Nvidia hollowed out Groq. Here's what happened to the company left behind.

After Nvidia carved up much of Groq, the AI chip startup is now eyeing its next act.

A $3.5 billion AI cloud with real infrastructure — and a harder fight to stand out without the signature chips that once made it a viable Nvidia challenger.

Last December, Nvidia paid roughly $20 billion to license Groq’s signature chip architecture and hire its leadership, stopping short of a full acquisition.

It was one of Nvidia’s biggest deals ever and strengthened its position in inference — the fast-growing market for running AI models, where rivals have gained more ground than in training.

Groq 2.0 has lost some of what once made it stand out.

Some customers said technical support and AI model choice worsened after the deal.

And as inference becomes commoditized, price and speed increasingly matter more than which company provides it.

The unusual structure has also drawn scrutiny, with the Justice Department investigating whether it was designed to skirt antitrust review, The New York Times reported.

Nvidia said in a statement that the deal was “the American system working as designed.” The surviving entity is hardly an empty shell.

Groq has 13 data centers, plans to quadruple capacity next year, and 6 million developers and thousands of companies using its cloud.

It’s also increasingly relying on Nvidia systems — a striking reversal for a startup that once pitched its chips as a faster alternative.

Groq is backed by $1 billion in fresh capital.

It was valued at $3.5 billion in its latest funding round, versus $6.9 billion before the Nvidia deal, though shareholders received payouts in between.

It’s betting on the sheer demand for AI compute. “We think that the needs for inference are great,” a source close to the company said. “It’s sort of like the restaurant business.

People just need food.” Nvidia said Groq “continues to be a separate and independent business.” When the Nvidia deal broke last December, Groq was thrown into limbo, former employees said.

Management played the acquisition talks “very close to the vest,” one said.

The deal came together in eight days, followed by a “mad dash” as former Groq employees became Nvidians overnight, a source close to the company said.

Groq quickly cycled through two CEOs.

Simon Edwards, who had recently joined as chief financial officer, initially took over.

Adam Winter assumed the helm in March, after which the company unveiled its new cloud-focused strategy.

Rather than walking away from its data-center commitments, Groq’s leadership quickly realized it could refit the sites with Nvidia hardware, a source close to the company said.

The deal reshaped Groq’s workforce.

Most of the hardware and software engineering teams — including the majority of the cloud engineering team that built GroqCloud — moved to Nvidia, while data center operations, legal, human resources, finance, and other business functions largely remained to keep the entity running.

The company has since added cloud and infrastructure veterans to its C-suite.

The upheaval didn’t mean employees walked away empty-handed.

Those who held equity received payouts from the deal, according to two former employees.

Before the deal, the company had been known for a “very, very hands-on” approach to customer support, a third former employee said.

Paul Richards, cofounder and CEO of the AI startup Recall, said uncertainty about Groq’s future prompted it to rely less on the company and work with other providers — though roughly a quarter of its inference still runs on Groq. “Every person I spoke to there left the company.

So that obviously was alarming,” he said.

Recall previously had a private Slack with Groq employees, Richards said, and after the deal, “it was crickets.” Richards also said that Groq discontinued many of the AI models Recall used and wasn’t adding the newer models he requested.

Willow cofounder Lawrence Liu saw key support staff move to Nvidia, requiring a “re-initiation period” with their replacements, though the transition was quick and the service remained reliable.

ScreenApp CEO Andre Dean Smith said Groq fell behind on some model updates and never delivered a planned feature.

That prompted his company to shift work elsewhere, though it has since expanded its use of Groq again, he said.

A source close to Groq said that while customers may have lost support contacts who moved to Nvidia, they were connected with replacements, and no complaints reached leadership.

Not everyone noticed the disruption.

Eoin McMillan, founder of Sourcetable, said it was “business as usual” after the deal. “From my end user experience, you would never know anything happened at all,” he said.

Groq is fighting for customers as inference providers proliferate and become harder to distinguish.

Its rivals span inference specialists like Fireworks and Together AI, neoclouds such as CoreWeave, and cloud giants like Amazon and Google.

Groq says it has unique experience running its AI chips, known as language processing units, at scale, and that it will be among the first to deploy Nvidia’s latest inference systems that incorporate them. “You’ve got a billion-dollar-financed inference-native cloud,” said early Groq investor David Levy of Porch Capital. “Shouldn’t this be a big deal?

I think it is.” The harder question is what will make customers choose Groq.

And companies like OpenRouter make it easier for developers to switch between AI providers. “Economics is the best thing they can offer their customers,” said Ashish Nadkarni, an IDC vice president focused on enterprise infrastructure.

Sourcetable’s McMillan said that Groq’s speed once made it an obvious choice for certain workloads.

Since then, competing services have gotten fast enough to narrow the advantage.

Groq doesn’t see that as a problem.

Demand is high enough to mint multiple winners, a source close to the company said: “You don’t necessarily need a moat.” Have a tip?

Contact this reporter via email at gweiss@businessinsider.com or Signal at @geoffweiss.25.

Use a personal email address and a nonwork device; here’s our guide to sharing information securely.

Advertisement