AMD buys Taalas to hard-code AI models straight into silicon

AMD just bought an AI chip startup whose processors run blindingly fast by permanently etching model weights directly into silicon.

AMD is acquiring Toronto-based startup Taalas, according to a report from The Register. Traditional AI chips constantly shuffle data back and forth between the processor and high-bandwidth memory. Taalas bakes an AI model's parameters straight into the hardware fabric, hitting speeds up to 16,960 tokens per second on an 8-billion-parameter test chip.

Why it matters: Hard-coding a model slashes power consumption and boosts output speeds for token generation. That makes running heavy workloads like code assistants and reasoning agents far cheaper. The tradeoff is permanent hardware lock-in: if you want to update the underlying model, you have to manufacture a new chip.

AMD plans to pair these custom accelerators with its Instinct GPU racks. The GPUs will handle complex prompt processing, while the Taalas chips take over high-speed token generation. Fortunately, updating a model doesn't mean starting from scratch—Taalas says changing just two metal layers is enough to re-spin the chip.

The deal is expected to close in the fourth quarter, provided AMD's customers are ready to commit to their AI models like a permanent tattoo.