NVIDIA pays a reported $20 billion to license Groq's inference chips and hire its founders

Groq designs a Language Processing Unit built around hundreds of megabytes of on-chip SRAM, which lets it serve open models at very high token rates. The company has said its chips run large language models up to ten times faster at a tenth of the energy of GPUs — a vendor claim that no independent lab has reproduced.

On 24 December 2025 Groq announced a non-exclusive licensing agreement with NVIDIA; founder Jonathan Ross, Sunny Madra and other team members moved to NVIDIA, and Simon Edwards became Groq's CEO. CNBC reported the consideration at $20 billion and described it as an acquisition; analysts noted the structure avoids a formal merger review. NVIDIA has since shown a licensed 'Groq 3 LPX' design within its Vera Rubin platform.

Groq's own footprint runs through GroqCloud sites in the US, Dammam in Saudi Arabia — backed by a $1.5 billion Saudi commitment — Helsinki and Sydney, with plans for more than a dozen new data centres. It publishes no environmental data; the index estimates its per-prompt energy at 0.1 Wh, crediting a 2.5× efficiency gain over GPUs rather than the claimed 10×.