Compute Concentration

Compute concentration is the claim that the capacity to train and serve frontier artificial-intelligence models is held by a small enough number of firms to matter for competition, security and price. Established The claim is true at the layer people usually mean, and this brief’s argument is that it is truest at a layer they usually do not: not the models, but the physical substrate under them — advanced packaging, high-bandwidth memory, grid interconnection and firm power — and the financing structure wrapped around that substrate.

Frontier This is a synthesis brief, and it exists because three neighbouring briefs each stop at the same edge. Digital Economies tests whether digital markets inevitably tip to one winner and finds, repeatedly, that they do not when users can multihome — and then states as an external requirement the one condition it cannot settle: an inference cost structure that permits more than a few model providers. Future Capital Markets shows that the instruments with real scale are old ones relocated out of disclosure regimes, and that index concentration puts an extraordinary share of household savings behind ten companies. AI Governance shows why the leading statute uses a quantity of arithmetic as its regulatory trigger: compute is countable and capability is not. None of the three owns the joint question. This brief does: does the cost structure of frontier compute produce durable concentration, at which layer does it bind, and what remedy attaches to that layer?

Frontier The answer this brief defends has three parts. First, model-layer concentration is real but is a lease rather than a moat: the gap between the best closed model and the best open-weight model has been measured in months, not years, and it has not widened. Second, substrate-layer concentration is the durable kind, because it rests on capacity that takes three to five years and a great deal of capital to duplicate, and on grid connections that are rationed by queue rather than by price. Third, the financing structure is doing analytical work that nobody has audited: a large share of 2024–2026 demand was created by suppliers, investors and customers who are partly the same parties, an arrangement with a specific and unhappy precedent in telecommunications vendor financing.

Established What this brief does not do. It does not re-argue platform tipping, market definition or self-preferencing, which Digital Economies owns with better evidence; it does not restate the private-credit and index-concentration numbers that Future Capital Markets owns; and it does not re-litigate whether compute thresholds are a good regulatory trigger, which AI Governance owns. It uses all three and adds the layer beneath them.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established Training compute for leading models has grown at roughly four to five times per year for more than a decade. The measurement comes from a small number of groups maintaining databases of notable models, principally Epoch AI, and it is an estimate: most figures are reconstructed from chip counts, training durations and utilisation assumptions rather than disclosed. The largest publicly discussed runs passed 10²⁵ floating-point operations around 2023 — the number the EU statute later adopted as a trigger — and the largest by 2025 are estimated in the 10²⁶ range. Nobody outside the labs can verify these numbers, which is itself a finding.

Frontier Cost per frontier run has grown more slowly than compute, and still fast. Hardware improvement absorbs part of the compute growth, so amortised cost per leading run has been estimated to roughly double to triple each year rather than quadruple. That trajectory takes the largest runs from tens of millions of dollars in 2022 to the high hundreds of millions by the mid-2020s, with a run costing US$1 billion in compute alone treated as a near-term rather than a distant milestone. These are estimates with wide error bars, and the labs have not contradicted them with disclosure.

Frontier The capital expenditure is not an estimate, and it has outrun disclosed artificial-intelligence revenue by roughly an order of magnitude. The largest US cloud and platform firms raised combined capital spending from roughly US$150 billion in 2023 to well above US$300 billion in 2025, with 2026 guidance higher again. Disclosed revenue directly attributable to generative artificial intelligence across the major model developers and clouds has been reported in the tens of billions of dollars annually over the same period. The gap is not by itself evidence of a bubble — infrastructure is built ahead of demand by design, and much of the spending serves non-AI cloud workloads — but it is the single largest unexplained quantity in the sector, and the parties reporting both numbers use different boundaries for each.

Established Cloud infrastructure is a three-firm market by revenue share, and has been for a decade. The tracking firms that publish quarterly shares, principally Synergy Research Group and Canalys, put the three largest providers at roughly two-thirds of global cloud infrastructure services spending, with the leader near 30% and the share structure stable through the period when total spending tripled. A stable share structure under rapid growth is the signature of a market where entry is possible but scale economies are real.

Established The accelerator layer is more concentrated than the cloud layer. One vendor has held the large majority of data-centre accelerator revenue through the period, and the parts that matter are not fungible: the accelerator, the memory stacked beside it, and the packaging that joins them. Advanced packaging capacity at a single foundry has been the reported binding constraint on shipments, with the supply-demand gap tracked publicly and reported narrowing from about 20% to about 10% during 2026. High-bandwidth memory comes from three suppliers, and the current generation standard was published by the industry body in 2025.

