NVIDIA Licenses Groq Inference and Hires Team

This deal showed that low latency inference had become important enough for NVIDIA to buy speed by licensing and hiring, instead of waiting for its own roadmap. Groq had built a purpose made stack for fast token generation through LPUs and GroqCloud, and NVIDIA pulled that capability into Vera Rubin so it could sell a fuller AI factory, from prompt processing on GPUs to fast decode on Groq derived accelerators.1 2 3 4 5

• The structure matters. Groq and NVIDIA signed a non exclusive licensing agreement on December 24, 2025, and Jonathan Ross, Sunny Madra, and other team members moved to NVIDIA. That let NVIDIA absorb key technology and operating talent without fully buying Groq, while Groq stayed independent.1 2 3
• The product logic is concrete. NVIDIA now pairs Vera Rubin for prefill and long context work with Groq 3 LPX for fast token by token generation. In practice, that means a customer can use NVIDIA for the heavy front half of a request, then switch to a specialized engine that streams answers faster and more predictably.4 5
• For SambaNova and Cerebras, this raised the bar. It meant specialist inference advantages were no longer safe if they stayed narrow. Once NVIDIA could package a specialist decode engine inside its own rack, startups needed either a full stack cloud and systems business, or a performance lead big enough that NVIDIA could not close it through partnership.2 4 5 6

Going forward, the market is moving toward split inference systems where one engine handles prompt ingestion and another handles generation. That favors vendors that can ship complete racks, cloud capacity, and software orchestration together. The winners will not just have a faster chip, they will control the whole serving workflow customers actually deploy.4 5 6