Alexandros Kazimirov discusses how Nvidia’s quasi-merger with Groq resembles a familiar pattern of regulatory evasion that Google, Microsoft, and Amazon have adopted with emerging artificial intelligence companies. He notes that his proposed remedy that was available to antitrust enforcers in the large language model market is not applicable to chip manufacturers like Nvidia.
For a time, Nvidia was viewed as Big Tech’s outlier. Unlike Google, Microsoft, Amazon, and Meta, it did not rely on software prowess alone, positioned itself further away from the end user in artificial intelligence, and refrained from pursuing aggressive acquisitions of emerging competitors in the way that its peers did. But Nvidia’s latest deal with Groq, a company that builds tech infrastructure for large language models, indicates growing eagerness to protect its market power in graphic processing units (GPUs) by adopting Big Tech’s corporate devices.
That device takes the form of a quasi-merger which is a combination of an acquihire with a licensing deal. Like an acquihire, the acquirer prioritizes the hiring of senior management and software developers. Unlike an acquihire, however, the startup survives the acquisition. In turn, a quasi-merger splits the startup in two to maximize extraction value and shield the acquirer from antitrust enforcement at the cost of some of the startup’s founder, employee, and investor interests. The acquirer then uses license agreements as a means to cash out the startup’s venture capital investors, offer stock redemptions to employees, and provide a temporary lifeline to the surviving entity. This new transactional device has the potential to be both cooptive and synergistic. The hiring of software engineers can improve the acquirer’s products and enhance its innovation processes, but it can also preempt an emerging competitor from becoming a force of disruption, resembling a killer acquisition.
Antitrust enforcers struggle to prevent preemption because foreseeability in nascent competition is inherently uncertain. Digital networks evolve and firms fail. The Federal Trade Commission’s recent defeat in its case against Meta for monopolizing personal social media through its acquisitions of Instagram and WhatsApp offers a cautionary tale. Between the uncertainty of challenging an acquisition early or revisiting a consummated merger after many years, market dynamics may render the government’s efforts either too speculative or simply too late. Still, certain clues can be informative in discerning whether quasi-mergers are designed to achieve anticompetitive goals.
In 2024, Big Tech companies engaged in quasi-mergers with three AI startups: Character, Inflection, and Adept were subsumed by Google, Microsoft, and Amazon, respectively. As a consequence of those transactions, the acquirers took a non-exclusive license of the startups’ intellectual property, hired most or all of their employees, and forced them to compete further downstream, developing applications instead of large language models. These transactions resembled the captain and chief officers abandoning ship, taking all the supplies, and leaving the crew to fend for themselves in stormy seas. Since then, out of the three ships only Character has stayed afloat.
In a paper last year, I examined the question of market proximity between large language models and search engines. I advocated for robust antitrust enforcement based on prioritization of risks and contextual comparisons of evidence, and proposed separating tech incumbents from AI frontrunners when they subsume other smaller competitors. The idea was simple: for every developer acquired through a quasi-merger, the acquirer should proportionately relinquish control over non-merger companies whose products it has previously relied on. If Google, Microsoft, and Amazon acquire engineering talent to roll out their own large language models, they diminish the need to rely on such technology from OpenAI and Anthropic. Permitting Big Tech companies to maintain their influence over frontrunners like OpenAI and Anthropic while subsuming smaller AI companies would practically ensure Big Tech’s entrenchment in a new market. The public servants of the U.S. Department of Justice were intrigued enough to launch an investigation into Google’s quasi-merger with Character, which arguably seemed the most anticompetitive out of the three.
But if the antitrust enforcers were intrigued then, they should be concerned now. In many aspects, Groq was in a much stronger position than Character, Inflection, and Adept. Groq raised $750 million at a valuation of approximately $7 billion a few months ago—more than Character, Inflection, and Adept combined. It was a mature company in its tenth year with reportedly no plans to exit, unlike Character, Inflection, and Adept. It was expecting revenue of $500 million and enjoyed growing momentum neither of which is indicative of a failing firm. Nvidia’s price premium to Groq’s investors—reportedly $20 billion—was much higher than what Google offered to Character’s investors.
There is also little doubt about Nvidia and Groq being direct competitors. Both were building infrastructure required to run large language models. Both were competing for large language model developers who rely on these inputs. Groq’s flagship product, the Language Processing Unit, was running inference-as-a-service as an alternative with improved performance over Nvidia’s GPUs. It is precisely this type of product complementarity that reinforces the argument of technological synergy here. But if that synergy does not materialize, large language model developers may have to incur the costs of a horizontal merger resulting in adverse unilateral effects which may undercut their efforts to diversify supply chains.
Despite these differences however, Nvidia’s adoption of the quasi-merger does not raise any new substantive questions. In theory, the combination of a high premium, direct competition, and the likelihood of Groq’s post-quasi-merger decline points toward a motive to preemptively acquire Groq to eliminate a threat to Nvidia’s market power. But it is also true that the hiring of Groq’s software engineers can speed up Nvidia’s process for large language models to complete inference-related tasks and alleviate the memory supply shortage and related price spikes. Whether their integration achieves the desired outcome and offsets anticompetitive harms is a familiar problem which has to stand the test of time.
What is more concerning is that assuming this quasi-merger warrants government action, the remedy that was theoretically feasible for large language model developers in the Character context is out of reach here. On top of the structural problems of antitrust enforcement when it comes to quasi-mergers—that the acquirer and the startup continue to operate nominally independent of each other and that the antitrust enforcers cannot use injunctive relief to restrict employee mobility—Nvidia doesn’t have the kind of horizontal relationships that Google and Microsoft have with Anthropic and OpenAI. Nvidia’s main advantage lies in its vertical partnerships with other Big Tech companies so the threat of sacrificing governance or product agreements with direct rivals rings hollow here. What is more, any scrutiny of Nvidia’s GPU infrastructure might imperil the progress achieved by large language model developers putting them at risk of staying behind foreign adversaries.
All together, these considerations make the prospect of government intervention here fairly improbable, not to say adventuristic. But whether such intervention would be justified is a different question. In ordinary times, Nvidia should have to explain why a simple non-exclusive license with Groq isn’t enough and why a quasi-merger between an emerging and an incumbent chipmaker is the optimal choice for innovation. Groq’s departing management should have to lay out a clear vision for Groq’s few remaining employees beyond a stock redemption. And finally, while patterns of creative regulatory evasion are becoming the new norm, they are not without flaws, and there are ways to enforce the antitrust laws without endangering access to critical tech infrastructure. It’s just that each market requires its own workable remedy.
Author Disclosure: The authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
