$20bn and the Word “Non-Exclusive” (10 September 2026): The DOJ Is Probing How Nvidia Structured the Groq Deal — and Why Hugging Face Was the Exception

The Justice Department is investigating whether Nvidia structured its roughly $20 billion Groq transaction to avoid antitrust review — first reported by The New York Times on 9 September 2026, carried by Bloomberg and Reuters the next day, and understood to have been opened shortly after the deal was announced on Christmas Eve 2025. Nothing has been alleged or proven. What the probe does is put a price on a piece of deal grammar most readers have never had a reason to notice: under the premerger notification rules, an exclusive licence is an asset and a non-exclusive one is not. That is the hinge. It is why Nvidia's largest AI commitments of the past nine months triggered no merger filing at all — and why Hugging Face, the one it agreed to buy outright on 2 September, comes with a closing condition and a wait until 2027.

What actually happened: a formal demand about a structure, not a price

On Wednesday 9 September 2026, The New York Times reported that the Justice Department is investigating whether Nvidia designed its Groq transaction to sidestep antitrust scrutiny. Bloomberg and Reuters carried it the following day. On that reporting, the department opened the matter shortly after the deal was announced and has sent Nvidia a formal demand for information; it could impose a fine if it concludes the transaction was mishandled, but is not expected to try to unwind it.

Hold the facts of the underlying deal steady, because they are the whole argument. Nvidia announced on 24 December 2025 that it would pay about $20 billion in cash to buy assets from Groq and take a non-exclusive licence to Groq's inference technology. Founder and chief executive Jonathan Ross, president Sunny Madra and other senior leaders moved to Nvidia. Groq itself did not disappear: it continued as an independent company under finance chief Simon Edwards as chief executive, and GroqCloud kept operating. Reuters puts the licence component specifically at $17 billion. It remains the largest transaction Nvidia has ever done.

Notice what is absent from that description. No company was acquired. No voting securities changed hands. The two things that did move — a licence and a group of people — are, on their own, not the kind of object the premerger notification system was built to catch.

The word the whole thing turns on: “non-exclusive”

Merger review in the United States runs through the Hart-Scott-Rodino Act, which requires parties to notify the FTC and the DOJ antitrust division before completing certain acquisitions of voting securities or assets above a size threshold, then wait out a review period. The 2026 size-of-transaction threshold is $133.9 million, effective 17 February 2026; when the Groq deal was signed in December 2025 it stood at $126.4 million. Those numbers are adjusted each year to changes in gross national product.

The operative phrase is voting securities or assets. Intellectual property gets there only by being characterised as an asset, and the FTC's premerger staff have a long-settled line on when it is. The grant of an exclusive patent licence is treated as the transfer of an asset, and therefore potentially reportable. So is a partially exclusive licence — one limited to a single field of use or a single territory. A non-exclusive licence is not regarded as an asset for HSR purposes, because the licensor retains the right to grant the same rights to anyone else. Even a co-exclusive licence, where licensor and licensee both hold rights, is treated as non-exclusive.

That is not a loophole somebody discovered last year. It is decades-old interpretive guidance, largely written with pharmaceutical patent licensing in mind, now load-bearing in a market it was never designed for. And it explains why the same adjective appears at the front of announcement after announcement.

Whether such a combination should be looked at whole rather than piece by piece is precisely the kind of question now under examination — and there is a recent, concrete illustration that regulators do sometimes look at the whole. On 13 July 2026 the FTC secured $12 million in penalties — $10 million from Edwards Lifesciences and $2 million from Genesis MedTech — over a transaction split into a $115 million purchase of JC Medical, just under the then-applicable $119.5 million threshold, alongside a $25 million investment in Genesis. Separately, each sat below the line; together, they did not. FTC Chairman Andrew N. Ferguson's statement was blunt: "Companies that try to sneak deals through without lawful FTC review should take notice. The FTC will be vigilant in enforcing the requirements of the Hart-Scott-Rodino Act and we will not hesitate to seek penalties for its violation."

That case is not this case, and a medical-device split purchase is not a technology licence. It is offered only as evidence of the enforcement posture that makes the Groq question a live one rather than an academic one.

