By Hamza Faisal
AMD is paying far more than anyone expected for its bet on a radically different kind of AI chip. A filing disclosed on Wednesday shows the company will pay about $2.9 billion in cash and stock to acquire Taalas, the Toronto startup that hard-wires a finished AI model directly into silicon instead of running it on a general-purpose GPU.
AMD announced the deal in August without disclosing terms. The filing now puts a price on it — and reveals that up to $1.4 billion of the total depends on unspecified conditions being met.
The price is striking for a company of Taalas's size. The startup, founded in 2023, raised just over $225 million in venture funding and was last valued at roughly $750 million in late 2025, according to filings and investor disclosures. AMD is paying nearly four times that figure for a team of around two dozen people.
A chip that is the model
Taalas calls its approach hard-coded inference. Rather than loading a model's weights from memory each time it runs — the step that sets the speed limit for almost every GPU-based AI system — its chips store the weights in the transistors themselves. The model is, in effect, etched into the hardware.
Its first chip, the HC1, is an 815-square-millimetre die built on TSMC's 6-nanometre process, packing 53 billion transistors. In Taalas's own benchmarks, a single HC1 served Meta's Llama 3.1 8B model at about 17,000 tokens per second, which the company says is roughly 48 times faster than Nvidia's GPUs. Those figures are the startup's own and have not been independently verified at production scale.
The trade-off is flexibility. A hard-coded chip runs one model brilliantly and nothing else; anything beyond a small fine-tuning adapter needs new silicon. Taalas says only two metal layers need to change between models, at a fraction of the cost of training a new one.
Why AMD wants it
AMD plans to pair Taalas's technology with its Instinct GPUs and Helios server racks, aiming at the fast-growing market for AI inference — the work of actually running models for users, as opposed to training them. A second-generation chip, the HC2, is designed to hold roughly 20 billion parameters each.
The deal is AMD's third AI acquisition in nine months, following inference-software firm MK1 in November 2025 and memory-optimisation startup Mext in June 2026. It follows Nvidia's $20 billion deal for Groq's assets — a sign purpose-built inference silicon is becoming a genuine second track alongside general-purpose GPUs.
AMD's shares jumped this week on strength in its AI data-centre business and a forecast that beat analyst estimates. The same quarterly filing that revealed the Taalas price also showed AMD now carries $48.8 billion in unconditional purchase commitments — four times the $12.2 billion it reported at the end of 2025 — with $33.4 billion due by the end of 2027 for wafers, memory, substrates and packaging.
A premium for a shortcut
Taalas was founded by Ljubisa Bajic, a former chip architect at both AMD and Nvidia and a co-founder of Tenstorrent, alongside engineers Drago Ignjatovic and Lejla Bajic. The startup stayed in stealth until March 2024 and raised $169 million in February 2026 from backers including Quiet Capital and Fidelity.
Whether the $2.9 billion price proves justified will depend on how quickly AMD can turn a technology demonstrator into a shipping product — and on whether the biggest AI providers decide that locking a model into silicon is worth the loss of flexibility.



