Nvidia's $20B Groq Deal: The Real Risk Isn't the DOJ Probe

The headline is designed to scare you: The Justice Department is investigating whether Nvidia structured its roughly $20 billion deal with AI chip startup Groq to dodge antitrust review. Antitrust against the world's biggest company. Nvidia subpoenaed again. It reads like a storm gathering over the stock.

Read the deal itself, and the storm shrinks. The probe can only end in a fine. The contract will almost certainly stand. The number that actually carries this story is not in the headlines at all: Nvidia paid about three times Groq's private valuation to remove its one genuine rival in the next stage of AI compute.

What the deal actually was

In late December 2025, Nvidia announced its largest purchase ever — larger than the roughly $7 billion it paid for Mellanox in 2019 — at about $20 billion. But it wasn't called a purchase. It was structured as a "non-exclusive licensing agreement" for Groq's inference technology, plus the hiring of Groq's founder Jonathan Ross, its president, and other key people. Groq itself stayed independent, its cloud service kept running, and a new CEO took over.

The form mattered. Groq had raised a $750 million round at a $6.9 billion valuation, with investors including BlackRock, Neuberger Berman, Samsung and Cisco. Nvidia paid roughly three times that for a license and a bag of top talent. Analysts at Hedgeye put it plainly: the deal was "essentially an acquisition of Groq without being labeled one".

Why regulators care about the label

Under the Hart-Scott-Rodino Act, deals above a size threshold must be filed with antitrust agencies before they close. That threshold was about $126 million, and any transaction over roughly $506 million is reportable regardless of the parties' size. A $20 billion cash acquisition would automatically trigger review. Nvidia filed nothing — the license was never submitted.

The law has a gap here, and Nvidia stepped through it. HSR treats the grant of an exclusive IP license as a reportable acquisition; a non-exclusive license generally is not. Here the license was, on paper, non-exclusive. That is the crux of the investigation: did the structure achieve the effect of buying out a nascent competitor while escaping the review a takeover would have received?

The concern is not new, and it crossed party lines. In March, Senators Elizabeth Warren and Richard Blumenthal formally asked whether the deal was designed as a "license" to bypass mandatory merger review. Now the DOJ is probing it as part of a broader, years-long antitrust investigation into Nvidia — the same DOJ that sent the company subpoenas in 2024 over bundling and switching costs.

The downside is a rounding error

Here is where the scare deflates. The probe could close with no action at all — the Bloomberg reporting is explicit on that. The worst realistic outcome is a civil penalty for failing to file, and HSR fines run at most about $53,000 per day of violation. Even if the DOJ won, that is pocket change against a company with a market cap above $5 trillion and some $48.5 billion of free cash flow in a single quarter.

The remedy that would actually hurt — forcing Nvidia to unwind the deal — is the one regulators almost never get on a contract that is already executed, and analysts note the agreement is unlikely to be undone. If the facts showed the deal was a real acquisition that skipped review, that is a missed-filing problem, not a "give back the company" problem. The realistic financial exposure here is measured in the millions on a company that moves tens of billions in a quarter.

The signal underneath

Step back from the legal noise and the economics are striking. The AI boom is shifting from training models to running them — inference, the moment the model actually answers you. Training compute is becoming abundant; low-latency inference is turning into the scarcer, higher-value layer. Groq's chip, the LPU, was the purest bet on exactly that layer, built by a founder who created Google's TPU. It was Nvidia's most credible threat at the precise moment that threat was about to become the market.

That context gives the premium meaning. Nvidia did not pay $20 billion to amuse itself — about 2.9 times the valuation a roomful of sophisticated investors set three months earlier. It paid to take the one rival in the scarcest layer of the next phase off the board. The abundance-scarcity logic points the same way: as GPUs pile up everywhere, control of the scarce inference layer is worth defending at a premium.

The read is not purely bullish, and the honesty matters. Paying three times to buy out a competitor can be read either as an efficient moat purchase or as a sign that Nvidia felt threatened enough to spend. A company secure in its dominance does not usually overpay to neutralize a startup with $500 million in target revenue. But that ambiguity is the point: the DOJ probe tells you little about the business, and the $20 billion tells you something real about where the competitive fight is moving.

The probe will make headlines every few months, and most will age into nothing. The move that matters happened in December, at roughly three times the last agreed value of a company whose one big idea — fast, cheap inference — is the next decade's problem. That is the number to weigh, not the summons.