Why Merger Law Misses Frontier AI's Ownership Structure

The views expressed here are the author's own and do not represent the views of Digital Policy Alert or the St. Gallen Endowment for Prosperity Through Trade.

Since late 2023, competition authorities in four jurisdictions have reviewed eight transactions tying the world's leading AI labs to a handful of large technology companies, and nearly all of them were found to fall outside of merger law. A small number of firms provide the capital, chips and cloud infrastructure that frontier AI labs run on, and merger control was designed to identify control, not the partial stakes, compute agreements, and board-adjacent access through which power over frontier AI labs is increasingly being consolidated.

The UK's Competition and Markets Authority (CMA) repeatedly concluded that no relevant merger situation had arisen, clearing Microsoft's deals with OpenAI and Mistral and Amazon's partnership with Anthropic, while also clearing Microsoft’s deal with Inflection on the merits. Germany's Federal Cartel Office, the country's national competition regulatory agency, concluded that Microsoft-OpenAI’s cooperation was not subject to merger control. Brazil's Administrative Council for Economic Defense (CADE) cleared Nvidia-Run:ai, Microsoft-Mistral and Google-Character.AI mergers on the same grounds in 2026. EU member states’ referral procedure concerning Microsoft’s acquisition of certain assets of Inflection was also closed.

The one exception: Brazil separately ordered a merger review of the Microsoft-Inflection deal, invoking a discretionary provision under Article 88(7) of the Competition Law that allows review of transactions falling below the revenue thresholds for mandatory notification, on the view that the arrangement's economic substance mirrored a conventional acquisition despite its unconventional form.

Why these deals don't count as mergers

Merger review is organized around a threshold inquiry: does the deal give one company enough power over another’s strategic decisions to count as control? Where the answer is affirmative, the transaction is notifiable and subject to scrutiny. Where it is not, the transaction generally proceeds, irrespective of how much influence has been acquired short of that threshold. This threshold has functioned adequately for the class of transactions that merger law was designed to capture: outright acquisitions, majority stakes, mergers of equals. Frontier AI's ownership structure presents a materially different case, characterized by influence accumulated through minority stakes, exclusive supply arrangements and board-adjacent access that regulators have repeatedly found significant, yet insufficient to establish control.

The reasoning the UK's CMA published on Microsoft-OpenAI shows the mechanism at work. The authority found that Microsoft held a high degree of material influence over OpenAI's commercial policy — arising from its multibillion-dollar investment and informal governance access, its status as OpenAI's exclusive compute supplier, and its exclusive license to OpenAI's IP — but concluded this fell short of de facto control. In reaching that conclusion, the CMA placed particular weight on recent developments that reduced OpenAI's dependence on Microsoft, including new third-party funding from investors such as Thrive Capital and a revolving bank credit facility, together with a renegotiated compute agreement under which Microsoft moved from exclusive supplier to a right-of-first-refusal model, enabling OpenAI to build additional capacity through the SoftBank, Oracle, and MGX-backed Stargate project.

Supplier diversification is generally viewed as procompetitive, and in this case it also happened to be the evidence that kept the deal outside merger jurisdiction. That evidentiary basis has since shifted. In October 2025, Microsoft and OpenAI entered into a further restructuring under which OpenAI reorganized into OpenAI Group PBC, Microsoft disclosed that it would hold an investment representing approximately 27 percent of the company on an as-converted diluted basis following the recapitalization, and Microsoft relinquished its right of first refusal to serve as OpenAI's compute provider. The dependency the CMA examined never resolved into independence. It evolved into a disclosed equity stake and a restructured set of contractual rights after two years of renegotiations that remained short of the threshold for control at each stage.

The Microsoft-OpenAI outcome is not an isolated case. Merger control asks whether one party has gained lasting power to determine another's decisions, a test built around discrete acquisitions of control. The CMA has observed that AI ecosystems are increasingly shaped by minority investments, long-term compute agreements and strategic partnerships that may fall short of the legal threshold for merger control individually, while nonetheless potentially reinforcing market power cumulatively. It is also noted that merger law has never claimed to capture every form of commercial influence, and that extending it further risks catching ordinary investment activity.

Regulators have begun looking more closely at the pattern. The US Federal Trade Commission opened a 6(b) study in early 2024 ordering five companies to provide information about their AI-related investments and partnerships, and the US Senate separately sought information from Nvidia about its agreement with Groq over concerns about competition in the AI semiconductor sector.

How these deals are structured is itself instructive: minority stakes, exclusive licensing, and compute commitments can deliver much of what an acquisition would provide, including board influence, priority access and first-look rights in future funding rounds, without triggering notification obligations or divestment risk, and advisory firms increasingly describe transaction structure itself as a primary tool for managing merger-control risk. It raises a fair question for policymakers: whether a legal test with a known, precisely defined boundary can meaningfully constrain conduct that is, by design, structured to stay clear of it.

