According to people familiar with the matter, speaking to the New York Times, the Justice Department sent Nvidia a formal demand for information about its Groq deal shortly after the deal was announced in December. No press conference, no dramatic announcement. Just a letter asking Nvidia to explain itself.
That detail tells you a lot. Regulators did not wait to see how the deal played out. They looked at the paperwork and had questions almost immediately.
If you write about AI agents for a living, as I do, this is the kind of story that seems like it belongs to lawyers and nobody else. I’d argue the opposite. This one sits very close to the machinery that makes your agents work.
What actually happened
Nvidia signed a $20 billion non-exclusive licensing agreement with Groq, a startup that designs inference chips. Reporting from PC Gamer describes it as the biggest deal Nvidia has ever done. The Justice Department is now investigating whether Nvidia deliberately structured that agreement to avoid antitrust review.
Read that carefully, because the distinction matters. The investigation is about the deal’s structure, not its content. Regulators are not currently arguing that the arrangement itself is illegal. They are asking whether Nvidia built it in a shape designed to slip past the review process that a deal of that size would normally trigger.
It’s the difference between being accused of speeding and being accused of taking back roads specifically to avoid the speed camera.
Why the structure question is the interesting one
Big acquisitions get reviewed. That is the deal companies accept when they buy each other outright. But a licensing agreement is not an acquisition. Neither is a non-exclusive arrangement, at least not on paper. Money moves, technology and rights move, but the corporate boxes stay where they were.
The DOJ’s apparent theory is that the effect can look a lot like a merger even when the documents say otherwise. Whether that theory holds up is the whole case.
The consequences are limited in one important way. Based on the reporting, if the DOJ finds fault, it may issue a fine but would not require the deal to be undone. So the $20 billion arrangement almost certainly stands. What is at stake is money and precedent, and honestly, the precedent is the bigger number here.
The agent connection nobody mentions
Here is where this touches the tools you actually use.
Every time an AI agent does something for you, that action runs on inference. Training is the expensive one-time process of building a model. Inference is what happens every single time the model gets used. When your agent reads an email, drafts a reply, checks a calendar, and then decides what to do next, each of those steps is an inference call.
Agents are unusually hungry in this respect. A chatbot answers your question once. An agent loops, reasoning through steps, checking its work, and calling tools. One task can mean dozens of inference calls. That makes inference cost and speed the two things that decide whether agents stay a demo or become genuinely useful.
Groq designs chips built specifically for inference. Nvidia dominates the chips used for training. A $20 billion arrangement between those two is a story about who controls the layer your agents will run on for the next several years.
What this means if you are not a lawyer
A few practical takeaways:
- Inference is becoming the main event. The money is moving toward the part of AI that serves users, not just the part that builds models. That shift favors agents.
- Consolidation shapes your bill. Fewer independent suppliers in the inference space generally means less pricing pressure. Your agent’s per-task cost traces back to decisions like this one.
- Deal shape is the new frontier. Licensing agreements, non-exclusive arrangements, and talent deals are how large AI companies are growing. Regulators are clearly paying attention to that pattern now.
- Nothing changes tomorrow. The deal is not being unwound. Your tools keep working exactly as they did yesterday.
What I’m watching
Two things. First, whether the DOJ’s structure-focused approach produces a result, because that would tell every other AI company which arrangements draw scrutiny and which do not. Second, whether inference capacity stays competitive enough to keep agent costs falling.
I want to be clear about the limits of what is public. We know an investigation is open, we know a formal information demand went out, and we know the focus is structural. We do not know what Nvidia has said in response, and we do not know the timeline.
Still, when regulators react to a deal within weeks of its announcement, they are signaling that the rules for how AI companies combine forces are being written right now. Agents are the product at the other end of that plumbing. Worth following, even if you never read a single filing.
🕒 Published: