AMD said on August 6 that it has agreed to acquire Taalas, a Toronto startup that builds chips with an AI model's weights physically wired into the silicon rather than stored in memory the chip has to fetch from — a bet that narrow, model-specific hardware can beat general-purpose GPUs on inference speed and power draw. Financial terms were not disclosed, and the deal is expected to close in the fourth quarter of 2026 subject to regulatory approval.
Taalas, founded in 2023 by former Tenstorrent chief executive Ljubisa Bajic, uses a specialized design process to hardwire a model's math, routing and parameters directly into CMOS logic, rather than routing data through high-bandwidth memory the way conventional accelerators do. The tradeoff is flexibility: a Taalas chip runs only the single model it was built for, and a new model means a new chip. In return, the company says, it gets speed — its HC1 chip served a small version of Meta's Llama 3.1 model at roughly 17,000 tokens per second, which Taalas has said is 73 times the throughput of Nvidia's H200 chip at roughly a tenth of the power draw.
The company says it can go from a finished model to a working chip design in about two months using in-house tools, since only a portion of a chip's layers need to change between designs. A follow-on chip, the HC2, is aimed at models around 20 billion parameters.
AMD's third AI deal in nine months
AMD plans to fold Taalas's designs in alongside its Instinct GPUs inside its Helios server racks and Epyc-based systems, managed through its ROCm software stack. The acquisition "strengthens our long-term AI roadmap with differentiated inference technology," the company said in a statement announcing the deal, framing AI inference — running trained models rather than training them — as one of the fastest-growing segments of the AI hardware market. It is AMD's third AI-related purchase in nine months, following the MK1 deal in November and memory-optimization startup Mext in June.
The deal puts AMD in a similar position to Nvidia, which reportedly struck a roughly $20 billion licensing arrangement with chip startup Groq in December and unveiled its own inference-focused processor in March. Taalas had raised about $219 million in venture funding, including a $169 million round in February from investors including Quiet Capital, Fidelity and investor Pierre Lamond, before agreeing to the AMD deal.
AMD shares rose modestly on the news. Whether model-specific silicon can find a durable place in a market still built around flexible, general-purpose GPUs will depend on how quickly AI labs iterate on their largest models, since every new version means Taalas's customers effectively need new chips.
