(By You Yunting) According to press reports, the U.S. Department of Justice has launched an investigation into Nvidia’s roughly $20 billion transaction with AI inference chip company Groq. The question is whether the deal’s structure — a technology licence coupled with the hiring of key personnel — was deliberately designed to circumvent antitrust review. Deals of this kind are now common in AI. Off the top of my head: Microsoft and Inflection AI, Google and Character.AI, Google and Windsurf, Meta and Scale AI. Why do some attract scrutiny while others do not? That is what I want to discuss today.
The structure Nvidia used is the one now widely used among Silicon Valley giants in chips and AI: the acquihire. To prevent monopolies by large companies, antitrust laws in most jurisdictions require that acquisitions of another company’s equity or assets exceeding a certain threshold be notified for antitrust review, for example, in China, the concentration-of-undertakings review conducted by the State Administration for Market Regulation (SAMR). To circumvent that review, many recent AI deals in the United States have adopted the acquihire model: no purchase of the business or assets, but hiring of the core team; no outright transfer of intellectual property, but a long-term IP licence; no extinction of the target’s legal entity, which survives and continues to operate independently.
There are many such deals and only so much enforcement capacity, so regulators pursue the cases most likely to be unlawful. The highest-risk transactions, in my reading, share three features: they are the first of their kind; the acquirer is a repeat antitrust offender; or the acquirer is the firm with the greatest market power in the relevant market. Nvidia falls into the last category. Let me take the deals above one by one:
The 2024 Microsoft-Inflection AI deal. In 2024 Microsoft paid $650 million to Inflection AI, a company known for emotionally intelligent AI interaction, for a non-exclusive licence to its models, not for the business. The talent it wanted was Mustafa Suleyman, who had led AI projects at DeepMind and Google. The relevant markets can be defined as the development and supply of foundation models, consumer AI chatbots, and enterprise AI models; at the time, Microsoft held no dominant position in any of them; otherwise it would hardly have needed to invest more than $13 billion in OpenAI.
Even so, this deal was investigated by the U.S. Federal Trade Commission. In my view, the main reason is that this was the first acquihire in AI. In the view of Lina Khan, then FTC Chair, the first signs of unlawful conduct in an emerging field had appeared, and other companies might follow suit; an investigation was therefore warranted. The investigation remains unresolved.
The 2024 Google–Character.AI deal. Google’s target at Character.AI, a company focused on AI role-play interaction, was not the business but the core team of Noam Shazeer and Daniel De Freitas — both former Googlers, and Shazeer a co-inventor of the Transformer architecture on which large language models are built. The relevant market for this deal is principally consumer-facing AI chatbots, where a Google acquisition would not normally be especially sensitive. But Google has a chequered antitrust record from the mobile-internet era: it has lost at least eight antitrust cases in recent years — four in the United States, covering default search agreements, ad platforms and the app store; four in the EU, covering comparison shopping, search pre-installation, advertising and the app store. The U.S. Department of Justice is therefore highly alert to Google’s expansion in emerging AI markets, and its investigation likely targets whether Google would leverage its existing dominance to quickly build a monopoly in the new AI field.
The 2025 Google–Windsurf deal. In 2025 Google took the same approach with AI coding company Windsurf—hiring the core team, while Windsurf’s remaining assets were acquired by Cognition. The relevant markets for this deal are AI code generation, AI coding assistants and agentic coding. But the leaders in this market are Anthropic and OpenAI and Google’s frontier models hold no advantage.
The June 2025 Meta–Scale AI deal. Meta paid $14.3 billion for 49% of Scale AI and hired its CEO, Alexandr Wang, to run Meta’s AI business. Scale AI focuses on data labelling, converting raw data that machines cannot learn from directly into the high-quality annotated material used to train large models. After the deal, Scale AI continues to operate independently, and Wang remains on its board. This deal is distinctive in being a vertical integration: the relevant market is the upstream input market for AI training data labelling, data curation and data production, together with the vertical relationship that creates. Meta is not strong in AI, however, so even if the deal produces vertical foreclosure and informational advantages, it would not necessarily change the competitive landscape quickly. In addition, Google moved to scale back its cooperation with Scale AI once the deal closed, which shows that Scale’s market power is limited and that there are substitutes for its services.
The Nvidia–Groq deal. Reportedly, in its $20 billion transaction with AI chip startup Groq, Nvidia bought a licence to Groq’s inference chip technology and took on its founder, former Google employee and TPU co-inventor Jonathan Ross, together with the core team. Groq offers a low-latency, cost-efficient AI inference compute platform built on its proprietary Language Processing Unit (LPU). Under new management, Groq has since raised a fresh round of financing and continues to run its GroqCloud inference business independently. The relevant market here is AI-specific inference chips and AI compute accelerators.
What sets this deal apart is Nvidia’s already formidable market power in AI chips: its ecosystem lock-in, developer stickiness and cluster solution have become almost the industry default. Groq’s LPU architecture, purpose-built for inference, could have become a serious challenger to Nvidia in that market. Instead, its core team and key technology were bought out before it could grow. For end users of chips, such as AI companies, cloud providers and internet platforms, Groq’s existence meant one more option when procuring AI accelerators. This deal erodes their bargaining power further and may leave them paying more for these services. When the strongest company in a market buys out the core team of the start-up best placed to challenge it on an adjacent technological path, at a price this high, regulators cannot stand aside.
The August 2026 Nvidia–Poolside deal. By contrast, Nvidia’s $6 billion deal with AI coding startup Poolside in August 2026 is relatively less likely to be investigated. Poolside’s core business is building specialized AI foundation models for software engineering. The relevant markets for this deal are AI code generation, AI coding assistants, and open-weight open-source foundation models. Nvidia’s core business is AI chip R&D, so even with coding and model technology in hand, it is unlikely to alter competitive dynamics in those markets in the near term. Regulators may well leave it alone.
Acquihires deliberately structured to evade merger control are within Chinese regulators’ reach too: SAMR can require such transactions to be notified. Pursuant to the SAMR’s Provisions on the Review of Concentration of Undertakings, two provisions are relevant here:
- Where a concentration falls below the notification thresholds, but there is evidence that it has or may have the effect of excluding or restricting competition, SAMR may require the undertakings to notify and shall so inform them in writing.
- A concentration also arises where an undertaking acquires control of, or the ability to exert decisive influence over, another undertaking by contract or otherwise. The factors for decisive influence include the purpose of the transaction and future plans, the appointment and removal of the target’s senior management, and whether significant commercial relationships or cooperation agreements exist between the parties. This captures precisely the questions an acquihire raises: whether the object was to acquire technology and talent or to remove a competitive threat; whether the target’s core team will be absorbed by the acquirer; and whether the premium paid for the target’s technology far exceeds a normal arm’s length price.
That said, acquihires are engineered to exploit the gaps in black-letter antitrust rules, so blocking one is not easy. But if a regulator is determined to block a deal, the parties have little recourse — antitrust enforcement is, at bottom, political, and where antitrust law cannot reach, other laws can. Meta’s acquisition of Manus is a case in point. Manus had re-domiciled in Singapore, in what appeared to be an attempt to circumvent China’s meger control review; yet China’s National Development and Reform Commission and Ministry of Commerce, acting under the Foreign Investment Law and the Measures for the Security Review of Foreign Investment, blocked a transaction that had already closed.
Disclaimer of Bridge IP Law Commentary
Short Link: