Ten billion dollars. That’s what Microsoft has injected into OpenAI since 2019. Amazon followed with four billion into Anthropic. These figures do not describe buyouts, nor even classical equity stakes. The FTC has documented something more subtle: these investments come with contractual obligations that force OpenAI and Anthropic to spend their computing budgets on Azure and AWS, the clouds of their respective investors. The laboratories remain formally independent. But their research infrastructure is bound.

This coupling between capital and infrastructure creates a form of dependency that current competition law cannot name with precision. Not a merger. Not an acquisition. Something else.

The Essentials

  • Microsoft has invested over ten billion dollars in OpenAI since 2019; Amazon has committed four billion to Anthropic. Both deals include clauses that tie the laboratories’ computing expenditures to their investors’ clouds.
  • The FTC opened an inquiry into these agreements in 2024, documenting that they function as tying arrangements between capital and infrastructure.
  • Classical competition law, designed for mergers and market share, lacks tools suited to markets where the critical resource is computing power, not the product.
  • If these structures stabilize, the emerging AI laboratories of the next five years will have to choose between infrastructural dependency and major competitive disadvantage.

Computing as a Critical Resource

To understand why these agreements change market structure, one must start with a simple economic fact: training a large language model costs between one hundred million and one billion dollars depending on the scale aimed for. The bulk of this cost is computing — chips, cooling, electricity, software orchestration. These resources are concentrated in a few global infrastructures.

Computing power has become the rare resource of the century, in the sense that its distribution determines who can do what in AI. A laboratory without access to computing cannot train or deploy competitive models. This is precisely the control point that tying agreements lock down.

The mechanism works as follows. The investor provides capital. In exchange, contractually or through the terms of use attached to the investment, the laboratory commits to consuming its computing resources on the investor’s cloud. When OpenAI wants to train GPT-5, it does so on Azure. When Anthropic wants to run Claude, the servers are AWS’s. These obligations are not marginal: they cover a substantial portion of the laboratories’ operating expenses, which run into billions annually.

The FTC used, in its investigation documents, the term “tying arrangements.” In American competition law, tying describes the practice of selling a desired product on condition that another be purchased. The jurisprudence on tying is old and robust. The novelty here is that the “desired product” is not an ordinary good or service: it is capital at a time when capital conditions access to the technological frontier.

Three Infrastructures, One Funnel

The overall picture is striking. Microsoft is linked to OpenAI. Amazon to Anthropic. Google, for its part, acquired DeepMind in 2014 — a laboratory that had been founded independently in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman — and operates its own Gemini models, while having invested in Anthropic via Google Cloud — a laboratory that therefore maintains capitalistic relationships with both competing clouds. Meta and Apple operate their own infrastructures. At the top of the AI market, almost everything passes through three or four clouds.

This is not yet a monopoly. But it is a funnel structure. The best laboratories — those that attract the best researchers, that produce the best models, that define the technical frontier — are contractually bound to the infrastructures of hyperscalers. An emerging laboratory wanting to compete with OpenAI or Anthropic must either find capital from an investor willing to forgo cloud counterparties, build its own infrastructure, or accept similar conditions from a competing cloud.

The third option reproduces the problem. The second requires resources that only hyperscalers possess. The first — raising capital without infrastructural obligations — becomes structurally harder as tying agreements normalize this model. Mistral AI, in Europe, has so far maintained relative independence. But Mistral also signed a partnership with Microsoft Azure for the distribution of its models — a distinct commercial agreement, but one that illustrates the gravity of infrastructural pull. Even actors who resist end up orbiting the same centers.

What Competition Law Does Not Yet See

The FTC and the British Competition and Markets Authority have both signaled their concern about these structures. The CMA published a 2024 report on foundation models explicitly pointing to the risk of dependency between laboratories and clouds. The European competition authority has issued similar observations.

But the problem is instrumental. Competition law has two main tools: merger and acquisition control, and prohibition of abuse of dominant position. The first does not apply here — there is no acquisition. The second assumes that a dominant position is established on a defined market, which is difficult when AI markets are only a few years old and their boundaries remain unclear.

Tying agreements, in American law, can be attacked under Section 1 of the Sherman Act if one proves they substantially restrict competition. But jurisprudence requires demonstrating sufficient market power in the market for the tying product — here AI investment capital — and a restrictive effect on the market for the tied product — here cloud. Both demonstrations are possible, but demanding. The era of digital antitrust advances, but it moves on terrain that regulators have not yet fully mapped.

It is not that regulators lack will. It is that the instruments were designed for markets in physical products, homogeneous services, dominant positions measurable by market share. AI introduces a new variable: dominant position in nascent markets is built before markets exist, through control of the resources that will condition their emergence.

The Barrier to Entry That No One Formally Built

What makes these arrangements particularly difficult to regulate is their organic nature. Microsoft did not declare that it wanted to control the AI market by tying OpenAI to Azure. It provided capital at a time when OpenAI needed it, in exchange for commercial counterparties that had immediate business logic. Amazon did the same with Anthropic, when the laboratory was seeking to fund its accelerated development in the face of OpenAI competition.

Each agreement taken in isolation seems reasonable. Together, they draw a barrier to entry that no one formally constructed but that everyone contributes to erecting. Economists call this an endogenous structural barrier: it emerges from the individually rational behaviors of actors, not from a deliberate conspiracy.

