Groq Cofounder Explains Whirlwind Deal With Nvidia

追加: 2026-05-13

[00:00:00]Today on Forbes, sometimes you don't want a GPU. Groq's co-founder explains whirlwind deal with Nvidia.

[00:00:09]Last winter, Groq co-founder and CEO Jonathan Ross walked into a meeting with Nvidia CEO Jensen Huang with a pitch to stop building AI data [music] centers as if every workload needs the same hardware. Training is bulk hauling, inference is [music] last-mile delivery.

[00:00:28]Inference is essentially the process of using AI rather than training it. GPUs can do both, but using the 18-wheeler [music] when you just need a van can be slower.

[00:00:38]Nvidia's general-purpose GPUs are the big trucks, while Groq's specialized LPUs [music] or language processing units that are designed to run models fast are the smaller vans.

[00:00:51]When speaking about which one to [music] use, Ross said, "The quote, the best answer is both."

[00:00:57]Ross wanted Nvidia's permission to buy around 100,000 Blackwell chips, [music] likely worth billions.

[00:01:04]Huang grilled him on the technical details, and the meeting ended. Three days later, Huang called back [music] not to discuss the order, but to cut to the chase saying, "We quote, we should probably move really fast."

[00:01:18]Three weeks later, Nvidia announced a $20 [music] 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 or [music] running into antitrust issues.

[00:01:39]Groq's remaining independent company and LPU cloud provider still exists and [music] is growing, said Ross, who is now Nvidia's chief software architect.

[00:01:50]At the time, [music] it wasn't obvious what Nvidia wanted beyond a very expensive signal that it was serious about inference. Even Wong said Groq had, "quote, a very hard time addressing the mainstream [music] part of AI factories, but, quote, in combination with us, they don't have to."

[00:02:10]Months later, Nvidia made its plans clear by dedicating ample airtime at its annual developer conference to Groq.

[00:02:18]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.

[00:02:28]With the deal, Ross [music] is taking home an estimated $950 million in cash after taxes, [music] and with Nvidia stock compensation, he'll be a new billionaire.

[00:02:39]Investor Chamath Palihapitiya's Social Capital held a similar stake to [music] Ross, and Groq's COO and President Sunny Madra, who Ross [music] credits with getting the deal done.

[00:02:51]The deal structure also means the US government will likely collect more than $6 billion in tax revenue, though Nvidia can also cash in on an estimated [music] $3 billion in tax deductions.

[00:03:04]At Nvidia's annual developer conference last month, Wong dubbed 2026 the year [music] of AI inference. He announced a new product integrating Groq's LPUs with Nvidia's newest GPUs. The subtext is the, quote, GPU does [music] everything era is colliding with a market that cares more about costs, latency, and throughput.

[00:03:26]This is Nvidia effectively blessing a heterodox idea that sometimes you want something that isn't a GPU.

[00:03:34]The chips are in full production and set to start delivering this summer. Nvidia declined to specify how many Groq [music] chips the company plans to make, but Ross said, "quote, this is not a pilot." According to Dion Harris, Nvidia's [music] senior director of high-performance computing and AI infrastructure, there is quote lots of interest though no buyers have been confirmed.

[00:03:58]The deal cements the market's [music] conviction in inference chips.

[00:04:01]Competitors like Cerebras, D-Matrix, [music] and Tenstorrent point to the same trend as the next phase of AI growth will come from inference. Nvidia's move shifts inference from quote nice idea to quote Nvidia supported.

[00:04:16]Groq started in 2016 [music] as an answer of superfast inference to a question the market hadn't yet [music] asked. CEO Ross said quote Groq nearly died many times. In 2023 [music] it generated $3 million in revenue on $88 million in losses. By mid-2024, [music] when it raised $640 million at a $2.8 billion valuation, revenue was still quote relatively negligible according to Mark Edwards, chief investment officer of Alumni Ventures, which [music] first invested in Groq in 2021.

[00:04:51]Around the time of the Nvidia deal, sources say revenue was closer to $100 million, far below earlier projections.

[00:04:59]Despite headwinds, Ross insisted he was always thinking big. He said quote I figured I was going to die at Groq. We wanted to deliver half of the world's inference.

[00:05:10]The deal almost didn't happen. Ross said he wasn't sure the chips would integrate well, but it ended up feeling like the perfect match.

[00:05:18]Now the bet moves from deal theory [music] to product reality. It's still too early to tell how the combined systems will perform at scale, but Wong said it could unlock massive revenue and expects a significant portion of GPU workloads to link up with Groq chips.

[00:05:35]For full coverage, check out Phoebe Lo's piece on forbes.com. [music] This is John Palmer from Forbes. Thanks for tuning in.

[00:05:55]>> Mhm.