NU · neighbordoorsrecords over spin
Open in NU's Reading Room →

Groq, the Company: The Man Who Set Out to Beat Nvidia, and the $20 Billion Deal That Ended With Him Joining It

Imagine spending nine years building a chip company on one bet — that you could out-run the most valuable company on Earth at the thing it was supposed to own. You hire a team, raise billions, sign up two million developers. And then, on Christmas Eve, the giant you set out to beat writes a $20 billion check, takes your patents, and takes you with it. That is, more or less, the actual record of Groq and its founder Jonathan Ross. Here is the business — the real numbers, the real twist, and the part the press release leaves out.

What it felt like to build

The pitch had a clean emotional core: inference should be fast and cheap, and Nvidia's GPUs — built for training huge models — were never the right tool for running them at scale. Ross liked to frame Groq as building "the American infrastructure" for the AI era, delivered "with high speed and low cost" [Groq, Sept. 17, 2025 funding release, attributed to Ross].

That is the feeling that pulled in capital and developers: not a faster GPU, but a different shape of chip for the job everyone was about to need most. Whether that bet paid off depends on how you score the ending — which we'll get to.

The founder: a TPU lineage

Groq was founded in 2016. The credential that mattered: Jonathan Ross was one of the engineers behind Google's Tensor Processing Unit (TPU), Google's in-house AI accelerator [Wikipedia, "Groq"; widely reported]. He co-founded Groq with Douglas Wightman, a former Google engineer [Wikipedia]. So the origin story is literally an ex-Google chip architect leaving to build a rival to the merchant-silicon king. Kooky-sounding, then credentialed.

The chip: LPU

Groq's hardware started as the Tensor Streaming Processor (TSP) and was later rebranded the Language Processing Unit (LPU) [Wikipedia]. The marketing claim is deterministic, low-latency inference — a simpler, more predictable architecture than a GPU. Reported gen-1 specs: a 14nm process part, with a later generation moving to Samsung's 4nm process (announced 2023) [Wikipedia]. Treat exact throughput figures as reported rather than independently benchmarked here — Groq's own funding materials, notably, carried no head-to-head speed-vs-Nvidia numbers [Groq release].

The business model: inference-as-a-service

The commercial engine is GroqCloud, soft-launched February 2024 — a managed platform where developers rent LPU compute through an API instead of buying chips [Wikipedia; Groq]. By the September 2025 raise, Groq said it powered over 2 million developers and Fortune 500 customers, with data centers in North America, Europe, and the Middle East [Groq, attributed]. That is the Nvidia-alternative play in one line: don't sell shovels, rent the dig — and rent it cheaper for the specific job of running models.

The money (the verified record)

Published rounds, by date:

So heading into late 2025, the verified picture was a fast-growing, well-funded inference challenger worth a reported $6.9B. Then the ending changed.

The twist: Nvidia's $20 billion deal

On December 24, 2025, Nvidia agreed to pay approximately $20 billion in connection with Groq — Nvidia's largest deal on record, eclipsing its ~$7B Mellanox purchase [CNBC, Dec. 24, 2025; multiple outlets]. But read the structure, because it is the whole story: this was not a clean acquisition. It was reported as a non-exclusive licensing agreement — Nvidia paying for a perpetual, non-exclusive license to Groq's patent portfolio and software stack, not buying the company outright [reported by Yahoo Finance, CNBC; "non-exclusive inference technology licensing agreement" per Groq's own framing].

The talent moved too: founder Jonathan Ross and president Sunny Madra, plus other senior leaders, are joining Nvidia to scale the licensed technology [CNBC; Yahoo Finance]. Groq, the corporate entity, is said to continue independently — reportedly to be led by finance chief Simon Edwards as CEO — and GroqCloud is said to keep operating [reported].

The part the press release leaves out

Here's where records beat spin. Analysts flagged the deal's odd shape as deliberate. CNBC reported one analyst's read that the license-not-buyout structure is built to "keep the fiction of competition alive" — i.e., to clear antitrust scrutiny by letting Groq nominally survive as an independent competitor while Nvidia absorbs the founder, the team, and the IP [CNBC, Dec. 26, 2025, attributed to analyst]. Some commentators bluntly called it a "hackquisition" — an acqui-hire dressed as a license [Spyglass].

Mark the line clearly between verified and interpretation: Verified — the ~$20B figure, the license structure, Ross and Madra joining Nvidia, the Dec. 24 timing [CNBC; Yahoo Finance]. Interpretation / contested — whether Groq remains a real competitor or is a competitor in name only. That second part is analyst opinion and worth watching, not banking [attributed].

What it adds up to

The honest scorecard: as a technology and talent bet, Groq paid off enormously — a high-school-and-Google-TPU lineage turned into a roughly $20B outcome and a seat inside Nvidia [reported]. As the "Nvidia alternative" it was sold as, the ending is ironic — the challenger's crown jewels now sit inside the incumbent's "AI factory" architecture, per Jensen Huang's stated plan to integrate Groq's low-latency processors [Yahoo Finance, attributed to Huang].

For anyone still building on GroqCloud, the live question isn't the founder's payday — it's whether an "independent" Groq under new leadership keeps investing in the platform now that its inventor works down the road at Nvidia. That one isn't on the record yet.

Sources: Groq newsroom (Sept. 2025 raise); Wikipedia, "Groq"; CNBC — Nvidia buying Groq assets for ~$20B; CNBC — "fiction of competition" analysis; Yahoo Finance — breaking down the $20B deal; DataCenterDynamics — $750M / $6.9B. Image: Groq logo, Wikimedia Commons, by Avivweinstein, CC BY-SA 4.0.

Note on freshness: AI-hardware facts move fast. Figures here are dated and attributed; the Nvidia deal terms reflect reporting around Dec. 24–26, 2025 and the May 2026 reported $650M intra-investor raise. Verify current status before acting on it.

NU original — sourced analysis of the public record. Read it in the interactive Reading Room, or browse more at nothingunseen.com.

Transparency: NU articles are AI-assisted and editor-reviewed, built from the cited primary sources. We label what's proven, alleged, and opinion.