The semiconductor company AMD is acquiring the Canadian chip specialist Taalas, which does not load language models onto graphics processors but instead burns them directly into the circuits of its processors. This was announced by AMD in a press release on August 6, 2026; the acquisition is expected to be completed in the fourth quarter. Financial details of the deal remain confidential.
Chip burns language models instead of loading data from memory
Taalas calls its processors Hardcore Models: Unlike traditional graphics processors, they do not load the weights of an AI model from a separate memory chip but burn them directly into the chip structure. One area of read-only memory holds the model weights, while a second area of random access memory handles intermediate results and session data. As reported by The Register, the first chip, HC1, manufactured by TSMC using a 6-nanometer process, processed Meta’s model Llama 3.1 with eight billion parameters at 16,960 tokens per second in February 2026 – according to its own statements, 48 times faster than comparable Nvidia graphics processors and 8.5 times faster than accelerators from Cerebras; these values have not been independently verified. A successor chip named HC2 is expected to process models with up to 20 billion parameters later this summer. A disadvantage of the approach: Once a model is burned in, it cannot be replaced. According to Taalas, a change requires only two new metal layers instead of a complete chip redesign – according to the company, about one hundred times cheaper than training a new top model.
AMD positions the acquisition against Nvidia’s market leadership
AMD manager Vamsi Boppana stated that Taalas’ technology and its engineering team strengthen the company’s AI portfolio through differentiated inference performance and efficiency. Taalas co-founder and CEO Ljubisa Bajic said that joining AMD provides his company with the scale, development resources, and global reach for faster growth. AMD plans to use the Taalas chips alongside its Instinct graphics processors in the future – for example, in the Helios server racks that the company is already integrating into Microsoft’s Azure Cloud. A split architecture is conceivable: The graphics processors handle the computationally intensive processing of incoming requests, while the specialized Taalas chips then take over the actual text generation. For AMD, the acquisition is another building block in the competition with market leader Nvidia, which remains far ahead in AI accelerators despite such advances. Inference is now considered the fastest-growing expenditure item for AI computing power in the industry – a field in which AMD holds a significantly smaller market share compared to Nvidia’s established GPU platforms. The acquisition is part of a series of similar bets in the industry on specialized inference hardware, such as SambaNova, which recently raised fresh capital at a valuation of eleven billion dollars.
Toronto loses four independent chip start-ups in two years
Taalas was founded in 2023 in Toronto, among others by Bajic and co-founders Lejla Bajic and Drago Ignjatovic – all three previously held leadership positions at the Canadian chip start-up Tenstorrent and some had worked at AMD itself. After exiting stealth mode, Taalas initially raised 50 million dollars from Quiet Capital and investor Pierre Lamond in 2024, followed by another round of 169 million dollars in February 2026 led by Fidelity. As the Canadian tech magazine BetaKit notes, the acquisition fits into a broader trend: In 2025, AMD acquired the Toronto start-up Untether AI, Nvidia bought CentML, and Tenstorrent itself moved its headquarters to California. The Ottawa start-up Hyperlume also went to the US provider Credo in 2025. Canada’s once-celebrated AI chip scene is thus continuing to thin out, while the acquired technologies are being integrated into the product lines of the large US providers. AMD emphasized that the acquisition reflects a continued commitment to retaining and further developing Canadian talent – how many of Taalas’ employees will remain at the Toronto location in the long term remains open.
It remains to be seen whether the coupling of graphics processors and model-specific chips will work as smoothly in practice as outlined – AMD has not yet provided details on the specific product integration. It will also be crucial whether the acquisition passes antitrust scrutiny as planned in the fourth quarter of 2026 and whether the successor chip HC2, announced for summer, actually delivers the promised leap to 20 billion parameters.
