Groq AI Inference: LPU Cloud, NVIDIA License, Company After Deal

Selling inference speed, not training runs

Groq was founded in 2016 by Jonathan Ross, an engineer who had worked on Google's original Tensor Processing Unit, and it built its own chip design rather than buying GPUs. The chip is the LPU, for Language Processing Unit, purpose-built for running trained models rather than training them, which is the workload called inference. Groq sells access to that hardware through GroqCloud, a cloud platform it says more than five million developers use, pushing trillions of tokens through it every week. [4]

Nvidia licenses the technology, hires the team

On December 24, 2025, Nvidia and Groq announced a non-exclusive licensing agreement covering Groq's inference technology. Founder Jonathan Ross, president Sunny Madra, and a portion of the engineering team moved to Nvidia to help scale the licensed technology there. Reporting at the time put the value of the arrangement at roughly $20 billion, a figure neither company confirmed. Groq itself continued to exist as an independent company, with GroqCloud stated to keep operating without interruption. [2][3]

Groq's chief financial officer at the time, Simon Edwards, stepped into the chief executive role as part of the same announcement. He led the company through roughly three months before moving on to become chief financial officer at Bloom Energy in March 2026, at which point Adam Winter, who had joined Groq in 2024 to run its international business, became chief executive. [2][4]

A fresh cap table for the company that stayed

The Groq that remained after the Nvidia deal raised new money twice within a few months: $650 million in June 2026 led by Disruptive and Infinitum, and $350 million more in August 2026, described by Groq as a Series A round for its inference-cloud business, at a $3.5 billion valuation with Nvidia itself among the participants. Calling a raise that size a Series A is unusual, and reflects that the post-deal Groq is being financed and described as a narrower business, an AI inference cloud operator, rather than the chip-design company it was before December 2025. [5][6]

Groq also formalized a partner relationship with Nvidia separate from the licensing deal: it became a certified Nvidia Cloud Partner on August 12, 2026, a status Nvidia grants operators that meet its reference architecture and operational standards. Twelve days later, Nvidia announced the Groq 3 LPX, an inference accelerator built on Nvidia's Vera Rubin platform that uses the Groq and LPU names under license from Groq, Inc., with Groq itself named as an early adopter and the cloud provider Nebius named as first to deploy it in production. [7][8]

One company, two bets on inference

Atlas interpretation: The unusual part of this sequence is that Nvidia ended up on both sides of the inference bet at once: it hired away Groq's founder and licensed the underlying technology for chips that now carry the Groq name, while also investing new capital into the Groq that remained. That is consistent with treating fast inference as a market worth owning outright and worth having a well-funded independent player compete in, rather than betting on one path alone. [2][8]