AMD acquires a startup that etches AI models directly into silicon

AMD said it had acquired Taalas, a Toronto-based startup founded in 2023 that designs inference chips with a model’s weights etched directly into the silicon rather than loaded and run as software on general-purpose hardware. The companies did not disclose a purchase price.

AMD’s stated rationale is efficiency at inference time: a chip built for one model’s weights can strip out the flexibility that general-purpose accelerators carry for training and multi-model serving, cutting the compute and memory overhead that comes with it. AMD said it plans to fold Taalas’s approach into its accelerator roadmap and pair it with its Instinct GPU line, positioning the acquisition as an addition to a stack that already spans EPYC server CPUs, Helios rack-scale systems and the ROCm software platform. AMD AI group senior vice president Vamsi Boppana said the deal would “strengthen our AI portfolio by delivering differentiated inference performance and efficiency”; Taalas co-founder and chief executive Ljubisa Bajic said joining AMD would give the company “the scale, engineering resources and global reach to accelerate” its work.

The acquisition is a small deal by dollar terms but fits a broader industry move toward specialised inference hardware, as the largest AI labs shift spending from training runs toward serving models at scale to paying users — a shift chipmakers are increasingly racing to capture with purpose-built silicon rather than general-purpose GPUs alone.