Taalas
A Toronto chip startup that etches a specific AI model directly into silicon, acquired by AMD in August 2026.
Taalas was founded in Toronto in 2023 and came out of stealth in February 2026 with a contrarian pitch. Rather than build a general chip that can run any model, it burns one model’s architecture and trained weights straight into the hardware. The payoff is speed: a demo chip running Llama 3.1-8B hit more than 16,000 tokens per second per user, far beyond conventional inference hardware. The cost is total inflexibility, since that chip can never run anything else.
AMD announced it was acquiring the company in August 2026, planning to fold the technology into its accelerator roadmap and sell it alongside its Instinct GPUs. The bet behind the deal is that top models now stay in service long enough, and serve enough traffic, to justify dedicating physical silicon to them. Google is reportedly pursuing the same idea for Gemini.