AMD Closes Taalas Deal to Add Model-Specific AI Inference Chips

Unite.AI
Read full story· 3 min
AMD finalized its $2.9 billion acquisition of Taalas, a Toronto startup specializing in model-specific AI inference chips that optimize dataflows and reduce bottlenecks. Taalas's HC1 chip hard-wires Llama 3.1 8B model, offering efficient inference without advanced cooling or packaging. AMD plans to integrate this technology with its Instinct GPUs and broader accelerator roadmap.

What happened

AMD completed its acquisition of Taalas, a Toronto-based startup, on November 17, 2026. The deal, valued at about $2.9 billion, brings Taalas’s model-specific AI inference chip technology into AMD’s portfolio. Taalas’s HC1 chip is designed for efficient operation with the Llama 3.1 8B model and does not require advanced cooling or packaging. AMD plans to incorporate Taalas’s technology into its Instinct GPU line and broader accelerator products.

Key details

  • The acquisition was valued at approximately $2.9 billion, including up to $1.4 billion in contingent consideration.
  • Taalas’s HC1 chip is built using TSMC’s 6-nanometer process, with an 815-square-millimeter die and 53 billion transistors.
  • The HC1 chip delivers 17,000 tokens per second per user on Llama 3.1 8B in lab tests.
  • Taalas’s chip design merges storage and computation on a single chip at DRAM-level density, eliminating the need for high-bandwidth memory or advanced packaging.
  • AMD will integrate Taalas’s technology with its Instinct GPUs, EPYC CPUs, ROCm software, and Helios rack-scale systems.

Why it matters

AMD customers in data centers and AI infrastructure will have access to more specialized inference hardware, potentially lowering operational costs and increasing efficiency for specific AI models. The acquisition expands AMD’s engineering capabilities in Canada and broadens its AI hardware offerings, giving clients new options for deploying model-specific solutions without the need for complex cooling or memory systems. Companies using AMD’s AI stack now face new choices between general-purpose and model-specific accelerators for their workloads.


More in Chips & Compute