How Does Groq Make Money: LPU Chips, GroqCloud Tokens, and the $20B Nvidia Licensing Deal

Nvidia paid roughly $20 billion in December 2025 to license Groq's inference chip technology and hire 80% of its staff, including founder Jonathan Ross. That's the short answer. The longer answer is more interesting.

Groq spent nine years building a real, growing business around a genuinely different chip — one that reached a $6.9 billion valuation just three months before its founder walked out the door to join the market leader it was built to challenge. Understanding how Groq made money before December 2025, and what's left of that business now, tells you more about how the AI inference market actually works than most head-to-head chip benchmarks ever will.

How Does Groq Make Money?

Before December 2025, Groq made money selling access to its Language Processing Unit (LPU), a chip architecture built specifically for AI inference rather than training. Revenue came from two lines: enterprise and sovereign-cloud customers buying GroqRack hardware — on-premise clusters of 64 to 576-plus LPUs for data-residency-sensitive workloads like Aramco Digital's multi-rack deployment — and developers paying per-token fees through the GroqCloud API, priced as low as $0.05 per million input tokens for models like Llama 3.1 8B, running 10-20x cheaper than equivalent GPU-based inference.

That business was real enough to more than double Groq's valuation from $2.8 billion in August 2024 to $6.9 billion in September 2025 on a $750 million raise led by Disruptive, with BlackRock, Neuberger Berman, and Deutsche Telekom Capital Partners joining existing backers Samsung, Cisco, D1, and Altimeter. If you're tracking how the rest of the private AI market is being priced against revenue, see our AI valuations dashboard.

Then, on December 24, 2025, Nvidia paid roughly $20 billion for a non-exclusive, perpetual license to Groq's LPU designs and software stack, and hired away 80% of the company's employees — including CEO Jonathan Ross and president Sunny Madra. Groq kept its cash, its board, and about a fifth of its team, and has been rebuilding around a much narrower business ever since.

Jonathan Ross Built the LPU to Beat the Chip He Co-Invented

Groq's origin story is unusually direct: Jonathan Ross helped design Google's first Tensor Processing Unit (TPU) before leaving in 2016 to co-found Groq with fellow Google X engineer Douglas Wightman. Where GPUs and TPUs are built for the parallel, throughput-heavy demands of training, Ross designed the LPU as a "software-defined," deterministic chip optimized purely for the low-latency, sequential math of running a trained model — inference, not training.

That focus paid off in raw speed: Groq's paid API has run open models like Llama 3 8B at over 1,300 tokens per second, a figure GPU-based competitors have struggled to match on equivalent hardware. It's also why Nvidia, not a scrappy startup rival, ended up being the buyer — the LPU solved a real gap in Nvidia's own GPU-first product line as inference workloads started outgrowing training workloads industry-wide.

The $20B Nvidia Deal: Licensing, Not an Acquisition

The Groq-Nvidia transaction was structured as a technology license and mass hiring event, not a formal merger — the same "hire and license 2.0" playbook Google used with Character.AI and Windsurf, and Microsoft used with Inflection AI. Nvidia paid for a non-exclusive, perpetual license to Groq's LPU intellectual property and brought over roughly 80% of Groq's staff, while Groq itself remained a legally separate company. We covered the mechanics of that broader acquisition pattern in our breakdown of Hire & License 2.0.

Of the roughly $20 billion headline figure, $17 billion was structured as cash paid across three installments through the end of 2026. Groq distributed the first tranche — $7.6 billion — to shareholders in February 2026, delivering a return to backers like Samsung, Cisco, D1, Altimeter, BlackRock, and Neuberger Berman years before any realistic IPO timeline would have paid them out.

What's Left of Groq's Business Model in 2026

With its founder, president, and most engineers now at Nvidia, the Groq that remains is a much smaller company betting entirely on the cloud side of its old business. On June 22, 2026, Groq raised $650 million led by Disruptive and Infinitum — both repeat investors — specifically to expand GroqCloud's inference-cloud data centers, rebuild its C-suite, and operate as a pure-play inference-as-a-service provider rather than a chip designer competing head-on with Nvidia.

Notably, Groq did not disclose a valuation for the June 2026 round — a silence that stands out against the company's habit of publicizing every prior raise. Most analysts read that as reflecting a business worth meaningfully less than the $6.9 billion it commanded in September 2025, now that its most valuable asset (the LPU exclusivity) has been non-exclusively licensed away and its highest-profile talent sits inside a competitor. Track how the rest of the AI infrastructure buildout compares on our Big Tech Earnings Tracker.

What Groq kept, beyond the cash, is the GroqCloud brand, its existing developer base, and a rebuilt executive bench brought in specifically to run a cloud business rather than a chip-design shop. Whether $650 million is enough runway to compete against Nvidia's own inference offerings, AWS Inferentia, and Google's TPU-backed Cloud AI depends almost entirely on whether GroqCloud's pricing and latency edge survives now that its underlying architecture is licensed to the very competitor it was built to beat.

Groq vs Cerebras vs SambaNova: Three Different Inference-Chip Outcomes

Groq, Cerebras, and SambaNova all set out to challenge Nvidia's dominance of AI inference with fundamentally different chip architectures — and by mid-2026 each has landed in a completely different place. Cerebras took the public-markets route, closing its first trading day at roughly a $66 billion fully diluted valuation after a $5.5 billion IPO. SambaNova took the opposite path, cutting staff in 2025, watching an Intel acquisition collapse, and raising a $350 million down round at a $2.2 billion valuation. Groq landed somewhere in between: a real exit event, just structured to look like anything but one.

Figures blended from Bloomberg, TechCrunch, DataCenterDynamics, and Hashrate Index reporting, 2025-2026. Private-company valuations are based on the most recent disclosed funding rounds and may not reflect current market value.

The Bull and Bear Case for What's Left of Groq

The bull case is that Groq's remaining team gets to run a leaner, cloud-only business with $650 million in fresh capital and zero chip-manufacturing overhead, at the exact moment inference demand — not training demand — is becoming the larger AI infrastructure spending category. A non-exclusive license means Groq can still sell GroqCloud access built on its own LPU designs; it just no longer has to defend that architecture against Nvidia's balance sheet.

The bear case is that Groq gave up the two things that made it valuable in the first place: proprietary technology (now non-exclusively licensed to the market leader) and the founder whose reputation drove the September 2025 valuation jump. An undisclosed valuation on the June 2026 round, against a public $6.9 billion comp from nine months earlier, is not the kind of silence a company stays quiet about when the number is bigger.