The structure is the story: five commitments, one merger agreement

Nvidia has spent the last nine months deploying balance sheet across the AI stack, and it has consistently chosen instruments that stop short of acquiring a company. Hugging Face is the exception, and the table makes the pattern legible.

The Poolside arrangement Bloomberg reported on 20 August 2026, citing Newcomer, is the closest sibling to Groq: a $6 billion non-exclusive licence covering the system Poolside uses to build its Laguna family of open-weight coding models, a $1 billion investment at a $12 billion pre-money valuation, offers to the engineers who built those models — and Poolside's three founders staying put, with the company continuing independently. Same grammar, smaller number.

The MediaTek deal is the sharpest contrast in the other direction. Nvidia put $3.5 billion into convertible bonds while MediaTek agreed to adopt NVLink Fusion, providing "hyperscalers, cloud service providers and frontier model developers with a prevalidated path to develop custom XPUs and bring them into NVIDIA NVLink-connected, rack-scale AI factories." Debt securities and an architecture commitment; no ownership, no licence, no filing question.

Then Hugging Face, where the filing carries the one line with a real condition: closing is "subject to the satisfaction or waiver of customary closing conditions, including receipt of required regulatory approvals." That is why the expected close is the first half of 2027 rather than this quarter. For readers tracking the same balance sheet through a different instrument, our earlier piece on the $105bn OpenAI residual value guaranty works through the contingent-obligation version, and the Q2 FY2027 results note picks out the $2,944m financing outflow labelled only "Groq, Inc." in the cash flow statement — the deal appearing in the accounts, quarter by quarter.

$11.9bn to holders, $1bn to keep the people

Nvidia entered into the definitive agreement on Wednesday 2 September 2026 and filed the disclosure the next morning under Item 8.01, signed by chief financial officer Colette M. Kress. The operative language: the transaction "includes an approximately $11.9 billion purchase price payable to Hugging Face stockholders, subject to certain adjustments, and an equity-based retention program of up to approximately $1.0 billion for Hugging Face employees joining NVIDIA."

Two details do real work. "Subject to certain adjustments" means $11.9bn is a headline rather than a wire amount. And the retention pool is equity-based and up to — not part of the price paid to owners, but the cost of keeping the people whose judgment is most of what was bought, and it only pays if they stay. In his announcement post, Jensen Huang gave the total to the hundred thousand — $12,930,300,000. Hugging Face chief executive Clément Delangue told CNBC's Squawk Box that the approach came from his side: "During the summer, I think we realized that Hugging Face and open source AI in general was at the turning point, and that it needed more, more resources, more scale, more visibility." CNN reported the company was valued at $4.5 billion in a 2023 round, and had previously declined a $500 million Nvidia investment at a $7 billion valuation.

What $12.93bn buys is not chips, fabs or a model. It is a registry: more than 18 million developers sharing more than 3 million models, 500,000 datasets and 1 million applications, with more than 200,000 companies using the platform "to discover, evaluate, customize and deploy AI." Nvidia is already its largest single contributor, with more than 500 open models and more than 250 open datasets.

The transmission to a semiconductor income statement is indirect, which is why it is worth spelling out. Open weights are hardware-agnostic by construction: a downloaded model is a file of parameters that will run on whatever accelerator the operator points it at. What the repository determines is not which chip but how much inference happens at all — how quickly a team finds a model that fits the job, how cheaply it can evaluate three of them, whether the path from "we should try this" to "it is running in production" is measured in days or quarters.

The new risk factor is the thesis, written by the buyer

Filings rarely hand you the reasoning. This one does, because Nvidia had to add a risk factor, and a risk factor has to explain what is at risk. Its opening sentence is the whole investment case stated in reverse:

"Demand for open-source foundation models and applications based on them promotes the use of our products worldwide and sustains the Hugging Face platform."

That is the mechanism, described by the party with the most exact knowledge of it. The threat that follows is not Broadcom, or Google's TPUs, or a rival hub. It is policy. Nvidia writes that "other parties are actively lobbying the U.S. Government and other stakeholders worldwide to adopt legislative or regulatory measures that would restrict or disadvantage open-source models and the customers of them," and that new requirements "could restrict the models or datasets available through Hugging Face, require changes to Hugging Face's platform or practices, delay or restrict offerings, increase compliance costs or result in investigations or enforcement actions."

Then the paragraph tying the acquisition to the export-control file: "Many of the world's most popular and successful open-source models originated in China and are then downloaded, revised, fine-tuned, and tested by developers in the United States and worldwide." Any regulatory control limiting Nvidia's ability to support models derived from any region, including China, it says, could materially affect both the platform and the business.

That is a company telling shareholders that a meaningful share of what makes its new asset valuable is produced in a jurisdiction its own guidance already excludes on the revenue line — the same 26 August release states Nvidia "is not assuming any Data Center compute revenue from China" in its $108.0 billion third-quarter outlook. The chips cannot go one way; the weights already come the other. Nothing about that observation requires a view on whether either restriction is wise. It only requires noticing that the two flows run in opposite directions through the same company — and that "investigations or enforcement actions", a phrase Nvidia put in its own risk factor on 3 September, acquired a concrete referent six days later.

The channel to what you actually trade

This is not a currency story, and pretending otherwise would be the wrong lesson. Its instruments are the US index complex — the S&P 500 and ES, the Nasdaq 100 and NQ — and the reason is concentration rather than the deal itself. An acquisition costing roughly 13% of one quarter's revenue at the largest index constituent is close to a rounding error in index arithmetic, and a civil fine, if one ever arrives, would be smaller still.

What is not a rounding error is the precedent. If a licence-plus-talent structure turns out to carry filing obligations, the cost is not the penalty; it is the return of a waiting period to transactions that currently close on announcement. Deal certainty and deal speed are what the structure buys, across an industry that has reached for it repeatedly — and that is a sector-wide input, not a single-name one.

For proportion, recall the announcement session itself. Nvidia rose about 1% on Thursday 3 September, while the Nasdaq Composite rose 1.4% to 26,584.06 and the S&P 500 gained 1.06% to 7,747.71. The stock announcing a $12.9 billion acquisition underperformed its own index, because the session belonged to rates: a 4.77% 10-year, falling after Federal Reserve Governor Christopher Waller told Reuters that "if this continues in the data due over the next two weeks, I would be inclined to support holding the target for the federal funds rate at its current setting." The semiconductor tape had its own dispersion that day, which we walked through in the Broadcom Q3 note.

That split is the discipline worth keeping. The FX channel operating on 3 September ran through rates, not chips — a falling yield and a governor leaning toward a hold is a dollar story on the interest-rate factor, one of the five the meter scores, and it would have happened with or without the announcement. The habit of making every large headline a dollar headline is how mechanisms get mistaken for narratives. More on how we frame that on the about page.

What would change the picture

Four things, none of them this week's share price. First, the outcome of the demand: whether the department closes the matter, seeks a penalty, or — the outcome reporting currently treats as unlikely — attempts anything structural. Second, whether the theory travels. The licence-plus-hiring pattern has been used widely across the industry; a conclusion that it is reportable would reprice deal timelines far beyond one buyer, which is why this is a sector question wearing a single company's name.

Third, the Hugging Face review itself: the filing's H1 2027 close and its "required regulatory approvals" condition mean there is a window in which conditions or commitments could be attached — and the open-platform and "support other silicon vendors" language reads like an opening position for exactly that conversation. Fourth, the policy file Nvidia named itself: any concrete measure restricting the distribution of open weights, particularly China-origin ones, would reach the asset's value directly, and Nvidia has said so in writing.

Until then, nothing changes operationally. Hugging Face runs as it ran, Groq runs as it runs, and Nvidia's $56.6 billion of cash, equivalents and marketable securities at the end of July absorbs any of these outcomes without strain. What changed on 9 September is narrower and more interesting than a headline risk: the deal grammar that let an industry move $20 billion without a filing is now being read closely by the people who wrote the grammar.

Educational macro context only — not investment advice.