AMD Buys Startup Taalas To Bake AI Models Straight Into Silicon

AMD is making a big bet on the future of AI hardware. The Santa Clara chip designer announced it has reached a definitive agreement to acquire Toronto-based startup Taalas, a company taking an unconventional approach to inference by burning AI models directly into custom silicon.

Financial terms of the deal were not disclosed, but the strategic intent is clear. As the costs and power demands of running trained AI models (inference) continue to rise, AMD envisions bypassing traditional hardware bottlenecks via specialized inference chips.

Excited to announce our agreement to acquire @taalas_inc. Phenomenal team working at the bleeding edge of AI inference. Looking forward to all we will do together. pic.twitter.com/2WTrb88eWW

— Lisa Su (@LisaSu) August 6, 2026

Most modern AI chips—from NVIDIA GPUs to AMD’s own Instinct accelerators—are general-purpose processors. They largely rely on high-bandwidth memory (HBM) to feed model weights into programmable compute engines every time a query runs. That flexibility lets a single server run thousands of different models, but it incurs massive power and memory bandwidth penalties.

Taalas flips that architecture on its head with a philosophy it calls "the model is the computer." The company's motto is a clever way of saying that instead of treating an AI model as software to be loaded into memory, Taalas uses a specialized design flow to physically hardwire a model's math, routing, and parameters directly into CMOS logic. By stripping out general programmability and storing weights inside the silicon logic itself, the startup bypasses (in part) the need for expensive HBM arrays.

There is a cost, though. The trade-off to this approach is zero flexibility, as a chip hardwired for a specific model can only run that model. However, the potential performance gains are too big to ignore. Earlier this year, Taalas demonstrated a test chip tailored specifically for Llama 3.1-8B that delivered over 16,000 tokens per second per user, blowing past the throughput limits of general-purpose GPUs.

That said, AMD doesn't plan to replace its broad-purpose GPUs with model-specific chips. Instead, the company plans to integrate Taalas’s technology alongside its Instinct GPUs, EPYC server processors, and Helios rackscale platforms as part of its full-stack deployment.

The acquisition also brings key engineering talent back into the AMD fold. Taalas was founded in 2023 by Ljubisa Bajic, an AMD alumnus and former CEO of AI chip design firm Tenstorrent. Bajic and his Toronto-based team will join AMD’s Artificial Intelligence Group under Senior Vice President Vamsi Boppana.

"We founded Taalas to rethink AI inference from the ground up by building the hardware around the model," said Ljubisa Bajic, co-founder and CEO of Taalas. "Our Canada-based team has combined deep technical expertise with a willingness to challenge conventional approaches. Joining AMD will give us the scale, engineering resources and global reach to accelerate our innovation."

AMD also touts this acquisition as further solidifying its R&D presence in Canada, where the company already maintains deep engineering roots stemming from its 2006 ATI acquisition.