The concentration of capability

Before asking whether a few companies control AI, ask what they control.

Training an advanced model, adapting an existing one, running it, reaching customers and setting the terms of a public service are different activities. A sector can have many applications built on a small number of models. Many models can depend on a few cloud services for their computing power, often called compute. A downloadable model can give a user more freedom while still requiring equipment or skills the user cannot afford.

This makes concentration a question about a particular resource or decision. Counting brand names is not enough. Nor is pointing to one successful newcomer. We need to know where an alternative exists, what it can do, and what it costs to use.

What exactly are we counting?

The analysis prepared for this brief examines a public dataset rather than relying on a headline about the number of important models. The calculation uses Epoch AI’s notable-model file, downloaded on 14 October 2026. It includes 544 records dated from 2020 through 2025. These are six complete calendar years; the incomplete 2026 year is excluded. The file contains selected notable releases, not every model made in those years. Its contents can also change as the curators add or correct historical records. Related model versions remain separate records; a unique name does not make each record an independent research event. Epoch model data

The question is narrow: how much of the recorded development credit belongs to the five largest contributors, and how does the answer change when we count joint work and organization names differently?

Imagine a model with four named developers. We could give each one a quarter of a model’s credit. Then each model contributes one unit in total. Or we could give each developer a full credit and divide by all developer credits in that year. The second method gives a jointly produced model more weight in the overall count. Neither method tells us who paid for the work, who controlled it or who did most of it. They answer different descriptive questions.

Names create another choice. Google, Google Brain, Google Research and DeepMind appear under different labels in historical records. Counting each label separately can make one connected group look like several independent sources of capability. But combining every firm into an ultimate owner requires a verified ownership history, which this analysis does not have.

Four versions were run. Two retain the listed names. Two combine a disclosed set of Google/DeepMind and Meta/Facebook name variants. Within each pair, one divides a model’s credit equally among its developers and the other counts all developer credits. The grouping is a sensitivity check, not a complete map of corporate control. It combines some historically distinct teams and leaves other related organizations separate.

Here are the results. Each entry is the share of development credit assigned to the five largest contributors under that method. The five can differ across years and methods.

The broad claim that concentration rose between 2020 and 2025 does not survive every counting choice. Three versions rise. The grouped-name, shared-credit version falls from 58.0% to 55.1%. All four versions rise between 2023 and 2025, within the same selected dataset. The full six-year sequence matters: it shows a dip and rebound, not a steady march in one direction.

The recent rise is a descriptive finding worth noticing. It is not proof that a specific firm can raise prices, block entry or dictate a government’s choices. Those claims require other evidence. The dataset does not measure sales, customer dependence, actual usage, compute ownership or the importance of one release compared with another. A model that transforms a field and a model that makes a modest improvement each receive one model unit in the shared-credit calculation.

There is another important limit. The difference among these four lines is not a confidence interval. It shows sensitivity to a few explicit choices. It does not cover every possible error or selection problem. This is not a random sample of all AI development, and a statistical margin of error would not repair that fact.

Download the inputs, code and calculation record. The package includes a factual projection of the selected records, an offline verification route and instructions for checking the original input hash. It does not reproduce source-paper text.

The useful result is partly about measurement. A reader should ask for the counting rule before accepting a trend. If grouping related names changes the conclusion, that fact belongs beside the headline.

The frontier and the price of a useful task

It is possible for entry into the most expensive kind of model development to become harder while useful access becomes cheaper.

Ben Cottier and colleagues estimated training costs for 41 models among those using the most computing power when released between 2015 and 2023. Their estimates rose sharply over time. But they measured the changing set of highly compute-intensive models, using uncertain public information about equipment and costs. They did not measure the cost of obtaining a fixed useful ability. Buying a computing cluster is also a different expense from assigning part of its cost to one training run. The rising costs of training frontier AI models

A September 2026 Epoch analysis approached the other side of the question: the cost of achieving a fixed score on selected benchmarks, or set tests of a model’s ability. It found large declines since 2023, including the extra processing used to work toward an answer. The authors describe the estimate as rough and identify limits in sampling, benchmark choice and the link to useful work. This is evidence against a simple claim that all AI capability is becoming unaffordable. It does not mean a customer’s bill falls at the same rate. The plunging price of thought

These findings can both be true. A more capable frontier can demand greater investment even as yesterday’s capability becomes cheaper. For a small research group, access to a good existing model may matter more than the ability to train the most advanced one. For a country concerned about dependence, the ability to operate, inspect and replace an adequate model may matter more than winning a global ranking.

The policy goal should match the need. If the problem is that researchers cannot run experiments, shared compute and support may help. If the problem is dependence on a supplier’s terms, a larger research grant may leave that dependence intact. If the problem is an essential service with no reliable alternative, the buyer needs a continuity plan and tested ways to move.

Cheaper access is valuable. It should not be dismissed merely because large firms remain important. It should also not be mistaken for a permanent right. A service can be affordable while its users lack the records, skills or contract terms needed to switch.

Dependence can sit below the model

A 2025 US Federal Trade Commission staff report examined three partnerships between cloud providers and AI developers. It identified several possible sources of dependence: funding linked to cloud spending, preferential access, technical integration and costs of changing infrastructure. It also recognized that partnerships provide capital, compute and distribution. The report was not a finding of illegality, and its dated evidence does not establish the current terms of every deal. Its value here is to make the possible mechanisms concrete. FTC partnership study

The UK Competition and Markets Authority’s 2025 cloud investigation also shows why product definitions matter. Its much-cited concentration figures for standard cloud infrastructure are not shares of the separate market for accelerated computing, which uses specialized chips for demanding tasks such as AI. The report distinguishes these services and examines switching, entry and other evidence alongside shares. We should not turn a regional standard-cloud estimate into a global AI-compute statistic. CMA final report

What would practical independence require? A buyer could test whether its records can be exported in a usable form. It could check whether a second provider can run the same workflow at an acceptable level of quality. It could estimate the work needed to change prompts, interfaces, security checks and staff routines. These are proposed procurement tests, not claims that moving is always easy or always the best choice.

There are costs on both sides. Supporting two systems takes time. A portable design may give up a feature that works well on one platform. Small organizations may benefit from an integrated service precisely because they cannot build every part themselves. The aim should be a proportionate ability to change course, especially where a failure or price change would disrupt an essential function.

Open releases do not settle the question

Research on the open-model ecosystem also complicates a simple story. Shayne Longpre and colleagues studied a large collection of Hugging Face model downloads. They found a period of wider participation followed by rising concentration in 2025. Their analysis also showed the importance of groups that adapt and repackage models. Downloads are an imperfect proxy for use, and assigning credit to a base-model developer or an intermediary changes what the measure means. The work does not cover the whole closed-model economy. Economies of Open Intelligence

The lesson is not that openness fails. It is that openness changes some powers without necessarily changing all of them. A person who can obtain model weights may gain the ability to run or modify the model. They may still depend on a few hardware suppliers, a dominant distribution channel, or a specialist able to make the system work.

We should therefore ask which layer a proposed reform reaches. A public model can widen access. Shared equipment can make that access usable. Training can help new groups take part. Clear rights and tested alternatives can reduce dependence after adoption. None of those steps should be credited with benefits that only another step can provide.

Who controls AI is not answered by a single percentage. The answer is a map of resources, relationships and options. The purpose of measuring concentration is to find the places where people lack a workable alternative, then investigate why. It is the beginning of that inquiry, not its verdict.