What does Taalas actually make?

Taalas designs chips that have an AI model’s parameters and math etched directly into the silicon, rather than stored in memory. Its first chip, the HC1, runs only Meta’s Llama 3.1-8B model but reportedly reaches over 16,000 tokens per second per user.

How much did AMD pay for Taalas?

AMD has not disclosed the financial terms of the acquisition.

Why can't a chip like the HC1 run different AI models?

Because the model’s weights and logic are physically built into the chip’s circuitry rather than loaded from memory, a given chip can only run the exact model it was designed for — trading flexibility for speed and efficiency.

AMD Acquires Taalas, an AI Inference Chip Startup

Q: What will AMD do with Taalas' technology?

AMD plans to integrate Taalas’ model-etched chip design into its accelerator roadmap, building system-level products alongside its Instinct GPUs, EPYC processors, Helios rack-scale platform, and ROCm software.

AMD has signed a definitive agreement to acquire Taalas, a Toronto-based startup that builds chips with AI models physically etched into their silicon, the company announced August 6. Financial terms were not disclosed.

Taalas, founded in 2023, takes an unusual approach to AI inference: instead of building general-purpose processors that shuttle model weights between memory and compute, it hardwires a model’s parameters and math directly into chip logic. The company calls this “the model is the computer.” Its first design, the HC1, runs only Meta’s Llama 3.1-8B model, but reportedly reached over 16,000 tokens per second per user in testing — well above what general-purpose GPUs typically achieve for that workload. The trade-off is inflexibility: a chip built for one model cannot run another.

Taalas was co-founded by Ljubisa Bajic, an AMD alumnus and former CEO of AI chip design firm Tenstorrent. “Joining AMD will give us the scale, engineering resources and global reach to accelerate our innovation,” Bajic said in AMD’s announcement.

AMD said it plans to fold Taalas’ technology into its accelerator roadmap, building system-level products alongside its Instinct GPUs, EPYC processors, Helios rack-scale platform, and ROCm software. “Taalas’ technology and world-class engineering team strengthen our AI portfolio by delivering differentiated inference performance and efficiency,” said Vamsi Boppana, AMD senior vice president.

Part of a broader inference push

The deal follows Anthropic’s own move to build an in-house chip team for Claude, as AI labs and hardware makers look for ways to cut the cost of running models at scale. AMD has also been expanding its hardware footprint through infrastructure deals such as its $14 billion agreement with Core Scientific for AI computing capacity, part of a wider effort to close the gap with Nvidia in AI data centers.

Part of a broader inference push

Frequently asked questions