Thomas Philippon, in his work on the concentration of American markets, has documented how this type of structure reduces competitive pressure in the long term — not through visible anticompetitive action, but through the accumulation of positions that make entry progressively less attractive. AI markets are reproducing this pattern at a speed incomparable to previous industries, because the critical resource — computing — is simultaneously costly, concentrated, and growing in importance.

Three Scenarios for the Coming Five Years

Long-term analysis must rest on what is measurable. Hyperscaler capital expenditures on AI datacenters reached approximately 220 billion dollars in 2024 according to aggregated estimates from financial reports by Microsoft, Amazon, Google, and Meta. This figure should double by 2027 according to sector analyst projections, under the assumption — worth naming — that computing demand continues to grow at the current pace. This assumption can be contradicted by a breakthrough in algorithmic efficiency, as DeepSeek illustrated in early 2025. But even in case of training cost reduction, inference costs — running models at scale — will remain considerable barriers.

In this context, three trajectories emerge. The first is silent normalization: regulators demarcate without intervening firmly, tying agreements multiply, second-tier laboratories integrate into the same structures to survive, and the AI market resembles the mobile ecosystem of the 2010s — plural in appearance, concentrated in infrastructure. The second is structural regulation: the FTC or European Commission impose functional separation conditions between laboratory investments and cloud purchasing obligations, requiring neutrality clauses. The third is disruption through efficiency: a generation of models trainable on less centralized infrastructures mechanically reduces the advantage of hyperscalers, as open source models have limited the premium of some proprietary actors.

These trajectories are not mutually exclusive. But their respective probability depends on a central factor: the speed at which regulatory instruments adapt to the real market structure.

What Regulators Are Concretely Doing

Inaction would be a misreading of the situation. The FTC published a report in 2024 that precisely maps these agreements. The British CMA developed an analytical framework for “foundational model markets” that explicitly recognizes computing as a critical resource and the risks of dependency. On the European side, while the Digital Markets Act has not yet formally extended its scope to AI infrastructure providers — AI as such not yet constituting a “core platform service” under the regulation — the Commission nevertheless issued in June 2026 a preliminary position on designating AWS and Azure as “gatekeepers” for their cloud services. This designation, which remains to be confirmed, concerns cloud services as such and not AI activities proper. No pure AI laboratory is targeted by such a designation.

At the academic level, economists like Fiona Scott Morton in the United States are working on analytical frameworks specific to multi-sided markets dominated by infrastructure. The idea gaining ground is that of “behavioral remedy” — imposing behavioral conditions on investors (no cloud obligations in AI investment agreements) rather than seeking to dismantle structures that do not have the classical characteristics of an acquisition.

Voices like those of Luigi Zingales and Daron Acemoglu, in their respective analyses of the capture of technological gains, point to the same recommendation: the issue is not preventing hyperscaler investments in AI, but ensuring these investments do not automatically translate into barriers to entry for actors not in their orbit. The distinction between financing innovation and controlling access to its infrastructure is the line that law must learn to draw.

The Question for Emerging Markets

This debate goes far beyond AI. It raises a fundamental question for markets in generative technologies: who can enter them, and on what terms? If capital-infrastructure tying agreements become normalized, the next laboratories — those working on computational biology, autonomous chemistry, next-generation reasoning systems — will face the same choice at their inception: seek capital from a hyperscaler, with the counterparties that come with it, or remain outside the perimeter where the technological frontier is defined.

This is where the issue transcends sectoral regulation. Emerging markets are those where competitive positions form before regulators have the data to intervene. The CMA and FTC acted quickly, by the standards of public institutions. But “quickly” for an administrative authority means two to three years after agreements are signed. In markets that evolve in months, this is a different temporality.

The real question is not whether Microsoft and Amazon acted wrongly. It is whether available tools allow distinguishing legitimate investment from an arrangement that preempts future competition — and, if the answer is no, how to build them before the next emerging market.

Sources

  1. ProMarket / Stigler Center — “AI’s Tying Arrangements Jeopardize the Market” (June 22, 2026): https://www.promarket.org/2026/06/22/ais-tying-arrangements-jeopardize-the-market/
  2. Federal Trade Commission — report on investments in AI models, 2024 (FTC.gov, no verified direct link)
  3. Competition and Markets Authority (United Kingdom) — “AI Foundation Models: Initial Review”, 2024 (GOV.UK, no verified direct link)
  4. Annual financial reports Microsoft, Amazon, Google, Meta — fiscal years 2023 and 2024
  5. FTC - Launch of Inquiry on AI Partnerships (Jan. 2024)
  6. FTC - Staff Report on AI Partnerships (Jan. 2025)
  7. Amazon - Official Investment in Anthropic ($4B, March 2024)
  8. Anthropic - Amazon Partnership Expansion (April 2026)
  9. OpenAI - GPT-5 Trained on Azure
  10. CMA - Update Paper AI Foundation Models (April 2024)
  11. Wikipedia - Google DeepMind (acquisition in 2014)
  12. European Commission - DMA Gatekeepers Portal
  13. Fortune - Hyperscaler Capex 2024 (~$200B)
  14. CNBC - Microsoft Total Investment in OpenAI ($13B)