Justice Department asks whether a licence and a hiring spree did the work of an acquisition

If Groq was the second name on your inference vendor list last year, the useful diligence question now is whether the roadmap still has authors. The company lost its founder and senior engineers after the Nvidia arrangement, and it has since been raising $650mn to rebuild around the team and business that remain ^[12].

The mechanism the investigation runs into is the gap between control and capability. US antitrust review generally turns on whether control of a company changed hands, and a non-exclusive licence does not transfer control, so enforcers would have to argue the arrangement functioned as a merger without being one on paper ^[7]. Nvidia's account is internally consistent on those terms, in that the licence is not exclusive, Groq still exists, and its engineers are free to join someone else ^[8]. Both that description and the $650mn rebuild can be true at once, and only one of them would ever have appeared in a filing.

The price tag is unsettled, and that makes it a weak signal for now. Reuters, following the New York Times, puts the Groq arrangement at $17bn, while The Next Web has reported $20bn, and the difference may turn on whether the figure covers the licence payment alone or includes the value of the hiring ^[10]. That $3bn spread between two published numbers is about 86 percent of the $3.5bn valuation Groq raised at in a round Nvidia itself joined ^[18]. When the disagreement between reporters is nearly the size of the company, the deal value is too shaky a number to build an argument on.

The Poolside deal is the cleaner arithmetic. Nvidia paid $6bn to license Poolside's model factory and hired 109 staff ^[9], which works out to roughly $55mn a head if you allocate the entire payment across the hires ^[19]. The reported consideration is for the licence rather than the people, so treat that as a ceiling rather than a salary. What it measures is how much money can move in this shape without a merger filing.

Two deals by one dominant supplier is a narrower pattern than the label "the new acqui-hire" suggests, and the material here does not show that every AI talent deal of the past two years was built this way. The Next Web's report argues that repetition is itself the problem, because once the same structure appears more than once it becomes harder to treat each instance as an isolated commercial decision ^[16]. Congress got there before the DOJ did, with Senators Elizabeth Warren and Richard Blumenthal questioning the Groq deal in March ^[13].

For the person who has to defend a vendor choice on Friday, the forcing function is short. The test is whether the roadmap survives the departure of the five people whose absence would stall your challenger vendor's next chip or model, including the scenario where the incumbent takes a non-exclusive licence to that technology and hires those five. If nothing changes, that is a real second source. If the vendor is reduced to one supplier plus a support contract, what was bought was a price quote, and a licence followed by a round of hiring looks on paper like two normal business decisions that nobody has to notify ^[5].