Groq raised $650M to rebuild the company Nvidia gutted
This is one of the stranger comeback stories in AI hardware. Last December, Nvidia did a $20 billion not-acqui-hire of Groq. It licensed Groq's chip IP and hired away founder-CEO Jonathan Ross, president Sunny Madra, and a chunk of senior staff, without actually buying the company. The investors reportedly made out great. The company left behind had to figure out what it even is anymore. On Monday it answered: raise $650 million, hire, and pivot.
The round was led by Disruptive and Infinitum, with existing backers reinvesting. Groq's new identity is its neocloud business, selling fast, cheap inference rather than just chips. It now runs 13 data centers across North America, Europe, the Middle East, and APAC, and says it serves over five million developers and thousands of AI companies. So the talent and the IP went to Nvidia, but the customer base and the cloud footprint stayed, and that's what the new money is betting on.
Why it matters for agents: inference speed is becoming the axis that decides whether agentic workflows are even usable. An agent that loops, calls tools, and re-plans is doing dozens of model calls per task, and latency compounds fast. Groq's whole pitch was always speed. The genuinely interesting question is whether a company can keep that edge after the exact people who built it walked across the street to the most powerful chipmaker on earth. groq.com
← Back to all articles
The round was led by Disruptive and Infinitum, with existing backers reinvesting. Groq's new identity is its neocloud business, selling fast, cheap inference rather than just chips. It now runs 13 data centers across North America, Europe, the Middle East, and APAC, and says it serves over five million developers and thousands of AI companies. So the talent and the IP went to Nvidia, but the customer base and the cloud footprint stayed, and that's what the new money is betting on.
Why it matters for agents: inference speed is becoming the axis that decides whether agentic workflows are even usable. An agent that loops, calls tools, and re-plans is doing dozens of model calls per task, and latency compounds fast. Groq's whole pitch was always speed. The genuinely interesting question is whether a company can keep that edge after the exact people who built it walked across the street to the most powerful chipmaker on earth. groq.com
Comments