The Christmas Eve agreement—billed as Nvidia’s biggest deal in its three-decade history—landed at a precarious moment for Groq. Now Nvidia is betting on Groq’s inference-speed tech inside a newly announced chip platform.

By Phoebe Liu, Forbes Staff. With reporting from Iain Martin, Rich Nieva and Anna Tong.

Last winter, Groq cofounder and CEO Jonathan Ross walked into a meeting with Nvidia CEO Jensen Huang with a pitch for the companies’ tech to work together. He now describes the synergy with a logistics analogy: stop building AI data centers as if every workload wants the same hardware. Training is bulk hauling; inference is last-mile delivery. GPUs can do both, but using the 18-wheeler even when you just need a van can be a lot slower. So: Nvidia’s general-purpose GPUs are the big trucks. Groq’s specialized chips—LPUs, or language processing units, designed to run models fast—are the smaller vans. “If you were building out a logistics network for the entire United States, and I told you your two options were all 18-wheelers or just delivery vans, which one would you pick?” Ross said. “The best answer is both.”

Ross wasn’t just pitching a worldview. He wanted Nvidia’s permission to buy around 100,000 Blackwell chips, likely worth billions. Huang grilled him on the technical details, and then the call ended.

 When Huang called back three days later, Ross expected a discussion about his GPU purchase order. Instead, the Nvidia CEO cut to the chase. “We should probably move really fast,” Ross recalled him saying.

Three weeks later, Nvidia announced a $20 billion Christmas Eve deal to license Groq’s product and hire most of its staff. In Silicon Valley terms, it read like a merger without the paperwork: take the team, secure the tech, and get the strategic benefit without inheriting every loose end in the corporate attic, or running into antitrust issues. (Groq’s remaining independent company, an LPU cloud provider, still exists and is growing, Ross—now Nvidia’s chief software architect—told Forbes; former employees said last month the company is expected to sell but hasn’t yet.)

At the time, it didn’t make sense to everyone—because it wasn’t obvious what Nvidia wanted with Groq beyond a very expensive “we’re serious about inference” press cycle. Even Huang’s own explanation stoked bafflement. “They had a very hard time addressing the mainstream part of AI factories,” he said at a conference earlier this year. “But in combination with us, they don’t have to.”

But three months later, Nvidia made their plans clear by dedicating ample airtime at its annual developer conference to Groq. The smaller company didn’t have to become mainstream on its own. It just had to become useful inside the most well-known AI company that already is mainstream.

Everyone had already gotten paid: Ross is taking home an estimated $950 million in cash (after taxes) based on his estimated 9% stake; when his Nvidia stock compensation (likely an outsized share of an estimated $3 billion set aside for employees who went to Nvidia) vests, he’ll be a new billionaire. Ross declined to comment on his net worth, but he’d be worth a lot more if the deal were a stock deal or asset sale rather than a double-taxed, mostly-cash licensing agreement. Other big winners are investor Chamath Palihapitiya’s Social Capital and cofounder Doug Wightman, who held similar estimated stakes to Ross, and Sunny Madra, Groq’s COO and president—who Ross credits for getting the deal done. There’s also the U.S. government: the deal’s structure also means Uncle Sam will likely pocket more than $6 billion in tax revenue, though Nvidia can also cash in on an estimated $3 billion in tax deductions.

At Nvidia’s annual developer conference on Monday, Huang dubbed 2026 the year of AI inference—the process of using AI rather than training it.

He then announced a new product to integrate Groq’s specialized LPU chips—known for doing inference extra fast—with Nvidia’s newest Vera Rubin generation of GPUs. The subtext: the “GPU does everything” era is colliding with the part of the market that cares less about training bragging rights and more about cost, latency, and throughput at scale. This is Nvidia, the patron saint of the GPU monoculture, effectively blessing a heterodox idea: sometimes you want something that isn’t a GPU.

The chips are in “full production,” and are set to start delivering this summer. Nvidia declined to specify how many Groq chips the company plans to make, though it sounds like a lot. “I will say it’s Nvidia-scale,” Ross said. “This is not a pilot.”

Ross said he’s been in “a lot” of customer meetings, though no buyers have yet been announced. “I don’t know if I can confirm any orders right now,” said Dion Harris, Nvidia’s senior director of high-performance computing and AI infrastructure. “It’s early, but there’s lots of interest.”

