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

AMD has recently acquired AI chip startup Taalas in a bid to enhance its AI hardware capabilities and challenge Nvidia's dominance in the industry. The acquisition, announced at market close on Thursday, is seen as part of a larger strategy to deliver faster and more cost-effective inference services for AI agents, similar to Nvidia's $20 billion deal with Groq last December.

Taalas, founded in 2023 and based in Toronto, specializes in a unique approach to inference, etching model weights directly into silicon, which the startup claims can boost performance by an order of magnitude or more. The deal's terms remain undisclosed, but AMD is expected to integrate Taalas' technology with its Instinct-based Helios racks, potentially creating a disaggregated architecture where compute-intensive tasks are handled by GPUs and token generation is offloaded to Taalas-based accelerators.

While this technology offers high-speed inference, it also presents challenges, such as the inability to easily update models without re-spinning the chips. However, Taalas estimates that updating a model's weights only requires modifying two layers of metal, making the process relatively less costly compared to training a frontier model.

The acquisition is expected to close in the fourth quarter, pending regulatory approval, and could have significant implications for the AI industry, potentially enabling AI model developers, their infrastructure providers, and a few inference providers to leverage Taalas' technology for improved performance and efficiency.