Groq just raised another $350 million, and the pace alone says something about how fast the AI infrastructure race is moving. This is the company’s second funding round in under three months, following a $650 million raise that closed not long ago. Disruptive led the new round as a returning investor, and Groq says Nvidia plans to join later, though the company hasn’t disclosed how much Nvidia intends to put in.
The Nvidia relationship here goes well beyond a simple investment, though. Back in December, Nvidia signed a $20 billion deal to license Groq’s chip technology and bring several of Groq’s leadership team on board. That deal reshaped Groq’s own business almost immediately.
Rather than staying purely a chip company, Groq pivoted toward running a public cloud platform built specifically for AI workloads, powered by the same chip architecture it developed before Nvidia came calling.
Nvidia, meanwhile, used that technology to build something called the Groq 3 LPU, which debuted back in March. Nvidia built it specifically for inference, meaning the work of actually running trained AI models in production rather than training them from scratch.
The chip pairs with Nvidia’s Rubin GPUs in a fairly clever way: the LPU handles the feed-forward network calculations, which store most of an AI model’s learned knowledge, while the Rubin chips handle attention calculations, the part that figures out what matters most in a given prompt. Nvidia argues that splitting the work this way runs more efficiently than dumping everything onto GPUs alone.
Groq packages the Groq 3 LPU into racks of 256 accelerators, which power GroqCloud, the company’s AI-focused cloud platform. Customers who want deep hardware control can run bare-metal environments, while less technical users can lean on GroqStack, a toolkit that automates a lot of the underlying infrastructure work.
Interestingly, Groq now says its platform can support AI training in addition to inference, even though the LPU chip itself doesn’t handle training directly. That workload will presumably run on the Rubin GPUs deployed alongside the LPU racks, coordinated through Nvidia’s Dynamo software, which automatically decides which chip handles which part of a job.
Groq plans to put the new funding toward expanding its cloud footprint, which currently spans 13 data centers globally. The company aims to grow its total capacity from 57 megawatts today to more than 200 megawatts by next year.
