Acquisition by Contract: DOJ Probes Whether Nvidia's $20B Groq Deal Was Built to Dodge Antitrust Review

The U.S. Justice Department is investigating whether Nvidia deliberately structured its deal with AI chip startup Groq to avoid antitrust scrutiny, the New York Times reported on September 9, 2026, citing two people familiar with the matter. Reuters and Bloomberg confirmed the probe, with Bloomberg noting the inquiry centers on whether Nvidia’s roughly $20 billion licensing agreement with Groq was engineered to stay below the radar of formal merger review.

The story matters far beyond one chip deal. It is the sharpest test yet of a deal structure that has quietly become the AI industry’s favorite escape hatch: pay billions for a “non-exclusive license,” hire away the founder and the senior engineering team, and never file a merger notification at all. Whether that maneuver survives contact with antitrust law will shape how every AI giant buys technology for the rest of the decade.

What the deal actually was

Announced in late December 2025, the Nvidia-Groq transaction was framed from the start as a licensing arrangement, not an acquisition. Nvidia paid a reported $17-20 billion for what both companies described as a “non-exclusive license” to Groq’s chip technology, and hired several of Groq’s executives — including founder Jonathan Ross, the ex-Google TPU architect who built the company around the Language Processing Unit (LPU) concept.

The technical logic was straightforward. The AI hardware market is undergoing what analysts call the “Inference Flip”: as the industry’s center of gravity shifts from training enormous models to serving them to billions of users, the workload demands low latency and deterministic throughput — exactly the properties Groq’s LPU architecture was designed around. Independent tests during Groq’s independent era showed its hardware generating tokens at several times the speed of GPU-based rivals on some models. For Nvidia, which holds an estimated 70-95% of the data center accelerator market, Groq was less a startup than a potential challenger to its core inference franchise.

That is precisely what caught regulators’ attention. The DOJ opened its investigation shortly after the deal was announced in December, according to the NYT report, and has since sent Nvidia a formal request for information — a step that signals the probe has moved beyond a preliminary inquiry.

The “acquisition by contract” problem

At the heart of the investigation is a question that sounds technical but is really about power: when does a license plus a talent transfer become a merger in everything but name?

Competition scholars have a name for the pattern — “licensing-based concentration,” “asset-light acquisition,” or “reverse killer acquisition.” The concern is not the license itself but the cumulative effect: the acquirer ends up controlling the rival’s technology, its revenue pipeline, and the people who could rebuild it elsewhere, while the rival continues to exist as a shell with no realistic path to competing. The formal structure says “commercial agreement.” The competitive effect may say “elimination of a nascent competitor.”

Nvidia’s playbook is not unique. Microsoft’s 2024 deal with Inflection AI — hire the founders and staff, license the IP — was examined by four different competition authorities, including the UK CMA (which cleared it), Germany’s Bundeskartellamt (which found it theoretically constituted a concentration but declined jurisdiction), and the European Commission (which also declined jurisdiction). The FTC’s investigation of that transaction reportedly remains open. Nvidia itself pulled a similar move in September 2025, licensing the technology of chip interconnect startup and hiring its team, as Senator Elizabeth Warren and Senator Richard Blumenthal pointed out in a March 2026 letter questioning whether the Groq deal was “an attempt to avoid antitrust laws.”

What makes Nvidia-Groq different is scale. A $20 billion transaction is not a quiet acquihire; it is one of the largest technology deals of the past decade, executed entirely through a structure that never triggered a Hart-Scott-Rodino merger filing. Legal commentators have noted that the deal sits at the “outer contours” of the rules governing when transactions that avoid HSR review can themselves be considered avoidance. If the DOJ concludes the structure was a deliberate end-run, the case could establish that regulator-shaping a deal’s form is itself actionable.

Nvidia’s defense: “the system working as designed”

Nvidia is not disputing that the probe exists, but it is pushing back hard on the framing. “The Groq story is a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers,” a company spokesperson said in a statement.

There is a real argument behind the slogan. Groq investors received a massive return; the LPU technology gets manufactured and distributed at a scale Groq could never have reached alone; and the license is formally non-exclusive, leaving Groq free — on paper — to license its technology to others. Nvidia will argue that licensing deals are ordinary commerce, that preventing large companies from licensing promising technology would chill exactly the kind of startup exit that venture capital depends on, and that the AI inference market is more competitive than the “70-95%” headline suggests, with AMD, custom silicon from the hyperscalers, and a wave of well-funded startups all pressing in.

The counterargument is that “non-exclusive” means little once the founder and the engineering brain trust have moved across the street. A license without the people who can extend it is a patent portfolio; a startup without its architects is a licensing vehicle. The DOJ’s question is whether Nvidia understood that as well as anyone — and structured the deal accordingly.

What happens next

According to the NYT report, the DOJ’s most likely remedies are financial rather than structural: the agency could fine Nvidia if it finds the deal mishandled, but is unlikely to seek to unwind the transaction. That would be consistent with precedent — unwinding a completed licensing deal that has already reshaped a company’s product roadmap (Nvidia’s Groq-derived LPX inference chips are now in full production) would be extraordinarily disruptive.

But the deterrent value may matter more than the dollar amount. A formal finding that Nvidia structured the deal to evade review would put every AI company on notice that license-plus-acquihire is not a regulatory free zone. The European Commission is already evaluating the transaction along similar lines, and a U.S. finding would give other jurisdictions — and Congress — a template for closing the loophole, whether through enforcement guidance or by amending HSR thresholds to capture de facto control transfers.

For the AI industry, the stakes are straightforward. The great consolidation of frontier technology is happening through contracts, not mergers — and for the first time, the enforcer responsible for policing it is saying out loud that it noticed.

Sources

• Justice Dept. Investigates Nvidia Deal With Groq — The New York Times
• DOJ probes Nvidia’s licensing deal with AI startup Groq, NYT reports — Reuters
• DOJ Probes Nvidia’s License Deal With Groq on Antitrust Concerns — Bloomberg
• Warren, Blumenthal Question Whether NVIDIA’s $20 Billion Groq Deal Is Attempt to Avoid Antitrust Laws — U.S. Senate
• NVIDIA-Groq License Agreement: A Legal Review — Göksu Safiişik Law Office