AMD is buying Taalas, a Toronto-based AI chip startup, for an undisclosed sum, according to Outlook Business. The deal is being read across the financial press as AMD's latest attempt to chip away at Nvidia's dominance in AI hardware.
What makes Taalas unusual is its approach. As CNBC reports, the company hardwires AI models directly into its chips — etching model weights into silicon rather than loading them from memory the way a conventional GPU does. CNBC notes that Taalas' current chip runs a small version of Meta's Llama 3.1, and that the company is working on chips for bigger and more advanced models.
The trade-off is specialization for efficiency. An MSN report says Taalas' chips run a single AI model at a tenth of the power of alternatives. That fits AMD's apparent plan: EE Times reports AMD intends to use Taalas chips alongside its own GPUs for AI inference, likely as an accelerator for the decode stage of running large language models.
Inference — actually running a trained model to answer queries — is where much of AI's day-to-day compute cost now sits, and it is the market Outlook Business says AMD is targeting with this purchase.
Markets were not obviously rattled. Barron's noted Nvidia stock had been on a roll ahead of its earnings, while a Stocks Down Under piece asked why Nvidia shares actually jumped after the AMD deal was announced.
Why it matters: if baking models permanently into silicon really delivers order-of-magnitude power savings, it could reshape what AI costs to run — and give the industry a credible alternative to buying ever more general-purpose GPUs.