buying its way in.

The company announced Thursday it has reached a definitive agreement to acquire Taalas, a Toronto-based startup founded in 2023 that builds chips with AI models hardwired directly into the silicon. Rather than executing model weights in software on general-purpose accelerators, Taalas' approach embeds elements of AI workloads into the hardware itself, promising dramatically lower latency and energy consumption for running trained models.

Financial terms were not disclosed. Taalas had raised about $219 million from investors.

The deal underscores a strategic shift in the chip industry: as frontier models get bigger, running them for billions of users — the inference side of the business — has become a battleground almost as important as training. AMD plans to integrate Taalas' technology into its accelerator roadmap alongside its Instinct GPUs, EPYC CPUs and ROCm software stack.

Nvidia still dominates AI accelerators by a wide margin. But AMD has increasingly used acquisitions to build a broader hardware-and-software platform rather than chasing Nvidia head-on with GPUs alone — and etching models into silicon may be its most distinctive card yet.