AMD agrees to acquire Canadian AI inference chip startup Taalas
AMD agreed on August 6 to acquire Toronto-based AI chip startup Taalas, expected to close in Q4 2026, whose technology embeds model weights directly into custom silicon for faster inference.
On August 6, AMD agreed to acquire Toronto-based AI chip startup Taalas. Taalas' approach abandons traditional HBM memory in favor of etching trained model weights directly into mask-ROM fabric on the chip, creating what the industry calls model-specific integrated circuits (MSICs). The approach dramatically accelerates AI inference.
In a February demonstration, Taalas's first test chip HC1, fabbed on TSMC's 6nm process, served Meta Llama 3.1 8B at a blistering 16,960 tokens per second, 48x faster than Nvidia GPUs and 8.5x faster than Cerebras wafer-scale accelerators at the time. The second-generation HC2 chip, expected this summer, targets 20 billion parameters per chip, meaning just 50 accelerators could serve a trillion-parameter model.
The downside is significant: once deployed, the chips are locked to a specific model, and any change larger than a LoRA adapter requires a chip re-spin. Taalas claims only two layers of metal need to be redesigned per new model, dramatically reducing cost and turnaround time. AMD plans to pair Instinct-based Helios racks with Taalas accelerators in a disaggregated architecture where GPUs handle prompt processing and Taalas chips handle token generation.
The deal is expected to close in Q4 2026 subject to regulatory approval. The acquisition is AMD's response to Nvidia's dominance in general-purpose GPU compute, specifically targeting the inference side of AI compute.