Established Power, not silicon, is now the gating input in the largest markets. Interconnection queues in the United States hold far more generation and storage capacity than the grid can process, with the backlog documented annually by Lawrence Berkeley National Laboratory; one large state paused new data-centre interconnections in the face of hundreds of gigawatts of requests; capacity prices in the largest US wholesale market were attributed substantially to data-centre load by its own market monitor. Local opposition has become quantifiable: a tracking project put US$64 billion of data-centre projects blocked or delayed. The build time for a gigawatt-scale site, measured rather than promised, is the constraint that converts capital into capability.

Established Electricity demand from this build-out is large, growing and smaller than the headline claims. The International Energy Agency’s assessment and the Lawrence Berkeley laboratory’s US inventory both place data centres in the low single-digit percentages of electricity consumption with a steep growth rate attributable to AI workloads. The Institute’s Ultra-Efficient Computing Energy Systems brief carries the efficiency argument in detail; what matters here is that the marginal megawatt is contested locally and priced through queues, which advantages incumbents who applied first.

Frontier Export controls created a natural experiment, and its first result cut against the strong concentration thesis. Successive US controls from October 2022 restricted the most capable accelerators to China, were tightened in 2023, and were reorganised again in 2025. In December 2024 and January 2025 a Chinese laboratory published a frontier-class model with a reported final training run of a few million dollars on a restricted-tier accelerator fleet. The reported figure covers the final run only — not prior experiments, not the fleet’s capital cost, not staff — and the debate over the true total has not been settled by disclosure. What the episode establishes is narrower and still important: a compute ceiling roughly an order of magnitude below the frontier did not prevent a near-frontier result within about a year.

Frontier The open-weight lag is the best available measure of how durable model-layer concentration is. Comparisons of open-weight releases against the best closed models on public benchmarks put the lag in the range of six to twelve months over 2023–2025, without a clear widening trend. Benchmarks are a weak instrument — the critique AI Governance develops at length applies here — and the comparison ignores reliability, safety tooling and serving cost. Read conservatively, it says the model layer behaves like a fast-following industry rather than a natural monopoly.

Frontier The partnership structures are the part competition authorities actually examined. The US Federal Trade Commission’s 6(b) staff report on artificial-intelligence partnerships and investments, published in January 2025, described arrangements in which compute credits functioned as investment consideration, exclusivity operated through cloud commitments rather than through equity, and information flowed between partner and portfolio company. The UK competition authority reviewed several of the same relationships and concluded in most cases that they did not amount to relevant merger situations, while its separate cloud services market investigation concluded in 2025 that the two largest providers each hold significant unilateral market power. Two authorities, similar facts, different instruments: the merger test found little, the market-investigation test found something.

Frontier Circular financing is the 2025–2026 development with the weakest public accounting. Through 2025 the sector announced accelerator vendors investing in model developers who commit to buying accelerators, cloud providers taking equity in customers who commit to multi-year capacity purchases, and headline contract values in the tens to hundreds of billions of dollars whose revenue recognition, cancellation terms and collateral are not public. The structure is not novel: telecommunications equipment vendors financed their own customers in the late 1990s, and the resulting receivables were written down heavily when the customers could not pay. Future Capital Markets shows the general pattern — instruments with real scale relocating out of disclosure regimes — and this is that pattern applied to compute.

Frontier Sovereign compute is a real programme line and a small share of the total. The European Union’s artificial-intelligence gigafactory and factory programmes, the United Kingdom’s research resource, Japanese and Indian procurement subsidies, and Gulf-state national champions all committed public capital between 2024 and 2026, in some cases at announced scales of tens of billions of euros. Two features recur: the compute is largely bought from the same two or three suppliers, and the announced figures mobilise private capital rather than appropriate public capital. Sovereign compute so far diversifies who owns the cluster, not who makes it.

Established Switching costs are real at the training layer and weaker at the serving layer. Training a frontier model on an unfamiliar accelerator costs months of engineering, which is the strongest lock-in in the stack. Inference is more portable: compiler and serving stacks now target several hardware families, and at least one large model developer has trained and served at scale on a cloud provider’s own silicon. Cloud switching costs were attacked directly by regulation — data egress charges for customers leaving were removed by the largest providers, and European rules require switching charges to be eliminated — and the early evidence is that egress fees were a symptom rather than the cause of stickiness.

