UK competition regulator opens AI foundation model probe

The UK’s Competition and Markets Authority (CMA) opened a formal market investigation into AI foundation models this week, citing concerns about concentration and barriers to entry.

The CMA investigation, announced by CEO Sarah Cardell, will examine:

  • Market concentration in AI foundation models (currently dominated by OpenAI, Google, Anthropic).
  • Barriers to entry for new AI foundation model providers.
  • Vertical integration between foundation model providers and downstream applications.
  • Compute access for new entrants.
  • Talent concentration.

The investigation is the first comprehensive UK competition review of the AI foundation model market. The CMA has the power to recommend remedies up to and including forced divestitures.

Why this matters:

The CMA is one of the most aggressive competition regulators in the world. Its investigations into digital markets (the Online Platforms and Mobile Ecosystems cases) have shaped global tech regulation.

The CMA’s interest in AI foundation models reflects a growing consensus that the AI market has the characteristics of a natural monopoly. Frontier model training requires:

  • Massive compute ($1B+ per training run).
  • Massive data (trillions of tokens).
  • Massive talent (dozens of top researchers).
  • Massive capital ($10B+ in funding).

These requirements create natural barriers to entry. New entrants face the cold-start problem — they need to spend $1B+ to build a competitive product, but they can’t generate revenue until they have a competitive product.

The CMA’s likely findings:

Based on the investigation’s scope, the CMA will likely find:

  1. Concentration. The top 3 providers (OpenAI, Google, Anthropic) account for an estimated 75%+ of AI foundation model usage. This is high concentration by any standard.

  2. Vertical integration. All three top providers offer downstream products (ChatGPT, Gemini, Claude) that compete with their customers’ products. This creates a fox-guarding-the-henhouse problem.

  3. Compute access. Compute is the critical input, and it’s dominated by Nvidia. New AI entrants depend on Nvidia chips, which limits their ability to compete with incumbents who have preferential Nvidia relationships.

  4. Talent. AI talent is concentrated in a handful of organisations. The top 1% of AI researchers work at fewer than 20 companies.

Possible remedies:

The CMA has wide-ranging powers. Possible remedies include:

  • Mandating interoperability between foundation models.
  • Requiring foundation model providers to license their models on FRAND terms.
  • Blocking vertical integration.
  • Forcing divestiture of downstream products.
  • Regulating compute access.

The most likely outcome is interoperability mandates. The CMA has historically favoured interoperability remedies over structural separation. Forced divestitures are reserved for the most egregious cases.

The global context:

The CMA investigation joins similar scrutiny in the US (FTC AI probe), EU (AI Act competition provisions), and France (competition authority AI review). The global regulatory consensus is that AI foundation models deserve competition scrutiny.

The AI labs are unlikely to welcome the investigation but won’t fight it directly. They’re more likely to argue that the market is dynamic and that today’s leaders are not guaranteed tomorrow’s. Whether the CMA accepts that argument is the question for the next 18 months.