AMD acquires Taalas, betting on AI chips with model weights etched into silicon

AMD said on August 6, 2026 that it has agreed to acquire Taalas, a Toronto startup that etches AI model weights directly into silicon rather than storing them in memory. Financial terms were not disclosed, and AMD said the deal is expected to close in the fourth quarter of 2026, subject to regulatory approval.

The bet behind the deal is that inference — running trained models, the bulk of AI’s day-to-day cost — gets far cheaper when a model lives in the chip itself. Founded in 2023, Taalas builds model-specific integrated circuits that hard-wire a given model’s weights, sidestepping the high-bandwidth memory (HBM) that bottlenecks conventional GPU inference.

Taalas’ first test chip, the HC1, built on TSMC’s 6-nanometer process, reportedly served Meta’s Llama 3.1 8B model at 16,960 tokens per second, a figure the company has published but that has not been independently verified.

AMD plans to pair the technology with its Instinct GPUs, EPYC CPUs, ROCm software and Helios rackscale systems, according to its acquisition announcement. Vamsi Boppana, AMD’s senior vice president of the AI Group, said Taalas’ technology and engineering team “strengthen our AI portfolio by delivering differentiated inference performance and efficiency.” Taalas co-founder and chief executive Ljubisa Bajic said joining AMD gives the startup “the scale, engineering resources and global reach to accelerate our innovation.”

The purchase deepens AMD’s challenge to Nvidia in inference, where custom silicon from startups and cloud providers is proliferating. Whether etched-in-silicon designs can keep pace with models that change every few months is the open question a single test chip does not answer.