Groq
San Francisco, CA
Early Stage
AI inference infrastructure
2020
60× Entry valuation multiple • $20B NVIDIA acquisition • 5 yrs Thesis horizon
Groq builds custom silicon for machine learning inference, specifically tensor streaming processors (TSPs) designed to run neural networks faster and more efficiently than GPUs. Founded by Jonathan Ross and others from the Google Brain team, Groq’s hardware is purpose-built for the latency demands of large language model inference at scale. In 2025, NVIDIA entered a major licensing and talent agreement with Groq, reportedly involving approximately $20B in assets, while Groq remained an independent company. The deal validated Groq’s strategic position in AI inference infrastructure.
In 2020, there was no consensus thesis for AI inference infrastructure. The generative AI wave had not arrived. Most investors were focused on the training side of the compute stack. Groq had no revenue curve that conventional investors could point to. What our network was telling us through ongoing conversations with product leaders across enterprise technology, was something no revenue dashboard could show.
CPOs and product leaders at AI-native companies consistently flagging inference latency as their primary product constraint, 18 months before it became a widely discussed infrastructure problem
Companies building products on top of AI models already hitting latency constraints that would only worsen as model usage scaled
Product practitioners with deep ML backgrounds in our network identifying TSP architecture as structurally superior for inference workloads compared to GPU-based alternatives
Jonathan Ross and team as Google Brain alumni with direct experience building the first TPU, ie the precise pedigree our network recognized as category-defining before the category existed
We closed the deal before there was a consensus thesis. Then the thesis formed around the investment. By the time NVIDIA moved in 2025, the inference category had become the most contested space in AI infrastructure. We had been in the investment for 5 years.
The network validated the technical thesis through CPO and product leader conversations that confirmed inference would be the AI bottleneck before any financial metric could confirm it.
The Groq exit unlocked over $50M in co-investment capacity for subsequent portfolio rounds and demonstrated the full power of the Thesis Frontier approach: we invest before consensus, the thesis forms around the company, and the exits reflect strategic necessity rather than financial math. NVIDIA did not acquire Groq because the revenue justified the price. They acquired because the alternative of not owning the inference infrastructure layer, was worse.
