I almost lost $10M on Groq... Then This Happened
An investor shares the story of investing $10M in Groq, a chip company founded by Google's TPU inventor Jonathan Ross, which struggled for 7 years without product-market fit before ultimately being acquired by Nvidia for $20 billion in 2025.
Summary
The speaker recounts their investment journey with Groq (GRQ), which began in 2015 when they became fascinated by Google's announcement of their Tensor Processing Unit (TPU). Through a friend's introduction, they met Jonathan Ross, the TPU's inventor at Google, who had an unconventional background - not graduating high school, getting a GED, working at a high-frequency trading firm, and eventually creating the TPU during his 20% time at Google. Despite the positive reception, the speaker invested approximately $10 million for nearly a third of the company. The early years were challenging, with the team struggling to recruit talent (eventually convincing 20 original TPU engineers to join), failing to secure government contracts, and having no customers for their first chip. The company endured seven years without product-market fit until the LLM revolution transformed their fortunes. Their chip proved perfectly suited for LLM inference, offering exceptional cost and speed advantages. The breakthrough came in May 2025 when Nvidia introduced Envy Link Fusion, allowing their chips to communicate with others. Groq's team experimented with this technology, leading to a conversation between the speaker and Jensen Huang. This ultimately resulted in Nvidia acquiring Groq for $20 billion by December 2025. The speaker reflects on organizational philosophy, arguing that traditional corporate structures are inefficient and that technical genius should drive company organization rather than conventional hierarchy.
Key Insights
- Jonathan Ross invented Google's TPU during his 20% time after an unconventional path from not graduating high school to working at a high-frequency trading firm
- Groq struggled for seven years without product-market fit, trying government contracts and various applications before the LLM revolution made their chip perfectly suited for inference
- The speaker argues that traditional organizational charts are vestiges of enterprise software that reward politicians rather than technical genius, citing Jensen Huang's 60-80 direct reports as a better model
- AI silicon requires different architectures for training versus inference, with Groq's use of SRAM being brilliant for the decode phase of inference but not necessarily optimal for training
- The Nvidia acquisition happened rapidly, with Jensen Huang promising $13 billion in the account by Tuesday after recognizing Groq's value in accelerating Nvidia's GPUs through Envy Link Fusion
Topics
Full transcript available for MurmurCast members
Sign Up to Access