AMD Acquires Taalas to Hard-Code AI Models Directly Into Silicon
The Decoder · rss · 2026-08-08
AMD has acquired Canadian startup Taalas, which specializes in hard-coding AI model weights directly into inference chips.
While this approach locks each chip to a single model, it delivers extreme performance. A demo chip achieved over 16,000 tokens per second per user running Llama 3.1-8B. Google is also reportedly developing a similar hardware approach for Gemini.
More from Venture
- Big Tech's AI Revenue Surges, but Capex Remains Higher Fueling Suppliers — Beth_Kindig · 2026-08-08
- AI Inference Shifts to Dense Small Hardware, Overestimating Data Center Demand — Ghost_Pilot_MD · 2026-08-08
- Traditional SaaS and Metered Usage Models May Not Fit LLMs — clarejtbirch · 2026-08-08
- Anthropic Hiring SEO Lead Sparks Debate: Is SEO Dead in the AI Era? — ayushtweetshere · 2026-08-08
- Will AI Get Cheaper? Competition and Compute Costs to Offset Subsidy Loss — intellectronica · 2026-08-08
- Polymarket Forecasts Anthropic 67% Likely to IPO by 2027, OpenAI at 16% — Polymarket · 2026-08-08