AMD’s acquisition of Taalas brings a different kind of inference silicon into its portfolio: chips whose logic is finalized around a single AI model. The August 6 deal, whose terms were not disclosed, adds the Toronto startup’s engineering team, led by co-founder and former Tenstorrent CEO Ljubisa Bajic, to AMD’s expanding accelerator group.
Taalas, founded in 2023, customizes only the final metal layers of a chip once a model’s weights are locked. That approach lets it merge storage and compute on one die, doing away with high-bandwidth memory, advanced packaging, 3D stacking, and liquid cooling. Its debut product, hard-wired to Meta’s Llama 3.1 8B, is claimed to deliver 17,000 tokens per second per user at 20 times lower build cost and 10 times lower power, though those figures are vendor-reported and the aggressive 3-bit quantization used to reach them trades away some output quality.
AMD plans to pair the technology with its Instinct GPUs, Helios rackscale systems, EPYC CPUs, and ROCm software. The deal lands two weeks after AMD unveiled its MI400 Series and announced up to 2 gigawatts of Instinct MI450 deployment for Anthropic, on top of a 6-gigawatt OpenAI commitment from October 2025.
Investors in Taalas included Quiet Capital, Fidelity, and chip-industry veteran Pierre Lamond, with total funding of $219 million. AMD has long had Canadian roots, tracing back to its 2006 purchase of ATI in Toronto.
The transaction still needs customary closing conditions and regulatory approvals before the technology formally enters AMD’s roadmap.
