Nvidia Pays Poolside $6bn to License Its Model Factory — Then Groq 3 LPX Ships

Last updated: 25 August 2026. Figures below are sourced to primary publications — NVIDIA’s newsroom, Groq’s newsroom, the US Senate, and Poolside’s own blog — linked inline. Where a number is self-reported by a company, or reported second-hand from a document we have not seen, it is flagged as such.

Two things happened to Nvidia inside five days last week. On 20 August 2026, a letter to Poolside’s investors, first reported by Eric Newcomer and later reported by the Wall Street Journal, which said it had reviewed the letter, revealed that Nvidia is paying the coding-model startup $6bn to license its “Model Factory” training system and hire 109 of its engineers. Four days later, on 24 August, the company announced that Groq 3 LPX had entered full production — a dedicated inference accelerator extending its Vera Rubin NVL72 platform, built on technology it licensed from Groq eight months earlier.

Treated separately, these are a product launch and a funding story. Treated together, they are one story at two points in its life cycle: Nvidia has found a way to buy companies without buying them, has now done it three times, and the second of the three has just shipped as a product bearing the licensed company’s name.

The 60-second version

• Nvidia is paying Poolside $6bn for a non-exclusive licence to its model-training system, plus $1bn invested in what remains of the company at a $12bn pre-money valuation, and is hiring 109 staff — the engineers who built Poolside’s Laguna models. The investor letter reportedly states this is “not an acquisition and it is not an acquihire”.
• Four days later, Nvidia confirmed Groq 3 LPX is in full production — the commercial output of the non-exclusive licence Groq itself announced on 24 December 2025, reported at $20bn and cited at that figure by the US Senate, alongside the hiring of Groq founder-CEO Jonathan Ross and president Sunny Madra.
• Across three deals — Enfabrica (September 2025), Groq (December 2025), Poolside (August 2026) — Nvidia has committed roughly $27bn in licence fees. No merger control filing for any of the three has been publicly reported.
• Senators Elizabeth Warren and Richard Blumenthal wrote to Jensen Huang in March 2026 arguing Nvidia had “effectively acquired Groq in all but name”. There is no public evidence of any regulatory action since.
• Nvidia’s press release says Groq Inc. — the nominally independent company — “plans to be among the platform’s earliest adopters” of the Nvidia product built on Groq’s own licensed technology.
• Nvidia reports Q2 FY2027 results on 26 August 2026. How these licences are accounted for is the number to watch.

The paper trail, in dates and figures

The structure, and why it keeps working

The mechanics are consistent across all three transactions. Nvidia pays a very large sum for a non-exclusive licence to a startup’s core technology, and hires the people who built it — in Groq’s case, including the chief executive. The target continues to exist, keeps its name and remaining staff, and distributes the licence fee to its investors. No controlling stake changes hands, and on that framing no merger control filing arises: on paper, no merger has occurred. That is precisely the framing Warren and Blumenthal challenged.

The Poolside investor letter’s reported insistence that the deal is “not an acquisition and it is not an acquihire” is the tell. Ordinary licensing deals do not need that disclaimer. It is there to be read by lawyers, not investors.

What the senators asked, and what the answer looks like now

In their letter of 19 March 2026, Warren and Blumenthal put the question plainly: “While Groq nominally continues to operate as an independent company, by licensing its technology and hiring its most important employees, NVIDIA has effectively acquired Groq in all but name.” They cited Nvidia’s control of roughly 90% of the market for high-end data centre GPUs and asked whether the structure was chosen specifically to avoid antitrust review. They requested answers by 3 April 2026. We could find no published Nvidia response and no evidence of a Federal Trade Commission or Department of Justice proceeding.

Five months on, Nvidia’s own press release supplies something close to an empirical answer. Groq 3 LPX is not a partnership or a joint venture. It is an Nvidia product, shipped as five purpose-built racks, working with Nvidia’s BlueField-4 DPUs, Vera CPU racks, Vera BlueField-4 STX storage and Spectrum-6 SPX Ethernet, sitting inside the Vera Rubin roadmap. And the release states: “Following Nebius, purpose-built AI inference cloud Groq plans to be among the platform’s earliest adopters.”

The independent competitor whose independence justified the deal’s structure now plans to buy the Nvidia product built from its own licensed designs. That is not a competitor; that is a channel partner. CNBC’s report when the Groq deal broke was headlined on an analyst’s view that the structure kept the “fiction of competition alive”. The 24 August release is the strongest evidence yet for that reading — though the reading is ours, inferred from published facts, not a claim Nvidia has made.

