AI and Market Concentration — FTC, CMA and JFTC Raise Three Competition Risks── The foundation model market is concentrating around compute, data and talent, and competition authorities in major economies are beginning to intervene in dynamics that entrench first movers

01Three companies capture 90% of enterprise AI spending

According to Menlo Ventures' mid-2025 market update, OpenAI, Anthropic and Google together accounted for roughly 90% of enterprise spending on foundation models. In 2023, OpenAI alone held 50% of the market. By the end of 2025, its share had fallen to 27%, while Anthropic rose to 40%. The names at the top shifted, but the combined share of the top three barely moved. The market grew while concentration persisted.

Why does the foundation model market concentrate? Three inputs drive the answer: compute, data and talent.

02Exponentially rising training costs create barriers to entry

Epoch AI estimates that the compute required to train frontier models has grown at roughly 5× per year since 2020. Translated into cost, training expenses have risen at 2.4× per year. The Stanford 2025 AI Index Report calculated the training compute cost of GPT-4 at approximately $78 million and Google's Gemini Ultra at $192 million. In 2017, the original Transformer cost about $670 to train — a 280,000-fold increase in seven years.

This cost structure functions as a barrier to entry. Only a handful of organisations can commit over $100 million to a single training run, funded either by hyperscaler capital expenditure or multi-billion-dollar fundraising rounds. Anthropic's CEO suggested in 2024 that models costing over $1 billion would appear soon; if current trends hold, Epoch AI projects trillion-dollar training runs by the end of the decade. A 2024 RAND study examined the economic attributes of the foundation model market as of January 2024 and concluded that the conditions for natural monopoly are relatively strong.

03Concentration operates across three layers simultaneously

Training cost alone does not explain the full picture. Concentration is reinforced across three layers — compute, data and talent — each with its own dynamics.

The three layers reinforce one another. Organisations with compute can train on larger datasets. Their results attract talent, and that talent develops more efficient training methods. This cycle makes late entry progressively harder. The combined annual revenues of the three leading foundation model providers exceeded $20 billion by early 2025, growing over 9× between 2023 and 2024 — scale that funds the next round of compute acquisition and widens the gap further.

04The FTC, CMA and JFTC are asking the same structural question from different angles

Since 2024, competition authorities in major economies have published findings on foundation model markets in quick succession.

In January 2025, the US Federal Trade Commission published a staff report under its 6(b) investigative authority. The report examined investment and partnership arrangements between hyperscalers and AI developers, finding that partnership terms could distort access to compute and talent. It flagged the risk that hyperscalers gain access to sensitive technical and business information from their AI developer partners — information unavailable to competitors.

The UK's Competition and Markets Authority published an update paper in April 2024, identifying three risks to competition: firms controlling critical inputs (compute, data and expertise) may restrict access to shield themselves; foundation model developers with market power may dominate downstream markets; and independent developers may be denied key resources. The CMA is building a 200-person monitoring team through the Digital Regulation Cooperation Forum.

Japan's Fair Trade Commission released a discussion paper in October 2024, followed by a fact-finding report (ver. 1.0) in June 2025 and ver. 2.0 in April 2026. The JFTC divided the market into three layers — infrastructure, models and applications — and identified tying arrangements and access restrictions by cloud providers as potential violations of the Antimonopoly Act.

05In pharmaceuticals, platform dependence is the concentration risk

Pharmaceutical companies are buyers, not builders, of foundation models. They use cloud providers' infrastructure and AI vendors' services. The AI-in-drug-discovery market was estimated at roughly $4.5 billion in 2025, growing at 20–30% per year, but the compute infrastructure underpinning it is concentrated among a small number of cloud providers.

This dependence creates concrete risks. When a cloud provider bundles a foundation model with its compute environment, pharmaceutical companies can become locked into a single platform. Migrating a fine-tuned model requires re-preparing data and re-validating results — a process that can take months and consume significant engineering resources. Switching the model used for clinical trial analysis mid-programme affects regulatory submissions, because the analytical methodology described in the filing must match the tool actually used. A change of platform therefore triggers not just technical migration but regulatory re-documentation.

The tying concerns raised in the JFTC's ver. 2.0 report apply directly to pharmaceutical AI usage. Adopting an AI tool that runs only on one cloud environment weakens the buyer's negotiating position on compute pricing.

06Regulation can entrench the first movers it aims to constrain

Competition policy carries an unintended dynamic: the cost of regulatory compliance functions as a fixed cost, and fixed costs burden smaller players disproportionately.

The EU AI Act, which took effect in August 2024, requires conformity assessments for high-risk AI systems. Building the legal and technical infrastructure for these assessments demands upfront investment. Incumbents absorb that cost across existing revenue; for new entrants, it is an additional barrier. A similar dynamic appears in the Digital Markets Act's gatekeeper regime. In March 2025, two gatekeepers were fined a combined EUR 700 million — a figure representing only a small percentage of their annual revenue.

In the US, the FTC's scrutiny of investment relationships could have an asymmetric chilling effect: existing partnerships are treated as historical fact, while new partnerships face heightened scrutiny. Whether regulation targets past or future arrangements determines whether it constrains or protects first movers.

Korinek and Vipra argued in a 2025 paper that competition authorities face two tasks: ensuring market contestability by monitoring strategic behaviour, and setting quality standards for frontier models. The first pushes against concentration; the second may reinforce it. This tension sits at the heart of regulatory design.

07Open-weight models and vertical integration define the competitive axis

Open-weight models offer a partial counterforce to concentration. Meta's Llama holds roughly 9% of the enterprise foundation model market and makes its weights freely available. DeepSeek attracted attention in January 2025 by achieving strong performance with relatively modest compute.

But open-weight models have limits. Training still requires large-scale compute, and publishing weights concentrates the training cost burden on the developer. The developer bears the full cost of the training run while competitors benefit from the result at no charge — a structure that is sustainable only for organisations large enough to treat it as a strategic investment rather than a commercial product. Even when organisations fine-tune open models for commercial use, inference requires cloud compute, so dependence on concentrated infrastructure persists.

Meanwhile, the major cloud providers are integrating vertically — from compute provision through model development to application deployment. As this integration deepens, switching costs rise and downstream users lose alternatives. The JFTC's three-layer framework was designed precisely to analyse this vertical integration through the lens of antitrust law.

For pharmaceutical companies, the practical response involves three measures. First, distribute AI workloads across multiple cloud environments so that no single provider holds all the data and all the trained models. Second, secure data portability and migration terms at the contracting stage — before lock-in takes hold, not after. Third, build internal capability to evaluate and select models independently, rather than accepting whatever the platform vendor bundles. These steps do not eliminate dependence on concentrated infrastructure, but they preserve negotiating power and the ability to switch when alternatives emerge.