Stephen Nellis
Dec 24 (Reuters) – Nvidia has agreed to license chip technology from startup Groq and hire its chief executive, a veteran of Alphabet’s Google, the company said in a blog post on Wednesday.
Groq focuses on what’s called inference, where already trained AI models respond to user requests. While Nvidia dominates the market for training AI models, it faces more competition in the inference space, with traditional rivals like Advanced Micro Devices and startups like Groq and Cerebras Systems aiming to challenge it.
Groq said Nvidia has agreed to a “non-exclusive” license to Groq’s technology. Groq said founder Jonathan Ross, who helped launch Google’s AI chip program, as well as Groq President Sunny Madra and other members of the engineering team will join Nvidia.
A person close to Nvidia confirmed the licensing agreement.
Groq did not disclose financial details of the deal. CNBC reported earlier that Nvidia had agreed to acquire Groq for $20 billion in cash, but neither Nvidia nor Groq commented on the report. Groq said in its blog post that it will continue to operate as an independent company, with Simon Edwards as CEO, and that its cloud business will continue to operate.
After completing a $750 million financing in September, Groq’s valuation more than doubled to $6.9 billion from $2.8 billion in August last year.
Groq is one of a number of upstarts not using external high-bandwidth memory chips, allowing them to escape the memory crisis affecting the global chip industry. This approach uses a type of on-chip memory called SRAM, which helps speed up interactions with chatbots and other AI models, but also limits the size of the serviceable model.
Groq’s main competitor in this approach is Cerebras Systems, which Reuters reported earlier this month plans to go public as soon as next year. Both Groq and Cerebras have signed big deals in the Middle East.
Nvidia CEO Jensen Huang spent much of his biggest keynote of 2025 arguing that Nvidia would be able to stay ahead as the AI market shifts from training to inference.
(Reporting by Stephen Nellis in San Francisco; Additional reporting by Harshita Mary Varghese in Bengaluru; Editing by Shailesh Kuber and Matthew Lewis)
