NVIDIA’s Groq Licensing Deal and the Inference Market

ZUSAMMENFASSUNG

In 30 Sekunden

Matthew Berman discusses NVIDIA’s December 2025 deal with Groq, the company behind language processing units, and argues that inference is becoming more important as AI usage grows. He speculates that the transaction’s structure was intended to avoid antitrust scrutiny. Groq described the arrangement as a non-exclusive technology license, with founder Jonathan Ross and other team members joining NVIDIA while Groq continued independently. A later NVIDIA filing disclosed $13 billion paid at closing and $4 billion payable within a year, including imputed interest; Berman’s $20 billion figure was a contemporaneous reported estimate.

1License, not a company purchaseBerman opens by calling the deal an acquisition but later explains it is a non-exclusive license. Groq said it would continue independently.
2Groq’s LPUHe discusses Groq’s specialized inference chips and founder Jonathan Ross’s earlier work on Google’s TPU.
3Inference demandBerman predicts that inference will generate most AI revenue as usage scales. That is a forecast, and GPUs are also used for inference.
4Deal structure and speculationThe announced license and employee moves are documented; Berman’s claim that the purpose was to avoid antitrust regulation is speculation.

Das Wichtigste

Groq and NVIDIA announced a non-exclusive inference technology license, while Groq said it would remain an independent company.
Jonathan Ross and other Groq team members joined NVIDIA as part of the arrangement.
A later NVIDIA filing disclosed $13 billion paid at closing and $4 billion payable within a year, including imputed interest.
GPUs support inference as well as training; Groq’s LPU is a specialized inference technology.
Berman’s claims about future revenue and the parties’ antitrust motivations are interpretations, not established facts.

ThemenSemiconductors Inference AI cost & performance AI Strategy

VOLLTEXT

Den bearbeiteten Text nach Kapiteln lesen

Transkript zeigen

Nur das Transkript

Groq’s inference technology

Berman introduces a reported deal between NVIDIA and Groq and describes the LPU. He contrasts specialized inference chips with GPUs.

FOLIE 1 / 2Contemporaneous report on the Groq dealThe article shown uses a $20 billion figure; later NVIDIA disclosures gave a more specific consideration breakdown.

NVIDIA has made a major deal involving Groq, the AI-chip company. Although I initially call it an acquisition, the [companies’ announcement](https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale) describes a non-exclusive license for inference technology, with Groq continuing independently. The roughly $20 billion figure I cite was reported at the time, not a disclosed purchase of the company.

FOLIE 2 / 2Groq LPU illustrationThe video illustrates the specialized chip technology discussed by Berman.

Groq makes specialized language processing units, or LPUs, designed for AI inference. Founder Jonathan Ross previously helped develop Google’s TPU before starting Groq. I contrast these chips with GPUs, but GPUs already run both inference and training; the question is how different architectures perform on particular workloads.

Inference economics and the deal

He argues that inference will become a major revenue driver. He then discusses the non-exclusive license, staff moves, and his view of the transaction structure.

Inference is the stage when a trained model processes an input and generates a response. I expect inference revenue to become increasingly important as AI usage grows, potentially larger than training revenue. That is my forecast about the market, not a result established by the agreement.

The arrangement gives NVIDIA a non-exclusive license to Groq inference technology, and Ross, Groq president Sunny Madra, and other team members are joining NVIDIA. Groq says it continues as an independent company. A non-exclusive license does not mean NVIDIA acquired all Groq intellectual property or equity.

I interpret the structure as an effort to avoid antitrust regulation, but the parties’ motivation is not established by the announcement. In a [later NVIDIA filing](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/R11.htm), the company disclosed $13 billion paid at closing and $4 billion payable within one year, including imputed interest, and said no customer contracts, existing products, or equity interests were purchased. I see the deal as a strategic move toward specialized inference hardware.