Regardless, the deal cements the market’s conviction in AI inference chips—even if the Groq-Nvidia partnership appears to have a leg up in distribution, according to other inference chip leaders. The new product, Nvidia Groq 3 LPX, “acknowledges that GPUs alone do not serve the faster-growing inference market well,” wrote Sam Fok, CEO of edge AI inference chip maker femtoAI. He believes cheaper, more accessible inference means will unlock higher demand. “He validated the market we’ve been pushing for a while,” said Sid Sheth, cofounder and CEO of d-Matrix. “The next phase of Nvidia’s growth is going to come from inference.” Groq competitor Cerebras announced deals with OpenAI and Amazon within the last two months. Adds Tenstorrent CEO Jim Keller: Groq “opened the door to heterogeneous computing … I’m sure somebody did some math on it.”

Which is all well and good, because the deal almost didn’t happen. Nvidia’s strategy is to be the platform the ecosystem can’t avoid—and inference chips are the latest add-on. It’s moving from “nice idea” to “Nvidia-supported.”

Groq started in 2016 as an answer—superfast inference—to a question much of the market hadn’t yet bothered to ask. “Groq nearly died many times,” Ross told Forbes in 2024. “We started maybe a little bit too early.”

In 2023—seven years in—Groq generated $3 million in revenue on $88 million in losses. By mid-2024, when it raised $640 million at a $2.8 billion dollar valuation, revenue was still “relatively negligible,” according to Mark Edwards, chief investment officer of Alumni Ventures, which first invested in Groq in 2021.

Two sources told Forbes that around the time of the Nvidia deal Groq’s annual revenue was closer to $100 million, far below its initial 2025 projections of a reported $2 billion, and later a revised $500 million. Its highest-profile deal and main customer was Saudi Arabia’s sovereign wealth fund. Ross said Groq was doing “okay,” hitting “some targets” and in conversation with additional big customers, but declined to comment on 2025 revenue.

Have a tip about Groq, Nvidia or another key character in AI infrastructure? Contact Phoebe Liu at pliu@forbes.com or phoebe.789 on Signal.

In the months leading up to the Nvidia deal, Ross publicly talked up its Saudi-linked partnership, saying it would generate massive revenue for the company. “We did not raise $1.5 billion,” Ross said on the 20VC podcast a year ago. “That’s revenue. That’s actually about 30% of the revenue of OpenAI.”

But a former employee told Forbes the headline number referred to “a cumulative value of the services they were going to be building.” That figure would likely include the total cost of building the data center that would house the Groq chips, the value of the chips themselves, and the value of the computing power the chips would generate (over an unspecified period of time).

Groq’s deal with Saudi Arabia—specifically Aramco Digital, a technology subsidiary of the state-owned oil company—was structured as a revenue share deal, according to three former employees. Groq would sell its AI chips to Aramco for cheap-ish; Aramco would fund the data center to house them; and the two parties would split revenue from the compute the chips generated.

And not every “partnership” looked like a clean, cash-for-product transaction. In one of Groq’s other high-profile partnerships, with data center REIT Equinix in Australia, Groq is the customer, paying Equinix up to $300 million for Groq to house LPUs and in turn sell to its cloud customers. “It felt a lot more like, yes, there was an exchange of funds, but there was a lot of bartering of goods as well,” one source told Forbes.

It’s unclear how these arrangements have continued under the Nvidia umbrella, or to what extent they were part of the Nvidia talks. Ross said it was only a factor insofar as Groq’s independent company is continuing to support its Saudi customers, according to Ross.

The integration Nvidia unveiled on Monday uses Groq’s third generation chip. Their delayed second generation hadn’t ramped up yet, according to Ross, but Groq moved to a new version because it would have taken as long to integrate it with Nvidia as if they started fresh with a third version.

Headwinds aside, Ross insisted that he was confident and dreaming big throughout. “I figured I was gonna die at Groq,” he recalled thinking. “I always told people that we wanted to deliver half of the world’s inference.”

When Nvidia initially proposed the deal, Ross wasn’t sure it was going to work. “We almost didn’t do it.” He said Madra, who is now a vice president at Nvidia, convinced him that it was a good idea. It ended up feeling like the perfect match—both business-wise and culturally, Ross said.

Now the bet moves from deal theory to product reality. It’s still too early to tell how the Nvidia-Groq system—dubbed the LPX—will perform since it’s positioned as an optional integration with Nvidia’s Vera Rubin platform and not yet being used at scale. At Monday’s conference, Huang said it could unlock $300 billion in annual revenue per gigawatt, and that he expects around 25% of GPU workloads to link up with Groq chips.

We were like “kids in a candy store,” Ross said of joining Nvidia. “All of a sudden, we had all of the access to all of the stuff.”

© Forbes