Groq Raises $350M Series A at $3.5B Valuation — The Money Goes to NVIDIA Clusters

Groq, the AI infrastructure company built around inference, announced a $350 million Series A on August 17, 2026. The round values the company at $3.5 billion. It was led by the global tech investment firm Disruptive, with planned participation from NVIDIA¹. Together with the $650 million raised in June, recent funding comes to $1 billion¹.

The line worth reading more closely than the headline number is what the company says the money is for. The capital “will support those seeking usage of medium and larger sized clusters of NVIDIA accelerated computing for training and inference”¹. A company known for delivering fast inference on its own silicon is naming NVIDIA hardware as the use of funds in its own press release.

What the deal says

The press release carries a San Francisco dateline, August 17¹. Disruptive led, and its CEO, Alex Davis, also serves as Groq’s Executive Chairman¹.

The headline says “Closes,” but the body adds that “the Series A fundraise is subject to customary closing conditions”¹. NVIDIA’s involvement is likewise phrased as “planned participation,” not as an investment already made¹. Both qualifiers are reasons not to treat the numbers as fully settled.

For scale, the company cites 13 data centers across North America, Europe, the Middle East, and Asia Pacific; more than six million developers; and Fortune 500 enterprises plus thousands of AI-native companies¹. Capacity is currently 54 megawatts, expected to scale to more than 200 megawatts in 2027¹. All of these are Groq’s own figures.

What happens to the LPU

Groq was established in 2016 for one thing — inference — and has delivered it on its own LPU (Language Processing Unit). We covered why the LPU design is fast and the case for AI chip alternatives to NVIDIA back in August 2025, when the framing was about reasons to pick hardware that is not NVIDIA.

Reading this announcement, that framing no longer holds unchanged. What should be stated precisely, though, is that nowhere in the official materials does Groq say it is dropping the LPU.

Five days earlier, on August 12, Groq announced it had joined the NVIDIA Cloud Partner (NCP) program — certification to design, deploy, and operate NVIDIA accelerated computing to NVIDIA’s reference architecture and operational standards². In that post, Groq describes the certification as following its “non-exclusive inference technology licensing agreement” and says it “clears the path to fit out our existing data centers with the latest NVIDIA accelerated computing technologies for inference”².

In the same announcement, Groq calls itself “the only team in the world with hands-on experience operating LPUs in production at scale”². The Series A release likewise includes a line in Alex Davis’s comment that “our team has unmatched experience operating LPUs at scale”¹.

So what the primary sources support is two things: that Groq is moving toward fitting NVIDIA configurations into its existing data centers, and that LPU operating experience remains a pillar of how the company describes itself. TechCrunch characterizes the move as a pivot from manufacturing its own LPU chips to operating cloud infrastructure on NVIDIA³ — but that is the publication’s reading, not something Groq stated.

How the valuation moved

The reported background numbers are worth noting too. According to TechCrunch, Groq was valued at $6.9 billion as of September 2025³. The $3.5 billion figure here is lower.

What happened in between, per the same outlet, is a roughly $20 billion licensing deal, as part of which NVIDIA brought on Groq’s founder and CEO, Jonathan Ross, and others³. In its August 12 announcement, the CEO Groq names is Adam Winter².

A Groq representative told TechCrunch the company does not view the lower valuation as a down round, describing it instead as establishing a new valuation for the “post-Nvidia-licensing-deal version of Groq”³. Both the valuation history and this framing come from TechCrunch’s reporting and are not in the official announcements.

If you use Groq as an API

If Groq sits in your stack as an inference provider, the officially stated answer right now is clear. In the NCP post, Groq writes that “as we look to bring NVIDIA accelerated computing online in the future, customers will get more capacity and the latest inference technology from Groq — without changing a line of code”². By the company’s own account, this is not a change to the interface you use.

What is not stated anywhere in the official materials is where price and speed land. If the reason for choosing Groq was “custom silicon produces economics other vendors cannot,” part of that premise is now worth re-checking. If it was simply “a fast, inexpensive inference API,” nothing published so far gives a reason to revisit the choice.

On the supply side of inference, August 13 also brought OpenAI’s Ultrafast API tier, with Cerebras providing the capacity. Sell your own cloud on specialized hardware, sit behind a major provider’s API, or stack capacity as an NVIDIA certified partner — the same “fast inference” now comes with meaningfully different places to stand as a business. Groq, on the evidence available, has taken the third.