Nvidia is being investigated over a deal it never had to report. The Justice Department has sent the company a formal demand for information about the roughly $20 billion arrangement it struck with AI inference chip startup Groq in December, and the question on the table is not whether the deal was too big. It is whether the deal was built the way it was built precisely so nobody in Washington would get to look at it.
Most of the coverage this week has treated that as an Nvidia problem, complete with a dutiful note that the stock slipped about 1% on Thursday. That framing is too small. The legal theory the DOJ is testing here does not stop at Nvidia, and it was never really about Nvidia. At least four other transactions that closed over the past two years used the same structure, none of them were filed for merger review either, and every one of them sits inside the reach of whatever conclusion the department comes to. This is the enforcement test case for the entire playbook.
What Nvidia Actually Signed
The mechanics matter, because the mechanics are the story. Nvidia did not buy Groq. It took a non-exclusive license to Groq’s inference chip technology, and it hired the people who built it: founder and chief executive Jonathan Ross, who is widely credited as the architect of Google’s original TPU, along with president Sunny Madra and a large share of the senior engineering bench. Groq itself continued to exist. Simon Edwards took over as chief executive, and the company remained independent on paper, still free to license its technology to anyone else.
Read those as two separate events and they are unremarkable. Companies license technology constantly. Companies hire executives constantly. Neither action, on its own, triggers a filing under the Hart-Scott-Rodino Act, which requires advance notice to regulators when one company acquires stock, assets, or a controlling interest in another. Nvidia acquired none of those things. It acquired the team, the tech and the strategic position, and left the corporate shell standing.
That is the whole trick, and we said so in December when the deal was announced: it was a deal that was not technically a deal. The price tag was roughly three times Groq’s most recent private valuation, which is not what a non-exclusive license usually costs.
A merger has to be reported and can be blocked. A licence followed by a hiring round looks, on paper, like two ordinary business decisions.
The Four Deals Standing Behind This One
Here is what almost nobody covering the probe has said out loud. If the DOJ concludes that a license-plus-hire package amounts to an acquisition of control that should have been reported, the finding lands on a whole cohort of deals that were structured identically and never filed:
- Microsoft and Inflection. Roughly $650 million to license Inflection’s models, plus the hiring of co-founders Mustafa Suleyman and Karén Simonyan and much of the staff. Britain’s Competition and Markets Authority opened its own inquiry into that one.
- Google and Character.AI. A licensing arrangement valued in the $2.5 billion to $2.7 billion range, with co-founders Noam Shazeer and Daniel De Freitas returning to Google. The Justice Department already opened a formal look at this one in 2025.
- Google and Windsurf. Same shape, same year.
- Meta and Scale AI. Same shape again.
Antitrust scholars have been circling this for a while. A University of Chicago Law Review analysis asked directly whether a mass resignation should be treated as a merger under Section 7, and senators including Elizabeth Warren, Ron Wyden and Richard Blumenthal have called these arrangements de facto mergers in everything but paperwork. The Nvidia inquiry is the first time a US enforcer has put real process behind the theory, sending a formal information demand rather than making a speech.
Where We Land
The structure was designed to avoid merger review. That is not a cynical read, it is the functional description of what a non-exclusive license plus a leadership hire accomplishes that a purchase does not, and the pricing gives the game away. Paying triple a company’s valuation for rights you are explicitly not getting exclusively is not a licensing decision. It is an acquisition with the reportable parts filed off.
Nvidia’s lawyers will say, correctly, that they followed the statute as written. That is the problem worth naming. HSR polices the transfer of stock and assets, and the most valuable asset in an AI company is roughly forty people who can walk across the street. The law is aimed at a corporate form that the industry has simply stopped using when it wants to avoid scrutiny.
Which is why a fine, the outcome reporting suggests is most plausible if the DOJ finds anything at all, would settle nothing. Officials are not expected to try to unwind the deal, and a penalty absorbed by a company with Nvidia’s balance sheet is a cost of doing business, not a deterrent. We have watched this movie before: KKR paid a record HSR penalty this year and the bill landed on its law firms rather than changing anyone’s behavior. If the remedy for skipping merger review is a manageable fee assessed years later, skipping merger review remains the rational choice.
The fix is legislative, not prosecutorial. Congress should amend the reporting thresholds so that a transaction transferring a target’s founders, core engineering staff and technology rights in a coordinated package is reportable regardless of whether any equity changes hands. Antitrust enforcers should not have to litigate their way to a definition of “control” that any working technologist could supply in a sentence.
What to Watch
The probe may end quietly. Investigations of this kind frequently do, and the department has signaled no appetite for forcing a reversal. But the information demand itself is the meaningful event, because it establishes that a structure the industry treated as settled is not settled. Every general counsel who signed off on a reverse acquihire since 2024 is now reading the same news.
For Nvidia, the near-term financial risk is negligible against a company of its size. The longer-term risk is that the cheapest route to acquiring a competitor’s capability, buy the people and rent the patents, stops being quietly available. In a market where the binding constraint on every AI company is talent rather than capital, closing that route would change more about how the sector consolidates than any single merger challenge ever has.