Nvidia is no longer just selling the shovels

The Poolside deal is a different kind of move, and the one with longer consequences. Nvidia is not buying chip designs this time. It is buying the ability to train frontier models — Model Factory is described in reports as the internal system Poolside used to build its AI models, including the Laguna coding family — and hiring the 109 engineers who built them. What those engineers will work on has not been stated by either company; Nvidia already ships an open-weight model family of its own, Nemotron, and The Decoder noted the deal puts Nvidia into more direct competition with some of its own customers.

A company holding roughly 90% of the high-end data-centre GPU market is moving further into building models that compete with the models its customers build on its hardware. Our reading of the incentive — not a position Nvidia has stated — is that it is not to win the model market outright but to keep capable open weights freely available, so no single model provider accumulates enough pricing power to squeeze the layer below it. Cheap, good, open models keep inference demand high. It is a commoditise-your-complement strategy, executed with a chequebook.

The timing sharpens the point. Open weights have become a scheduling decision rather than an ideological one — as our reporting on OpenAI pausing its largest training run over cyber risk, with open weights trailing behind set out, the leading labs now stage releases against safety review and competitive timing. Nvidia carries no such constraint: it wants capable open weights out, and has just bought the machinery to produce them.

This is the same structural logic we traced in our analysis of Stripe’s pending acquisition of OpenRouter: infrastructure companies with entrenched positions are buying the layer above them to stop that layer becoming a chokepoint. Nvidia is running the same play with better lawyers and two orders of magnitude more money. For readers weighing what any of this means for their own stack, our comparison of open-source models versus closed APIs sets out the trade-offs that Nvidia is now actively subsidising one side of.

What most coverage is getting wrong

Three things are being missed or muddled.

The two stories are one story. Reported four days apart, they were covered separately — Groq 3 LPX as a chip launch, Poolside as a funding round. The Groq launch is the proof that the licence-and-hire structure delivers integrated product — which is precisely why Nvidia is now confident enough to run it a third time, on a different layer of the stack.

The headline figures are not comparable. Several outlets wrote that Nvidia “acquired Poolside for $6bn”. It did not. This is a non-exclusive licence plus a separate minority investment; Poolside keeps its founders and, in theory, the right to license the same technology elsewhere. Whether it has any commercial reason to do so is another matter.

Poolside was not negotiating from strength. Coverage has largely omitted what happened to Poolside’s 2GW Project Horizon campus, announced in October 2025 with CoreWeave as anchor tenant for a first 250MW phase under a 15-year lease. Data Center Dynamics reported on 2 April 2026, citing the Financial Times, that the partnership “began to fall apart late last year as Poolside was unable to stand up the first cluster of chips to CoreWeave’s timeline”, and that Poolside “was unsuccessful in its fundraising bid” — a roughly $2bn round reportedly including up to $1bn from Nvidia. The $12bn pre-money figure should be read against that, not as a clean step up. Nvidia was the party with options.

Practical takeaways if you build, publish or run software

• Assume open weights get cheaper and better, not scarcer. Nvidia already publishes open weights under Nemotron, and has now paid $6bn for the machinery used to train a competitive open coding model. If your architecture assumes you must pay frontier-API prices for coding or agentic workloads, re-run that assumption every quarter. The floor is being pushed down deliberately.
• Price inference on latency, not just tokens. Groq 3 LPX is a bet that interactive agent workloads are bottlenecked on time-to-completion, not raw throughput. Nvidia reports 3,400 output tokens per second on Gemma 4 31B at 100,000-token context, per Artificial Analysis benchmarking, and “4x faster responsiveness for agents and latency-sensitive workloads than the nearest alternative platform”. The Register’s Tobias Mann identifies that alternative as Cerebras, at 882 tokens/second under the same conditions on Artificial Analysis’ leaderboard, and adds an important caveat: Gemma 4 31B is close to a best case for Nvidia here, because it fits inside a single LPX rack of 64 accelerators, where Cerebras serves the same model on one or two chips. He also notes Cerebras announced its faster CS-4 accelerators the week before. If your agent loop makes many sequential calls, latency shifts your unit economics more than a headline price cut. Worth reading against the broader AI hardware roadmap before committing to a serving architecture.
• Do not treat vendor independence as durable. If your risk register lists a supplier as an alternative to Nvidia, check whether that supplier has since become an Nvidia licensee, customer or employer of record. Groq shows the label can outlive the substance.
• Watch exclusivity language in your own contracts. “Non-exclusive” is doing enormous work in these deals. If you license technology to a much larger partner, the word protects you very little when that partner also hires the team that maintains it.
• For publishers: this is a governance story, not a chip story. The reporting angle with legs is the gap between what merger law can see and what these transactions actually do — a gap we examined in our piece on whether AI governance can keep pace with the industry.

