Every time a chatbot answers a question, it runs a process called inference: the step where an AI model takes input and produces output. Advanced Micro Devices (Nasdaq: AMD) announced on August 6 that it has agreed to buy Taalas, a chip startup building hardware designed specifically for that workload. Deal terms were not disclosed.
What Taalas actually does
Taalas takes a different approach from general-purpose chips. Instead of running AI models on flexible graphics processors, the company hardwires entire AI models directly into silicon. Its HC1 demonstrator reportedly delivers around 17,000 tokens per second per user when running the Llama 3.1 8B model. Tokens are the word-fragments AI systems process; higher tokens per second means faster, more responsive output.
The company also claims its systems cost 20 times less to build and consume 10 times less power than conventional alternatives, because the design avoids high-bandwidth memory (HBM), advanced packaging, and liquid cooling. Those are expensive components common in today's high-end AI servers.
AMD plans to fold Taalas's technology into its accelerator roadmap and develop system-level products that pair it with AMD Instinct GPUs. Vamsi Boppana, senior vice president of AMD's Artificial Intelligence Group, said the company is building a full-stack AI platform that gives customers flexibility to deploy the right compute solutions for every AI workload.
A string of inference bets
The Taalas deal is the fourth inference-focused acquisition AMD has made in recent months. The company bought MK1, an AI software startup specializing in high-speed inference, in November. It picked up MEXT in June and added FastFlowLM to its AI group in July.
Nvidia (Nasdaq: NVDA), AMD's primary rival, signed a $20 billion deal with Groq, a designer of high-performance AI chips, and has already incorporated that technology into its own roadmap.
The limitation that matters
The same design choice that makes Taalas fast creates a real constraint. Hardwiring a specific AI model into silicon means switching to a newer model requires building new chips. In an industry where model architectures change quickly, that is a meaningful risk.
Investors do not yet have financial terms, revenue projections, or a commercialization timeline. The actual impact will become clearer once Taalas's technology is integrated into AMD's broader product roadmap. As of the end of the first quarter of 2026, 134 hedge funds held stakes in AMD, compared with 275 holding positions in Nvidia.