Groq raises $650 million, targets 200 MW of inference capacity by 2027

Groq has raised $650 million to scale its inference cloud. The round was led by Disruptive and Infinitum, joined by investors who chose to reinvest. The company now runs 13 data centers in North America, Europe, the Middle East and the Asia-Pacific region. More than five million developers use the platform, and trillions of tokens pass through it every week. The target for 2027 is 200 MW of capacity. The release does not name the company's valuation.

Groq narrowed its focus after the NVIDIA deal

The company dates its current phase to December 2025, when it signed a non-exclusive licensing agreement with NVIDIA. At this year's GTC, NVIDIA unveiled its next-generation LPX platform with Groq's inference technology built in. After that, the board and lead investors, together with management, reduced the strategy to a single goal: building the leading cloud for inference.

The breakdown of spending is telling, too. Part of the money will go toward retrofitting existing sites with the latest technology, including NVIDIA's LPX system. For almost ten years the company pushed its own LPU processor. Now it is installing in its data centers a solution built on its own work but released under the NVIDIA brand. Groq says it remains an independent American company and works with data center operators around the world, deploying capacity together with them.

The investors' thesis rests on an estimate of demand. According to the company, inference will require 15–20 times more compute than model training. Most AI clouds, it says, were built for training, and there is no clear leader in inference yet. That is Groq's own position, and the release offers no independent confirmation. Still, the direction is clear: the bet is on response speed and price per token.

New leadership team

The company is led by CEO Adam Winter, with Matt Eng as CFO. Board chair Alex Davis also heads Disruptive. Alan Rice is the new COO. He previously worked at xAI (now SpaceXAI) and Meta Datacenters, and began his career on US Navy nuclear submarines. In a business where the key resources are megawatts and uptime, that background looks like a deliberate choice.

Starting in July, CTO Sinclair Schuller and Chief Product Officer Rakesh Malhotra will join the team. They worked together at Apprenda, a cloud platform that Schuller founded and later sold to Atos. They then created Nuvalence, a software development and digital transformation company that EY acquired in 2024. Malhotra spent about ten years at Microsoft, where he was responsible for cloud products, data center management and enterprise storage. This skill set is geared toward selling inference to enterprise customers.

Key figures from the release:

• New capital: $650 million, round led by Disruptive and Infinitum
• Infrastructure: 13 data centers in North America, Europe, the Middle East and APAC
• Audience: more than 5 million developers and thousands of AI-native companies
• Load: trillions of tokens per week
• Capacity target: 200 MW by the end of 2027
• Hardware: retrofitting sites with NVIDIA's new LPX system
• Company background: founded in 2016, GroqCloud

What developers and infrastructure teams should keep in mind

First: don't treat the 200 MW as a product you can buy today. It is a plan for the end of 2027, and for now it only promises more available capacity and, possibly, looser limits. If your product is hitting quotas or latency limits, it makes sense to check GroqCloud's current limits and regional availability, then compare them on price per million tokens and time to first token using your own prompts.

Second: watch for when NVIDIA's LPX shows up in Groq's cloud and how pricing and performance change. For teams choosing today between GPU instances and specialized inference, it is a chance to get comparable hardware as a ready-made cloud service.

Third: plan for a backup provider. The company has just changed strategy, is replacing part of its leadership and plans rapid expansion. Multi-provider request routing will cost less than an urgent migration if something doesn't go as planned.