Hardwiring an AI model into a chip means embedding the model's structure into the hardware at the design stage, rather than running it as software on a general-purpose processor. AMD is acquiring Taalas, a startup that has done this with a product already running a compact version of Meta's Llama 3.1. The company is also developing chips for larger and more advanced models.
What the chip actually does
Standard AI chips are general-purpose. A model loads into memory, the chip executes it, and results come back. The hardware carries no permanent relationship with any particular model.
Taalas built its product differently. The chip does not execute Llama 3.1 as a software task loaded from above. The model's structure is embedded in the chip's physical design from the start.
That approach trades range for efficiency. A processor designed around one model's specific architecture can run that model faster and at lower power than a general-purpose chip, which carries the overhead of adaptability. The cost is real: a different model would require a redesigned chip.
Where the Llama 3.1 chip fits
Meta's Llama 3.1 is an open-weight AI model released in multiple sizes. Taalas' current product is built around the smaller end of that family.
A chip designed for a compact model is well suited to edge deployment: hardware that processes AI locally on a device, without routing every request to a remote data center. The smaller the model that can handle a given workload, the more places that kind of chip can go. Companies building AI into products without reliable cloud access, or those managing the recurring cost of high-volume queries to server infrastructure, pull demand in that direction.
Taalas is also developing chips for bigger and more advanced models. No timeline or technical specifications for those products appear in what has been publicly disclosed.
What AMD bought, and what it has not said
AMD competes in the data center AI chip market. The Taalas acquisition gives it a team that has already built a working model-in-silicon product, not a concept on a slide deck. What AMD has not disclosed is how it plans to incorporate that technology into its chip lineup, or which market it intends to target first with it.
The concrete asset in the deal is what Taalas has already built: a chip that runs Meta's Llama 3.1 today.