In December, Nvidia signed a $20 billion deal with Groq, the inference-chip maker whose hardware had become one of the few credible alternatives to Nvidia’s own. Groq’s founder, its president and much of its technical staff moved to Nvidia. Nvidia licensed Groq’s technology. No merger was announced, because no merger occurred: nothing changed hands that Hart-Scott-Rodino recognises as a transaction. There was no acquisition of voting stock, no purchase of assets crossing a dollar threshold, no filing, no thirty-day waiting period, no chance for the Federal Trade Commission to ask a single question before the ink dried. And yet, by any commercial measure, Groq as an independent competitive force had ceased to exist.
This is now a pattern rather than an anomaly. Meta hired Scale AI’s chief executive and took a large minority stake in the company last June while standing up its own research lab around the incoming talent; Scale AI subsequently cut roughly 700 full-time and contract roles. Google poached Windsurf’s two co-founders and its research team in July 2025 and licensed the coding-assistant technology it needed, leaving a shell that Cognition AI later picked up for what remained. Google’s $2.7 billion arrangement with Character.AI, which brought founder Noam Shazeer back in-house along with core technology rights, has been under Department of Justice review since mid-2025. Four deals, four different hyperscalers, one structure: hire the people who make the asset valuable, license the intellectual property you actually need, and leave the corporate entity behind as an empty container. On 4 February this year, Senators Elizabeth Warren, Ron Wyden and Richard Blumenthal wrote to the FTC and DOJ describing these as “de facto mergers” engineered to dodge review. FTC Chairman Andrew Ferguson has said the agency is “beginning to look very closely at how these things work.” His fellow commissioner Mark Meador went further, warning that firms may be acquiring talent “not to utilize it productively but to preempt rivals from accessing it.”
That is a serious allegation, and it may well be true in some of these cases. But treating the reverse acqui-hire purely as a loophole, something clever general counsels found and drove a truck through, understates what is actually happening. The more useful way to read it is as a category error built into merger law itself, one that the AI economy has simply made impossible to ignore.
A statute built for factories, applied to weights and people
Merger control, in the United States and in most jurisdictions modelled on it, was designed around a specific theory of what makes a company powerful: its assets, its market share, its balance sheet. The Hart-Scott-Rodino Act triggers review when voting securities or assets change hands above a set value, currently $133.9 million. That architecture assumes the thing worth regulating is ownership. It has no native concept for a company whose entire competitive value sits in forty engineers’ skulls and a set of model weights that can be licensed without anyone buying anything. When Nvidia pays $20 billion for a licence and a hiring spree instead of a company, it isn’t dodging a rule so much as discovering that the rule was never written with this kind of asset in mind. Antitrust law still asks, in effect, “who owns the factory?” The AI economy’s honest answer is closer to “nobody owns it, we just hired everyone who knew how to run it.”
This is precisely why the political reaction has been so uneasy. Regulators are not confronting a clear violation they can point to and prosecute; they are confronting a structural blind spot they built themselves, decades before anyone anticipated that market power could be concentrated by employment contract rather than share purchase agreement. DOJ official Omar Assefi has called these deals a “red flag,” which is candid but also revealing: a red flag is what you raise when you suspect a problem, not when you can cite the section of the code it violates.
Who actually bears the cost
The corporate press releases around these deals emphasise talent and technology. What they omit is who gets left behind. When the founders and the venture investors of an AI startup are made whole through a licensing payment and a hiring package, and the remaining company is sold off in pieces or wound down, the common shareholders and the option holders, generally the engineers who were not deemed essential enough to hire, are the ones absorbing the loss. This is a meaningful inversion of how technology exits have traditionally worked, and it should concern any CEO or founder currently building a company on the assumption that a strong acquisition offer protects the whole capital stack. It increasingly protects the top of it. Boards of AI startups, and the executives negotiating with hyperscalers on their behalf, would do well to treat this as a live structuring question rather than a hypothetical one: who exactly is being bought, and who is merely being left behind, in the term sheet under discussion.
India got here first, by accident and by design
The most interesting regulatory contrast is not between Washington and Brussels but between Washington and Delhi. India’s Competition (Amendment) Act, passed in 2023 and effective from September 2024, introduced a deal-value threshold, roughly ₹2,000 crore, that triggers review regardless of the target’s revenue or the formal structure of the transaction. It was written to catch exactly the scenario merger lawyers now describe in the context of Meta, Google and Nvidia: a high-value, low-turnover, data- or capability-rich deal that a traditional asset-and-turnover test would wave through. The Competition Commission of India’s own 2026 market study on AI explicitly flags vertical, data-driven transactions, including minority stakes and partnerships, as warranting scrutiny beyond conventional financial metrics, and cites the UK Competition and Markets Authority’s review of Google’s Anthropic partnership as the kind of precedent Indian regulators are watching. None of this was written with Nvidia-Groq or Meta-Scale AI specifically in mind; it predates all of them. But it means a structure currently invisible to US antitrust law would, in principle, already be reportable in India, and is already under active review in the UK.
For any multinational technology or industrial company, this is the part worth sitting with. The same transaction can be legally uneventful in one jurisdiction and a notifiable combination in another, not because the deal changed but because the underlying test for what counts as a merger differs by country. Corporate development teams have spent the past two decades building HSR and EU merger-filing muscle memory. That muscle memory is now an incomplete map. A capability-acquisition strategy engineered around avoiding one regulator’s tripwire will not automatically avoid another’s, and the jurisdictions least associated with fast-moving tech regulation, India among them, are in some respects better instrumented for this specific manoeuvre than the one that invented it.
What this means at the top of the house
None of this requires a company to abandon talent-driven deals; it requires treating them with the governance seriousness their economic substance deserves rather than the light touch their legal form currently permits. Boards approving these arrangements should assume retroactive scrutiny is more likely, not less, as the political temperature around Big Tech AI consolidation rises through this year. Corporate development functions should be running multi-jurisdiction analysis on capability acquisitions the way they already do on conventional ones, rather than defaulting to US thresholds because the deal happens to be denominated in dollars. And any executive building a company that might one day be worth acquiring for its people rather than its balance sheet should assume the exit, when it comes, may not look like the exits described in the textbooks. The law will eventually catch up to what these transactions actually are. The only real question for the people signing them now is whether they are prepared for the version of events regulators settle on once it does.
