Concentration of Power in AI: Who Owns the Stack
4 min read · updated August 3, 2026
“AI is concentrated” is too coarse to act on. Concentration is a property of a market, the stack contains at least eight of them, and they have almost nothing in common economically. This page maps them by structure rather than by share, because shares change and structure does not.
The stack as a set of markets
| Layer | Description |
|---|---|
| fabrication tooling | The equipment that makes advanced chips, above all lithography. The most concentrated layer in the entire economy by most measures, on the basis of decades of accumulated process knowledge. |
| foundries | Leading-edge manufacturing. Capital costs per fab at the frontier are extraordinary and yield learning compounds, so the number of firms able to operate at the newest node has fallen with every generation. |
| accelerators | The chips themselves, plus the software ecosystem around them. Concentration here is as much about developer tooling and accumulated kernels as about silicon. |
| memory and interconnect | High-bandwidth memory and networking. Frequently the binding supply constraint on accelerator output, and less discussed than the chips it feeds. |
| cloud capacity | Buildings, power, and the balance sheet to pre-commit to hardware. Capital intensity and, increasingly, access to interconnection are the barriers. |
| foundation models | Training runs at the frontier. High fixed cost, near-zero marginal cost of copying, and an unusually fast rate of obsolescence. |
| applications | The competitive layer, with low entry costs and correspondingly thin defensibility. |
| distribution | Operating systems, browsers, app stores, productivity suites. Where an installed base can convert into default usage without any advantage at the model layer. |
What concentrates a layer, and what disperses it
Concentrating forces. Extreme fixed costs with low marginal costs, which favour whoever can spread them over the most output. Learning curves, where yield and quality improve with cumulative production, so a lead compounds rather than decays. Scarce complementary inputs — advanced packaging, memory, power connections, a small pool of people who have trained at scale. Ecosystem lock-in, where the software written against one platform is the barrier rather than the hardware. And capital access, since the ability to pre-commit billions is itself a scarce resource that only some balance sheets have.
Dispersing forces, which get less attention. Open weights, which put a capable substitute in circulation permanently and cap what a closed provider can charge for comparable capability. Rapid obsolescence, which repeatedly resets any lead at the model layer — a frontier position is a lease, not a freehold. Falling cost per unit of capability, which lowers entry over time. Standardised API shapes, which reduce switching costs to a configuration change and make routing between providers practical. And specialised silicon entrants, which attack the accelerator layer at specific workloads rather than head-on.
The layers therefore behave very differently. Tooling and foundry concentration is grounded in decades of accumulated capability and is extremely durable. Model-layer concentration is grounded in capital and talent, both of which move, and it has repeatedly proven less durable than commentary at the time assumed. Treating these as one phenomenon is the analytical error to avoid.
There is a further asymmetry between the layers that is easy to miss. Concentration at the bottom of the stack is a supply constraint: it determines who can build at all, and it is enforced by physics and capital. Concentration at the top is a demand constraint: it determines which product a user encounters by default, and it is enforced by habit and by where the software already is. The two call for completely different remedies — the first is addressed, if at all, by industrial policy and diversification of supply, the second by interoperability, defaults and switching costs. A policy debate that treats “concentration in AI” as one problem tends to apply the wrong instrument to whichever end it was not thinking about.
Integration without a merger
The structural feature most worth watching is that the relationships binding these layers together are frequently not acquisitions, and therefore frequently do not trigger the review process built for acquisitions:
- Compute-for-equity. An investment paid substantially in credits for the investor’s own capacity. It funds the recipient, guarantees the investor’s utilisation, and creates a dependency, all without a change of control.
- Exclusive or preferential capacity. Long-term supply arrangements that determine who can train at scale, negotiated privately.
- Allocation of scarce hardware. When supply is short, the supplier chooses the customers, which is a form of market power exercised through a queue rather than a price.
- Distribution bundling. Placing a model as a default inside software with an existing installed base, where the competitive advantage comes from the distribution rather than from the model.
- Acqui-hire and licensing. Absorbing a team and licensing its technology while leaving a shell entity in place, which transfers the capability without the transaction that would be reviewed.
Whether these arrangements are anticompetitive or are ordinary contracting under supply scarcity is genuinely disputed among competition economists. What is not disputed is that they achieve coordination between layers, and that the standard review machinery is keyed to transactions rather than to arrangements.
What competition law can reach
Merger control needs a notifiable transaction meeting a threshold, usually based on turnover or deal value. Arrangements designed not to be one are not reviewed, though several authorities have asserted powers to call in transactions below thresholds.
Conduct rules — abuse of dominance, monopolisation — require establishing a dominant position in a defined market. Market definition is where these cases are actually won and lost, and defining a market in a stack that is being rebuilt annually is exceptionally difficult. Cases take years, and the layer in question can restructure before judgment.
Ex ante regulation of designated large platforms imposes obligations without proving an abuse first, trading precision for speed. Whether such regimes should extend to AI layers is an active question and the answer differs by jurisdiction.
The underrated levers are not competition law at all. Procurement can require multi-vendor capability and portability as a condition of public contracts. Interoperability and data portability requirements lower switching costs directly. Public compute provision changes the outside option for researchers and small entrants. And export controls, discussed elsewhere in this cluster, shape the hardware layer far more than any competition authority does.
Judging a concentration claim
Four questions before accepting any claim in this area. Which layer — they are not the same market. What unit: revenue, capacity, tokens served, installed base and research citations give different pictures and are chosen to. Is the barrier durable capability or a temporary supply shortage, which is the distinction between a structural problem and a cyclical one. And what is the counterfactual: concentration relative to what alternative arrangement that was actually available.
No current shares appear on this page deliberately. They change faster than a written page and quoting one would be presenting a stale snapshot as structure. For current figures use regulatory filings, competition authority market studies and financial disclosures, which are the primary sources and are dated.