The DOJ Wants to Know Why Nvidia’s $20B Groq Deal Wasn’t a Merger

The US Justice Department is examining whether Nvidia structured a $20 billion licensing agreement with AI chip startup Groq to avoid antitrust review. The Nvidia Groq antitrust probe was reported on 10 September 2026. Investigators have sent Nvidia a formal demand for information.

Groq presented the December arrangement as a non-exclusive licence. It gave Nvidia access to chips optimised for AI inference.

Two Groq executives moved to Nvidia as part of the same deal. Founder and chief executive Jonathan Ross and chief operating officer Sunny Madra both joined. Groq remained a separate company.

What the Nvidia Groq antitrust probe is testing

Merger review in the United States is triggered by acquisitions above certain thresholds. A licence is not an acquisition. Neither is hiring executives.

The question is whether those pieces together achieve what an acquisition would, without the filing that an acquisition requires. Regulators have looked at similar structures across the AI sector over the past two years.

The department opened the probe soon after the arrangement was announced, according to the reporting. That timing suggests the structure itself drew attention rather than any later conduct.

Why inference chips are the contested ground

Training a model and running it are different workloads. Training happens once and demands enormous parallel compute. Inference happens every time someone uses the model.

Inference volume grows with adoption. That makes inference silicon the larger long-run market, and the one where challengers have had the clearest opening against Nvidia.

Groq built its business on that opening. A deal that gives Nvidia access to its inference-optimised designs while moving its leadership across narrows the field, at least on paper.

None of that establishes a violation. A formal demand for information is an investigative step, not a finding.

The acqui-hire pattern regulators keep meeting

The structure at issue has become common across artificial intelligence. A large company licenses a startup’s technology, hires its senior leadership, and leaves the startup standing as a separate entity.

Each element looks ordinary on its own. Licences happen constantly. Executives change jobs constantly. Neither triggers a merger filing.

Combined, they can transfer the two things that matter most: the technology and the people who built it. What remains at the startup is a name, a cap table and whatever staff stayed behind.

Regulators in the United States, the United Kingdom and the European Union have all examined arrangements of this shape since 2024. Outcomes have varied, and several closed without action.

That history explains the department’s interest here without predicting where it lands. Structure alone does not establish harm, and the law asks about effects on competition rather than about form.

How this fits Nvidia’s wider regulatory picture

Nvidia is under pressure on several fronts at once. Export controls shape what it can sell and where, a subject we covered in our report on the chip export loophole.

Supply is the second front. Our note on the memory chip shortage sets out the constraint running through the whole accelerator market.

Litigation across AI companies is the third. Our coverage of Tennessee’s patent suit against Anthropic shows how quickly legal exposure has spread through the sector.

What happens after a demand for information

A civil investigative demand compels documents and answers. Companies typically negotiate scope before producing anything.

Most such investigations close without action. Some end in a consent agreement that changes how a deal is structured. A minority reach court.

Timelines run long. Antitrust reviews of this kind typically run for quarters rather than weeks.

Publicity is its own factor. Companies frequently adjust terms once an investigation becomes public, without ever conceding a problem.

Nothing in the reporting suggests the department has reached a view. It has asked for records, which is where these matters normally begin.

Nvidia has not commented publicly in the reporting reviewed here. Neither has Groq. Both descriptions of the deal so far come from the companies’ own framing at announcement.

That framing deserves scrutiny precisely because it is theirs. Calling a licence non-exclusive tells buyers Groq can still sell to others. It says nothing about whether Groq retains the people needed to do so.

Customers watching the probe face a practical question rather than a legal one. Anyone who designed a deployment around Groq silicon wants to know the roadmap survives the departures.

A Brazilian startup founder running inference at scale cares about supply and price, not about filing thresholds. Regulatory outcomes reach that founder slowly, if at all.

What people are asking about the investigation

• What is being investigated? Whether Nvidia structured a $20 billion licensing deal with Groq to avoid antitrust review.
• What was the deal? Groq described it in December as a non-exclusive licence giving Nvidia access to chips optimised for AI inference.
• Did Nvidia buy Groq? No. Groq remained a separate company, though its chief executive and chief operating officer moved to Nvidia.
• What step has the department taken? It sent Nvidia a formal demand for information.
• Does this mean a violation occurred? No. An investigative demand is a request for records, not a finding.
• Why does AI inference matter here? Inference workloads scale with usage, making inference chips the larger long-term market.

Continue reading on Tamara News

See our reporting on the Nvidia chip export loophole, the memory chip shortage and Tennessee’s patent case against Anthropic.

Sources

• Bloomberg — DOJ Probes Nvidia’s $20 Billion License Deal With Groq on Antitrust Concerns. bloomberg.com
• The Daily Guardian — US DOJ probes Nvidia’s licensing deal with AI startup Groq. thedailyguardian.com