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

Advanced Micro Devices announced its acquisition of Toronto-based startup Taalas on August 6, 2026, to integrate its model-in-silicon technology into AMD's Instinct and Helios rack-scale systems. Taalas etches AI model weights directly onto silicon, achieving inference speeds up to 48 times faster than traditional GPUs. While this approach sacrifices flexibility by freezing the model on the chip, it targets the high-volume, low-latency inference market where AMD is actively challenging Nvidia's dominance.

AMD acquisition of Taalas

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

Model-in-silicon inference technology

• ▪Taalas's first test chip, the HC1, was manufactured on Taiwan Semiconductor Manufacturing Company's 6-nanometer process and served Meta's Llama 3.1 8B model at 16,960 tokens per second in initial benchmarks.
• ▪Taalas's processors are comprised 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.
• ▪Taalas designs model-specific integrated circuits that etch AI model weights directly into silicon, eliminating the reliance on High Bandwidth Memory to store model weights.
• ▪Taalas is developing its second-generation HC2 chip, which aims to support up to 20 billion parameters on a single accelerator.

AI inference chip competition

• ▪The acquisition of Taalas by Advanced Micro Devices follows Nvidia's $20 billion licensing and asset acquisition deal with high-performance AI chip designer Groq in December 2025.
• ▪Advanced Micro Devices CEO Lisa Su stated that approximately 60% of AI computing power will be used for inference workloads by 2026.

AMD rack-scale AI strategy

• ▪Advanced Micro Devices plans to integrate Taalas's chips and technology into its Instinct GPU roadmap and its Helios rack-scale systems.
• ▪Advanced Micro Devices has built its rack-scale AI capabilities through the 2024 acquisitions of AI model developer Silo AI for $665 million and server manufacturer ZT Systems for $4.9 billion.
• ▪Advanced Micro Devices plans to deploy a disaggregated architecture where compute-heavy prompt processing is handled by Instinct GPUs and token generation is offloaded to Taalas accelerators.

Performance versus flexibility tradeoffs

• ▪Taalas CEO Ljubisa Bajic stated that the company's platform can realize a previously unseen AI model in custom hardware in only two months.
• ▪Hardwiring AI models into silicon offers high-speed, low-latency inference but sacrifices flexibility, as any model update larger than a LoRA adapter requires a physical chip re-spin.
• ▪Taalas claims that updating a hardwired model does not require starting from scratch, as only two metal layers of the chip need to be changed to implement a new model.