What we still don’t know

• Neither company has confirmed the Poolside deal on the record. Every figure in the Poolside half of this story traces to an investor letter that Poolside has not published and that we have not seen. The Wall Street Journal reported it had reviewed the letter, but no primary document is public. The $6bn, $1bn, $12bn and 109 figures are reported, not verified against a primary document.
• Whether the $1bn investment is new money or the revival of the collapsed round. Nvidia was reported to be investing up to $1bn in Poolside as part of a roughly $2bn raise that, per Data Center Dynamics in April 2026, was unsuccessful. The figure is identical to the one now reported. Nobody has clarified the relationship.
• How Nvidia accounts for a $20bn non-exclusive licence. Capitalised intangible, or expensed? Over what life? This materially affects reported margins, and Q2 FY2027 results on 26 August 2026 are the next opportunity to find out.
• Whether Nvidia ever answered the Senate. The 3 April 2026 deadline passed without any published response we could locate. Correspondence may exist and simply not be public.
• What Groq Inc. competes on now. Groq said in December 2025 that it would continue as an independent company, with Simon Edwards stepping in as chief executive and GroqCloud “operating without interruption”. Eight months on, its founder and its president are at Nvidia and it plans to be among the earliest adopters of the Nvidia product built on its own licensed designs. What that leaves as an independent competitive position is unclear.
• Whether the “4x faster” comparison holds up as configurations change. The figure does trace to Artificial Analysis — Cerebras at 882 tokens/second against LPX’s 3,400 — but it is a single model at a single context length, on hardware where Nvidia needs 64 accelerators to Cerebras’s one or two, and Cerebras has since announced faster silicon. Broader independent testing across models and context lengths has not yet appeared.
• Whether any regulator is looking. Investigations are frequently non-public until they are not.

FAQ

Did Nvidia buy Poolside?

No. On the reported terms, Nvidia bought a non-exclusive licence to Poolside’s model-training system for $6bn, invested a further $1bn for a minority stake at a $12bn pre-money valuation, and hired 109 staff. Poolside continues to exist with its three founders in place.

Why does the structure matter?

Because merger control is triggered by acquisitions of control, not by licences or hiring. A structure that transfers the technology and the people without transferring the company can achieve much of what an acquisition achieves while remaining outside the review process. That is the concern Warren and Blumenthal put to Nvidia in March 2026.

Is Groq still an Nvidia competitor?

Formally, yes — Groq Inc. remains a separate company, with Simon Edwards as chief executive since December 2025. Practically, Nvidia’s own 24 August press release says Groq “plans to be among the platform’s earliest adopters” of Groq 3 LPX, the Nvidia product built on Groq’s licensed technology. Draw your own conclusion.

What should I watch next?

Nvidia’s Q2 FY2027 results on 26 August 2026, for how the licence payments appear in the accounts; the first independent benchmarks of Groq 3 LPX against Nvidia’s claims; and any Nemotron release whose capabilities visibly reflect Poolside’s Laguna lineage.

Sources

Primary

• NVIDIA Groq 3 LPX Now in Full Production With World-Class Speed for Agentic AI — NVIDIA Newsroom, 24 August 2026
• Letter from Senators Warren and Blumenthal to Jensen Huang on the Groq deal (PDF) — US Senate, 19 March 2026
• Warren, Blumenthal Question Whether NVIDIA’s $20 Billion Groq Deal Is Attempt to Avoid Antitrust Laws — US Senate press release, 23 March 2026
• Groq and NVIDIA Enter Non-Exclusive Inference Technology Licensing Agreement — Groq Newsroom, 24 December 2025
• NVIDIA Sets Conference Call for Second-Quarter Financial Results — NVIDIA Newsroom (call 26 August 2026; quarter ended 26 July 2026)
• Announcing Project Horizon: Why we’re building a 2 gigawatt AI campus in Texas — Poolside blog, 15 October 2025

Reporting

• Poolside Strikes $6 Billion Licensing Deal with Nvidia — Newcomer, 20 August 2026 (originating report; partly paywalled)
• Nvidia pays Poolside $6bn to license its model factory and hire 109 staff — TNW, 21 August 2026
• Nvidia is acquiring Poolside’s “Model Factory” and 109 employees for $6 billion — The Decoder, 21 August 2026
• Nvidia is paying $6 billion to license AI model software from startup Poolside — Quartz, 24 August 2026
• What Nvidia’s first Groq 3 LPU benchmarks tell us about its $20B gamble — The Register, 24 August 2026
• Poolside seeks partners for data center in Texas after CoreWeave deal falls apart — Data Center Dynamics, 2 April 2026
• Poolside AI launches open-weight Laguna coding models — Open Source For You, 8 July 2026
• Nvidia-Groq deal is structured to keep ‘fiction of competition alive,’ analyst says — CNBC, 26 December 2025

Please follow and like us: