AMD Acquires AI Chip Startup Taalas to Boost Inference Performance with Model-in-Silicon Technology

Aug 6, 2026

Advanced Micro Devices announced an agreement on August 6, 2026, to acquire Toronto-based AI chip startup Taalas for an undisclosed sum. Taalas specializes in model-specific integrated circuits that etch AI model weights directly into silicon, bypassing traditional memory bottlenecks. While this approach sacrifices flexibility—requiring a chip re-spin for major model updates—benchmarks of Taalas' HC1 test chip demonstrated inference speeds 48 times faster than Nvidia's GPUs. AMD plans to integrate this technology into its Instinct roadmap and Helios rack-scale systems to challenge Nvidia's dominance in the rapidly growing AI inference market.

AMD acquisition of Taalas

▪Prior to its acquisition by Advanced Micro Devices, Taalas had raised a total of $219 million in venture funding since its founding in 2023.
▪Advanced Micro Devices announced an agreement to acquire Toronto-based AI chip startup Taalas on August 6, 2026.
▪Advanced Micro Devices did not disclose the purchase price for the Taalas acquisition, which is expected to close in the fourth quarter of 2026 subject to regulatory approval.

Model-in-silicon inference technology

▪Taalas CEO Ljubisa Bajic stated that the company's platform can realize any previously unseen AI model into custom silicon hardware in only two months.
▪Taalas designs model-specific integrated circuits that etch AI model weights directly into silicon rather than relying on High Bandwidth Memory to store them.
▪Taalas' chip architecture consists of a mask-ROM recall fabric where model weights are etched and an SRAM recall fabric where KV caches and fine-tuning adapters are stored.

HC1 chip performance benchmarks

▪The Taalas HC1 chip served Meta's Llama 3.1 8B model 8.5 times faster than Cerebras' accelerators in initial benchmarks.
▪Taalas' first test chip, the HC1, was manufactured on TSMC's 6-nanometer process technology and ran a small version of Meta's Llama 3.1 8B model.
▪In initial benchmarks, the Taalas HC1 chip served Meta's Llama 3.1 8B model at 16,960 tokens per second, which was 48 times faster than Nvidia's GPUs.

AI inference market competition

▪The acquisition of Taalas by Advanced Micro Devices follows Nvidia's $20 billion acquisition of Groq-related assets through licensing and talent recruitment in December 2025.
▪Advanced Micro Devices CEO Lisa Su stated that approximately 60% of AI computing power will be used for inference by 2026.
▪Advanced Micro Devices has expanded its AI capabilities through several acquisitions since 2024, including Silo AI for $665 million and ZT Systems for $4.9 billion.

Flexibility versus performance trade-offs

▪The primary disadvantage of Taalas' technology is a loss of flexibility, as any model change larger than a LoRA adapter requires a costly and time-consuming re-spin of the chip.
▪Taalas claims that updating a model on its silicon does not require starting from scratch, as only two metal layers need to be changed to implement a new model.
▪Taalas is developing a second-generation HC2 chip, scheduled for release in summer 2026, which aims to support up to 20 billion parameters.

Rack-scale system integration

▪Advanced Micro Devices plans to integrate Taalas' technology into its product roadmap, pairing the custom chips with its Instinct GPUs and CPU systems.
▪Advanced Micro Devices has begun shipping its Helios rack-scale AI systems, which are designed to compete with Nvidia's server racks, to customers including Microsoft and Meta.
▪Advanced Micro Devices plans to deploy Taalas' chips in a split architecture where compute-heavy prompt processing is handled by GPUs and token generation is offloaded to Taalas accelerators.