Nvidia-Groq AI Deal: If It Quacks Like a Merger and Acts Like a Merger …  

Can an AI licensing and talent arrangement spark the very harms to competition merger law was designed to prevent?

The U.S. Department of Justice is reportedly investigating whether Nvidia structured its licensing transaction with AI-chip startup Groq in a way that avoided the type of antitrust review typically associated with major acquisitions. Reuters, following reporting first published by the New York Times, put the transaction value at roughly $17 billion, while other public reports placed the figure closer to $20 billion. Reuters reported that Nvidia announced a non-exclusive license to Groq’s chip technology, hired several Groq executives, including founder Jonathan Ross, and later received a formal DOJ request for information about the arrangement. 

The reported investigation is significant because it spotlights a broader pattern Mogin Law has been tracking in our AI Deal Table: major technology companies are increasingly pursuing AI capabilities through licensing arrangements, talent transfers, infrastructure commitments, minority investments, and strategic partnerships rather than formal mergers. Mogin Law has been tracking transactions and investments inside the AI infrastructure space: data centers, chips, cloud capacity, energy, cooling, copper, and other physical inputs needed to develop and deploy advanced AI systems. That makes Groq important not only as another AI startup, but as a company positioned in the infrastructure layer that can determine how trained AI models reach users. 

AI’s Value Chain: Development, Training, and Inference 

Groq is a privately held AI infrastructure company founded by former Google engineers and known for developing the Language Processing Unit, or LPU, a specialized chip designed to accelerate AI inference. Unlike OpenAI, Anthropic, or xAI, Groq does not primarily develop consumer-facing AI models, but focuses on the computing layer that enables AI systems to deliver answers, images, code, and other outputs to users in real time. 

This is an important difference. AI systems generally depend on two different compute functions: training and inference. Training is the resource-intensive process of building a model by exposing it to large amounts of data so it can recognize patterns and generate useful outputs.  

Inference comes later, when the trained model is put to work. Each time a user asks a chatbot a question, prompts an image generator, writes code with an AI tool, or runs an AI-powered application, the system uses compute infrastructure to apply the trained model and produce a response. Put differently, the system uses the trained model to “infer” the most likely, useful, or appropriate output. It is not thinking like a person, but it is applying learned patterns to your prompt. Groq competes in this second layer. Its business is tied to the chips, servers, and software systems needed to run trained models quickly and efficiently at scale. 

This supports viewing AI competition across several related layers of the value chain: model development, training compute, and inference compute. Those layers are related, and they may overlap, but they raise different competitive questions.  

• Model development concerns who builds the AI systems.  
• Training compute concerns who has access to the chips, data centers, cloud capacity, power, and capital needed to build them.  
• Inference compute concerns who controls the infrastructure needed to run trained models for users at scale.  

Whether any of those layers constitute legally cognizable antitrust markets would, of course, require detailed market-definition analysis. 

That layered structure has analogues in other technology industries. Semiconductor markets distinguish among chip design, fabrication, packaging, and distribution. Cloud computing separates infrastructure, platforms, and software applications. Internet and media markets distinguish among content creation, distribution infrastructure, user-facing platforms, and access points. AI may develop similar layers, with competitive pressure arising not only where models are built, but also where they are trained, deployed, and delivered to users. 

This not just a U.S. story. Concerns about acqui-hires and “stealth mergers” have also been raised by competition agencies in Europe.

AI Deals That Capture Invaluable Assets 

The Nvidia-Groq arrangement warrants study because it sits between categories. Rather than an outright purchase, Nvidia reportedly secured access to Groq’s technology while hiring key personnel associated with the company. That makes the reported DOJ inquiry less about the transaction’s label and more about the deal’s practical effects. 

The reported investigation arrives during a period of intense scrutiny of investments, partnerships, licenses, and talent-acquisition arrangements involving major AI companies. Mogin Law has examined this issue through the lens of Google’s reported $2.4 billion Windsurf arrangement, in which Google obtained a nonexclusive technology license while hiring Windsurf’s CEO, cofounder, and several AI researchers to join Google DeepMind, without acquiring Windsurf outright. Similar questions have been raised about Microsoft and Inflection AI, Amazon and Adept, Google and Character.AI, and Meta and Scale AI. The common thread is access to AI talent, technical know-how, IP, compute resources, data, distribution, and strategic capabilities. 

Legislators recognize the problems. Senators Elizabeth Warren, Ron Wyden, and Richard Blumenthal have urged federal antitrust enforcers to examine whether some AI reverse-acqui-hire arrangements function as de facto mergers.

FTC Chairman Andrew Ferguson has indicated that the agency is examining these transactions, and Acting Assistant Attorney General Omeed Assefi calls them red flags. 

David T. Wong’s Yale Law Journal analysis supplies one legal framework for that concern. Wong argues that Section 7 of the Clayton Act should not turn only on whether a company bought stock or formally acquired another company. If human capital, technical know-how, and related capabilities are economically significant assets, he argues, then a transaction that transfers those assets may warrant Section 7 scrutiny even when the startup remains legally independent. 

Of course, whether these arrangements are good or bad for competition will depend on the facts of each case. They may be procompetitive when they provide startups with capital, commercialization opportunities, compute resources, or access to larger platforms while preserving independent rivalry. They may raise antitrust concerns when they remove an emerging competitive threat, transfer the startup’s core capabilities to a dominant firm, reduce customer choice, or allow market power to accumulate through contracts and hiring rather than through a reviewable merger. 

The Nvidia-Groq matter appears to place those questions squarely before federal antitrust authorities. 

Nvidia Hails the General Value of the Arrangement 

A spokesperson for the Santa Clara, Calif.-based Nvidia said the Groq arrangement demonstrates “the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers.” There is much we don’t know, however, like the terms and rights of the intellectual property license, whether any governance rights accompanied the transaction, or the extent to which the arrangement affected future competition between the companies. 

A Potential Test Case for the AI Economy 

This could become an important test case in the development of antitrust policy regarding artificial intelligence. 

Traditional merger law was largely designed around stock purchases, asset acquisitions, and corporate combinations. Modern AI transactions can look very different. As this deal demonstrates, valuable technology may be transferred through licensing arrangements. Competition may be affected when engineering teams move between companies. Strategic control can sometimes emerge through a network of contracts and commercial relationships rather than a formal acquisition agreement. Whether existing antitrust frameworks adequately capture those realities remains an open question, but it appears federal enforcers may intend to answer it.   

The framework described above suggests that AI competition may be developing across several related but potentially distinct value-chain segments: model development, training compute, and inference compute. Nvidia is already central to training compute because of its dominant position in AI chips. Groq, by contrast, has positioned itself as a challenger in inference compute. If inference capacity becomes concentrated among a small number of firms, competition concerns could arise even if multiple AI models remain available to users. 

Control of inference capacity may become as strategically important as control of the models. The Nvidia-Groq investigation therefore fits within a larger competition-policy question: whether dominant firms can consolidate AI control points through infrastructure deals, licenses, hiring, and strategic relationships even when no conventional acquisition occurs. 

Edited by Tom Hagy, Editor-in-Chief, Mogin Law Blog.