Antitrust enforcers have opened their first real examination of the deal structure that has replaced the acquisition in AI: Nvidia has received a formal Justice Department demand for information about the roughly $20 billion it paid Groq for a chip license and a hiring spree, an arrangement that never triggered merger review because, on paper, nothing was bought.
Key takeaways
- Nvidia has committed close to $27 billion across three license-and-hire deals in about nine months โ roughly $20 billion with Groq, $7 billion with Poolside, and around $900 million with Enfabrica โ and filed no merger notification for any of them.
- Groq carried a $6.9 billion valuation months before Nvidia paid roughly three times that figure for non-exclusive rights plus its founder and senior bench.
- Hart-Scott-Rodino notification attaches to transfers of equity and assets above $119.5 million, a definition that engineers and non-exclusive licenses sit outside of.
The gap in the statute
Everything about this inquiry turns on a definitional question rather than a factual one. Hart-Scott-Rodino forces a filing and a waiting period on deals above a threshold now set at $119.5 million, low enough that essentially any consequential startup exit clears it โ but the statute reaches share purchases and asset transfers, and a technology license bundled with employment offers is neither.
So the DOJ is not asking whether Nvidia did what it did. It is asking whether the plain text of a 1976 law still describes the thing that happened, when the assets that actually moved were people and intellectual property. Investigators reportedly began looking shortly after the December 2025 announcement, and the formal information demand came to light through The New York Times and subsequently Bloomberg. No findings have been issued, and the file may close without an enforcement action.
What Nvidia bought, and what stayed behind
Groq built Language Processing Units, accelerators tuned for the low-latency inference work where general-purpose GPUs are least efficient โ which made it one of the few companies selling a genuine architectural alternative to Nvidia rather than a cheaper version of the same thing. Nvidia licensed that technology non-exclusively and hired founder Jonathan Ross together with president Sunny Madra and much of the senior engineering group. Groq continued to exist, under a new chief executive named Simon Edwards.
How quickly the technology was absorbed is the detail that makes the license framing awkward. By its March GPU conference, barely three months later, Nvidia was demonstrating a new inference processor built on the Groq designs. Groq meanwhile raised $650 million in June 2026 to become a renter of GPU cloud capacity instead of a maker of chips, an outcome that leaves the original competitive thesis retired.
Nvidia's public position is that none of this needs defending. The company has characterised the arrangement as the American system functioning exactly as intended โ rewarding entrepreneurs, encouraging innovation, benefiting consumers.
Why Hugging Face makes the pattern legible
The strongest evidence that structure is being selected deliberately is Nvidia's own inconsistency about when it files. On September 3 the company announced a conventional $12.9 billion purchase of Hugging Face, the model repository used by millions of developers, and submitted it to standard review on both sides of the Atlantic with the argument that the platform will deconcentrate AI rather than concentrate it.
A company that reaches for a real merger filing when a repository is involved, and a license when a chip competitor is, has demonstrated that it understands the difference the paperwork makes. That asymmetry, more than the Groq terms in isolation, is what an enforcer would put in front of a judge.
Nor is Nvidia the only practitioner. Microsoft ran a version of this with Inflection, Google with Character.AI and Windsurf, and Meta with its roughly $14 billion Scale AI arrangement. Senators Elizabeth Warren and Richard Blumenthal โ who told chief executive Jensen Huang in a March 2026 letter that his company had effectively acquired Groq in all but name and appeared to be evading regulators โ have since joined Senator Ron Wyden in asking the FTC and DOJ to look at the template industry-wide. The Groq file is functioning as the test case for all of it.
What to watch next
Reversal is close to off the table, which is precisely what makes the structure attractive: by the time anyone examines it, the engineers have onboarded and the designs are shipping, leaving fines or conduct conditions as the available remedies. Two signals matter more than any ruling. One is escalation from information gathering to a civil enforcement action. The other is quieter โ acquirers voluntarily filing on license-plus-talent deals, which would reveal that outside counsel has already moved even though the law has not. The pressure Groq was built to apply is now being applied instead by buyers designing their own inference silicon to escape Nvidia's pricing.
FAQ
What is a reverse acquihire?
A large company licenses a startup's core technology and hires its founder and key engineers, while the startup remains legally independent. Because no shares or assets change hands, merger notification typically does not apply even though the competitive effect can resemble a takeover.
Could regulators unwind the Nvidia-Groq deal?
Almost certainly not. The team has already moved and the license is already operating, so financial penalties or behavioural conditions are the realistic outcomes. The significance is precedential โ whether deal structure alone keeps working as a shield from review.
Does Groq still exist as a company?
Yes. It operates independently under chief executive Simon Edwards and closed a $650 million round in June 2026, but its business has shifted from selling its own inference accelerators to providing cloud capacity, after its founder and much of its senior team moved to Nvidia.