The competitive stakes

This isn't a purely legal curiosity. If a handful of firms can hold equity and supply relationships across most of the labs capable of training frontier models, without any single relationship ever tripping the control threshold, the practical result looks a lot like concentration without a chokepoint anyone can act on. Fewer genuinely independent frontier labs get built, because the fastest path to compute and capital runs through one of the same three or four providers. Later entrants face a market where their most credible funding sources are also the incumbents' commercial partners. These arrangements can also give a compute supplier visibility into a competing lab's technical roadmap and usage patterns that an arm's-length investor would never have, simply through the ordinary mechanics of billing and board access. Further, infrastructure choices made early, while a lab is still small, harden into lock-in well before any authority gets a chance to weigh in.

Conduct enforcement is the more active response, but it arrives after that structure is already in place. Meta has drawn conduct enforcement attention from four authorities at once. In January 2026, CADE ordered Meta to suspend its new WhatsApp Business Solution Terms, amid concerns that the changes could limit market access, exclude competing AI providers, or favor Meta's own Meta AI tool. In June 2026, the European Commission ordered interim measures against Meta to restore free access to the WhatsApp Business API for rival AI assistants, finding that its fee-based access policy was tantamount to a ban and risked entrenching its dominance in the AI assistant market.

Italy and Türkiye reached the same conclusion independently: Italy ordered Meta to suspend its WhatsApp terms restricting third-party AI chatbots, citing likely harm to competition and customers, though it later closed its own probe when the Commission expanded its EEA-wide investigation to cover Italy. Türkiye separately ordered Meta to open WhatsApp to third-party AI chatbots on equal terms while it investigates the same conduct.

Google faces a parallel set of cases over content. In 2025, the European Commission opened a case against Google over its use of publisher and YouTube content for AI purposes, examining whether AI Overviews and AI model training leverage its market dominance to extract content without compensation while denying rivals equivalent access. Brazil escalated a probe into whether Google's AI Overviews use journalistic content without compensation. Separately, South Korea launched a market-wide study surveying dozens of AI developers and integrators over competition and consumer harm in AI services.

A few jurisdictions have begun exploring tools that sit closer to that earlier point. The UK's willingness to examine certain partnership and investment arrangements under its material influence standard, rather than limiting scrutiny to formal acquisitions, is one such adaptation, although the Microsoft–OpenAI decision illustrates both what that standard can capture and where its legal boundaries lie. The CMA acknowledged that there is no "bright line" between factors giving rise to material influence and those giving rise to de facto control. The decision therefore demonstrates both the flexibility and the limits of the material influence test: substantial commercial dependence may justify close scrutiny, but merger jurisdiction ultimately depends on whether the statutory threshold for material influence or control is satisfied.

Ex-ante regimes that impose obligations on firms in advance of their market actions rather than penalizing violations after they occur, like the EU's Digital Markets Act, were designed to address a different problem, gatekeeper conduct, rather than the accumulation of gatekeeper-adjacent power through investment. The DMA's obligations attach to firms that meet defined scale and entrenchment thresholds and have been formally designated as gatekeepers; it says nothing about the minority stakes, compute agreements, or exclusivity arrangements through which a non-designated firm might come to depend on one. Some proposals extend the DMA's logic upstream by designating cloud computing providers and foundation models as core platform services. More structural remedies proposed include barring cloud providers from also competing in foundation models, preventing chip designers from investing in AI development companies, and applying heightened scrutiny to mergers, acquisitions, and partnerships involving Big Tech hyperscalers and private-equity firms.

Whether that framework, or something comparable, could be adapted to address structural dependency at an earlier stage remains an open question that regulators have not yet had occasion to test.

The blind spots of merger law

The pattern holds across every jurisdiction examined here, and it has held for more than two years: a legal test built around discrete acquisitions of control is being applied to a market where power accumulates through partial stakes, compute agreements, and board-adjacent access. Each individual deal has survived review under the existing standard, and the industry's ownership structure has continued to consolidate around the same handful of relationships regardless.

Were merger control, or an ex ante regime modeled on the DMA, extended to capture accumulated dependency rather than discrete acquisitions of control, regulators would gain the capacity to intervene while a laboratory's strategic dependencies are still emerging, before compute and capital lock-in harden, and before a small number of capital and infrastructure providers acquire privileged visibility into the strategic direction of laboratories with which they may also compete. That gain would carry a genuine cost: a standard sufficiently broad to identify dispersed influence is also sufficiently broad to encompass ordinary investment activity, requiring regulators to develop new evidentiary frameworks capable of distinguishing between the two.

Absent such an extension, the present pattern is likely to continue. The frontier AI ecosystem is likely to remain concentrated around the same small group of capital and compute providers, in breach of no single rule, while conduct enforcement, of the kind now underway against Meta and Google, remains the principal instrument available to address competitive concerns only after those dependencies have become established.