Here is a puzzle that has kept antitrust lawyers busy for two years. A company pays billions of dollars. It takes a license to a rival’s core technology. It hires that rival’s founder, its president and much of its senior engineering talent. The rival keeps its name, keeps a corporate shell, and keeps operating, at least on paper. Has an acquisition occurred?
Under the letter of US merger law, the answer has been no, and that answer has been worth more than $20 billion to the largest technology companies in the world. The Justice Department is now investigating whether Nvidia structured its deal with AI chip startup Groq specifically to land on the comfortable side of that line.
What makes this one different from the inquiries that came before is that Nvidia is not a company that needed the loophole. It is the most valuable chipmaker on earth, and Groq was one of the few credible challengers to it in AI inference. That combination is exactly what merger review was built to examine, and it never got the chance.
Quick facts
- The DOJ is investigating whether Nvidia structured the Groq transaction to avoid antitrust review
- The deal was announced in December 2025 as a non-exclusive inference technology licensing agreement
- Reported values vary between roughly $17 billion and $20 billion depending on the outlet
- Groq founder and CEO Jonathan Ross and president Sunny Madra moved to Nvidia
- Groq continues as an independent company under new CEO Simon Edwards, retaining its GroqCloud business
- The DOJ opened the inquiry shortly after the December announcement and has sent a formal request for information
- Because a non-exclusive license is not treated as an asset under HSR rules, no premerger filing was required
- Fines are a realistic outcome; forcing the deal to unwind is considered unlikely
Why no filing was required
The Hart-Scott-Rodino Act is the machinery behind US merger review. Above a certain deal size, companies must notify the agencies before closing, wait out a review period, and accept that the government can sue to block the transaction. It is a genuinely powerful tool, and it has one large structural assumption baked into it: that buying a company means buying its stock or its assets.
A non-exclusive license is neither. The FTC’s Premerger Notification Office has held for years that a non-exclusive license does not count as an asset for HSR purposes, and that entering into one is not reportable no matter what it costs. Hiring people is not reportable either. Employment is not an acquisition, and antitrust law has generally treated the labor market and the product market as separate conversations.
Put those two rules side by side and a structure appears. License the technology non-exclusively instead of buying it. Hire the team instead of buying the company. Leave the corporate shell standing. Nothing reportable has happened, even if the practical result closely resembles a merger.
What Nvidia actually got
The specifics matter, because they are what turns a legal theory into an investigation.
Nvidia took a non-exclusive license to Groq’s inference technology, the architecture behind the language processing units that made Groq notable for running AI models at unusually high speed. Jonathan Ross, who founded Groq after working on Google’s tensor processing unit, went to Nvidia. So did president Sunny Madra. Reporting on the deal indicates it covered Groq’s assets broadly, with the exception of the cloud business, and GroqCloud has continued operating. Simon Edwards took over as CEO of what remained.
Reported deal values have ranged from about $17 billion to about $20 billion, a spread that itself says something about how little of this was disclosed through the usual channels. A conventional acquisition of that size would have produced a precise, publicly filed number as a matter of course.
One analyst told CNBC at the time that the arrangement was structured to keep the “fiction of competition alive.” That is a pointed way to put it, but it captures the regulator’s concern cleanly. On paper, Groq still exists and Nvidia still faces a competitor in inference chips. In practice, the competitor’s founder now works at Nvidia and Nvidia holds a license to the technology that made the competitor worth worrying about.
Groq’s language processing units were among the few architectures positioned as a serious alternative to Nvidia in AI inference. Photo via Pexels.
Nvidia is not the first, it is just the biggest
This structure did not appear in December 2025. It has been the dominant way Big Tech has absorbed AI startups since 2024, and by one count the four largest US technology companies spent more than $20 billion between March 2024 and January 2026 hiring away founding teams without formally acquiring a single company.
| Acquirer | Target | Regulatory response |
|---|---|---|
| Microsoft | Inflection | FTC investigation opened mid-2024 |
| Amazon | Adept | FTC probe followed |
| Character.AI | DOJ opened a formal investigation in 2025 | |
| Windsurf | Part of the same pattern of talent and licensing deals | |
| Meta | Scale AI | Cited by lawmakers as a de facto merger |
| Nvidia | Groq | DOJ inquiry opened December 2025, formal information request sent |
Senators Elizabeth Warren, Ron Wyden and Richard Blumenthal have described these arrangements as de facto mergers. The agencies themselves have started to move: on March 25, 2026, the FTC and DOJ jointly requested public comment on whether premerger notification rules should be rewritten to capture acquihires, reverse acquihires and non-exclusive intellectual property licensing deals. That request is an admission that the current rules do not reach this conduct, which is precisely why the Nvidia inquiry is interesting.
The distinction that decides this case
Using a lawful structure is not itself illegal, and companies are entitled to organize transactions efficiently. The DOJ’s reported focus is narrower and harder to defend against: whether the deal was deliberately designed to evade a review it would otherwise have faced. Intent is the difference between clever structuring and avoidance, and intent tends to live in internal emails.
The contrast sitting right next door
There is an unusually clean control case here, and it belongs to Nvidia itself. Earlier this month the company agreed to buy Hugging Face, the platform some 18 million AI developers depend on, in a deal valued around $13 billion. That one is a conventional acquisition, and it will go through the antitrust review that the Groq structure never triggered, despite being worth less money.
Two deals, same acquirer, same year, same sector. The smaller one gets scrutinized because of how it was papered. The larger one did not. Whatever the DOJ concludes about Nvidia’s intent, that comparison is the clearest possible illustration of why the agencies are asking whether the rules still measure what they were meant to measure.
What the DOJ can realistically do
Expectations should be calibrated. Reporting suggests that if investigators find wrongdoing they could seek fines, while forcing Nvidia to unwind the transaction is considered unlikely. Unwinding is an extraordinary remedy in the best of circumstances, and it is close to meaningless here, since you cannot order engineers to go back to a company they left a year ago or make a licensee forget technology it has already absorbed.
The realistic outcomes are narrower and still consequential. A penalty for failure to file would set a price on the structure, which changes the arithmetic for the next deal. A consent agreement could impose behavioral conditions on how Nvidia uses the licensed technology. And an enforcement action of any kind would give the FTC and DOJ a concrete case to point at while rewriting the notification rules, which is likely the most durable effect of all.
There is also the market context. Nvidia has spent the year making commitments at a scale that invites attention on its own, including roughly $500 billion in AI infrastructure financing that critics argue amounts to the company funding its own demand. A regulator already curious about how much of the AI stack runs through one firm will not view the Groq structure in isolation.
Why this matters beyond one deal
Merger review works on a simple premise: before a dominant firm absorbs a challenger, somebody neutral gets to look. The reverse acquihire does not break that rule so much as route around it, and it has now been used often enough, by enough companies, that the exception is starting to look like the norm.
Regulators tend to get there eventually, though rarely quickly, and rarely without a change in the underlying rules. Europe demonstrated this year how far an enforcement regime will go once it does decide to act, to the point where Google degraded its own European search results to comply with the Digital Markets Act. American enforcers are earlier in that arc. The Nvidia inquiry is worth watching less for what happens to this particular transaction than for whether it becomes the case that finally forces the definition of an acquisition to catch up with how acquisitions are actually done.
