Groq Raises $650 Million and Recasts Itself as an AI Cloud After Its Nvidia Deal
Groq, the artificial-intelligence chip company that licensed its technology to Nvidia Corp. in one of last year’s most unusual deals, has raised $650 million and will refocus its business on selling AI inference as a cloud service, the company said Monday.
The round, announced in California on June 22, comes about six months after Groq agreed to license its Language Processing Unit architecture, known as LPU, to Nvidia for $20 billion in a deal that also saw key Groq engineers and its then-chief executive, Jonathan Ross, join the chip giant. Groq kept its intellectual property, its cloud business and its independence, but the arrangement forced a question: what does an inference company do when the largest chipmaker in the world owns a license to its core technology?
Groq’s answer is to lean into the cloud. The company said it will operate as an AI inference cloud service provider, renting computing capacity to developers who want fast, low-cost responses from large language models. Its GroqCloud platform, which already serves more than two million developers, will become the center of the business, and the new capital will go toward expanding capacity, hiring sales staff and building the infrastructure that running a cloud service requires.
The pivot is a bet on the market’s shape. Training AI models is dominated by Nvidia, and the licensing deal gave Nvidia the rights to Groq’s designs for exactly that kind of work. But running models in production—inference, the part where users actually get answers—is a faster-growing, more fragmented business, and Groq believes its architecture’s speed and efficiency give it an edge in serving responses at scale.
The LPU is an unusual design. Unlike a GPU, which contains thousands of small processing cores and relies on external memory, the LPU is a single, large tensor processor with hundreds of megabytes of memory on the chip itself. That design gives it enormous bandwidth and very low latency for the kind of sequential, token-by-token work that language models do when they generate responses, and it uses a fraction of the power of a comparably fast GPU.
The technology has attracted attention and scrutiny in equal measure. Senators Elizabeth Warren and Josh Blumenthal opened an inquiry into the Nvidia deal in March, questioning whether a transaction structured as a licensing agreement was in fact a way to avoid antitrust review. Groq has said the arrangement is straightforward and that it retains full ownership of its intellectual property, and the company’s new funding round suggests investors accept that account.
The round also marks a shift in Groq’s management and direction. With Ross at Nvidia, the company is now led by Chief Executive Simon Edwards, who has emphasized the cloud business over chip development. Groq is still working on its next-generation hardware—the Groq 3 LPU, unveiled in March and expected to ship later this year—but the company’s future, as Edwards has described it, is defined by the services built on top of the chips, not the chips themselves.
The financial picture is improving. Groq raised $750 million at a $6.9 billion valuation in September, with BlackRock and Samsung among the investors, and the new round comes at a higher valuation, according to people familiar with the matter. The company has said its revenue has grown sharply as developers moved from testing its platform to running production workloads on it.
Groq’s roots explain its stubbornness. The company was founded in 2016 by Jonathan Ross, a former Google engineer who helped design the Tensor Processing Unit, and its early backers included some of the same investors who funded Nvidia’s rise. Meta was among the first companies to deploy Groq’s chips, using them to run inference for its open-source Llama models, a validation that helped establish the company’s reputation for speed. The licensing deal with Nvidia turned that rival relationship into a financial windfall, and the remaining company now has the unusual task of building a business from a position its former rival already understands.
The competitive field is crowded. Nvidia itself offers inference through its own clouds, and Amazon, Microsoft and Google all sell AI computing by the hour. Groq’s pitch is specificity: for developers who need the fastest responses at the lowest cost per token, an LPU-based service offers performance that general-purpose clouds cannot match, and the company says its utilization rates are high because its architecture is designed for exactly the workload it sells.
The $650 million round gives Groq roughly three years of runway, according to people familiar with its plans, time enough to prove whether an independent inference cloud can survive alongside the platforms it once hoped to challenge. The company’s new identity—not a chipmaker that happens to have a cloud, but a cloud built around one distinctive chip—is the clearest sign yet of where its management believes the value lies.
