AMD acquires AI chip startup Taalas to boost inference performance by etching models into silicon
AMD has acquired AI chip startup Taalas, aiming to dramatically speed up AI inference by etching model weights directly into silicon rather than storing them in conventional memory. The move is part of AMD's ongoing effort to challenge Nvidia's dominance in AI hardware, mirroring the logic behind Nvidia's own $20 billion licensing deal with Groq in December, and is intended to make premium, high-speed inference services for AI agents such as coding assistants faster and cheaper to run.
Founded in Toronto in 2023, Taalas builds what it calls model-specific integrated circuits (MSICs), which bake weights straight into the chip instead of relying on HBM memory. Its first test chip, the HC1, reportedly served Meta's Llama 3.1 8B model at nearly 17,000 tokens a second, far faster than Nvidia and Cerebras hardware, while a second-generation chip with 20 billion parameters is expected this summer. AMD is expected to combine this technology with its Instinct-based Helios racks, though it has not disclosed financial terms of the deal, confirmed as a full acquisition rather than an acquihire.
- AMD buys Taalas, whose chips etch AI model weights into silicon
- Taalas claims up to 17,000 tokens/second, beating Nvidia and Cerebras
- Deal targets faster, cheaper premium AI inference for agents like coding assistants
New here? Start with this
AI systems have two stages: training, when a model learns from data, and inference, when it uses what it has learnt to answer a question or perform a task. Inference powers services such as chatbots and coding assistants, and its speed and cost affect how many people can use them.
Most AI chips fetch a model’s learned numerical settings, called weights, from memory while running. Taalas, a Toronto startup founded in 2023, designs chips that build those settings into the silicon itself; this can make a chip very fast for a particular model, though it may be less suited to running different models.
AMD makes processors and other computing hardware, including chips for AI data centres, and competes with Nvidia, the leading supplier of AI accelerators. Faster, less costly inference could help companies offer AI services at greater scale, while specialised chips may add another option alongside general-purpose hardware.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
AMD's acquisition of Taalas addresses a genuine market need for extreme optimisation in fixed-model, high-volume inference workloads. For organisations running massive scale on stable, predetermined models, embedding weights directly in silicon delivers compelling advantages in latency, throughput, and cost per inference. The demonstrated 17,000 tokens per second performance suggests this is meaningful differentiation against Nvidia precisely where it matters: in the economics of premium inference services.
The case against
Model-specific integrated circuits, despite impressive performance in isolated scenarios, fundamentally conflict with the industry's practical requirements and natural evolution. Most organisations need to serve multiple models and update them frequently, making purpose-built silicon economically irrational and operationally inflexible. The premium economics only apply to organisations with both enormous scale and static model choices—a narrow niche rather than a sustainable basis for competitive advantage as AI models continue to proliferate and evolve.
AI Art Business Celebrity Companies Culture Entertainment Technology