Frontier Does scaling continue to pay at the margin? The concentration thesis assumes that the next order of magnitude of compute buys enough capability to be worth its cost. If returns to scale in pre-training flatten while post-training and inference-time compute deliver more per dollar, the advantage shifts from whoever can afford the largest run to whoever serves most efficiently, which is a different and less concentrated market.

Frontier Is the depreciation schedule right? Accelerator fleets are depreciated over roughly five to six years in the major filings, and that assumption determines reported profitability across the sector. If useful economic life is three years because newer parts are so much cheaper per unit of work, reported earnings are overstated and the capital cost of a frontier run is higher than the estimates in section 2. This is an accounting question with a factual answer that will arrive on its own schedule.

Frontier Which layer should a remedy attach to? Interoperability and open weights address the model layer, which this brief argues is the least durable. Remedies that would bite at the substrate layer — packaging capacity, memory supply, interconnection priority — sit with industrial policy and energy regulators who do not think of themselves as competition authorities.

Speculative Does compute concentration convert into political concentration? The governance literature treats compute as a governance handle precisely because it is countable and held by few parties. The same property makes the holders unusually useful to states and unusually exposed to them. Whether that produces regulation, capture or informal alliance is not settled by any measurement available here.

Established Advanced packaging is the narrowest physical point. Joining logic to stacked memory at current densities is done at commercial volume by a very small number of lines, and capacity additions run on multi-year construction schedules. Whoever holds allocation holds the queue position for everyone downstream, which is why allocation terms, not list prices, are the commercially interesting variable.

Established High-bandwidth memory is a three-supplier market with qualification friction. Qualification of a memory stack into an accelerator programme takes many months and is specific to the pairing, so supply is less fungible than the three-supplier count suggests. The published generational standard sets the interface; it does not make the parts interchangeable in a shipping product.

Established Grid interconnection is rationed by queue, not price. The published queue statistics show multi-year waits and low completion rates, and the effect on market structure is direct: a firm that filed three years ago holds an option a new entrant cannot buy at any price. Behind-the-meter generation is the workaround, and it converts a computing decision into a permitting decision.

Frontier Skilled labour for the build is scarcer than capital. Electricians, high-voltage technicians and commissioning engineers gate the schedule of large sites in every market where these projects cluster, and the training pipeline responds on a multi-year lag. This is the least discussed bottleneck and the one least amenable to money.

Established Leading-edge foundry capacity is the upstream dependency for everything here. One firm produces the great majority of leading-edge logic, and public subsidy programmes on three continents are attempting to broaden that base; the largest US awards were finalised in January 2025. Fabs built now produce parts late in this decade, so nothing in the current cycle depends on them.

Established Critical minerals and transformer supply sit under the power dependency. The grid equipment needed to connect large loads — large transformers, switchgear, turbines — is on multi-year lead times, and the materials outlooks that track the inputs show tight markets. The compute build competes for these with electrification generally.

Frontier Measurement of compute itself is a dependency this brief inherits unresolved. Training compute figures are reconstructions, not disclosures. Every regulatory threshold, every concentration statistic and every cost estimate in the public debate rests on a small number of estimating groups, and no auditable reporting regime exists. AI Governance documents the consequence for statutory triggers; the consequence here is that the competition analysis is being done on estimated data.

Frontier The decisive result is the one already running: whether models trained under the export-control compute ceiling stay within about a year of the frontier through 2026 and 2027. This is the cleanest natural experiment available on the central question, because it separates access to compute from everything else that varies between laboratories. If capped laboratories keep converging, then compute concentration does not convert into capability concentration, the model layer is a fast-following industry, and remedies belong at the substrate. If the gap widens through two successive model generations, the strong concentration thesis is confirmed on evidence rather than on inference from cost curves. The measurement needed is unglamorous: comparable, independently run evaluations of capped and uncapped models released in the same window, reported with dates and with training-compute estimates attached.

Frontier The second test is the depreciation cycle, which will report itself. The 2024–2026 accelerator fleets reach the end of their assumed useful lives on a schedule nobody controls. Whether they are still earning at that point, and at what utilisation, settles the profitability question that the capital-expenditure-versus-revenue gap only poses. No new instrument is needed, only consistent segment disclosure, which is exactly what has not been provided.

Speculative The third is a switching trial run by a large buyer and published. A public-sector purchaser of significant scale could port a serving workload between two accelerator families and two clouds and publish engineering cost, performance delta and elapsed time. Switching cost is asserted by every party to the debate and measured by none, and a single credible measurement would discipline both the lock-in and the contestability arguments.

Frontier The fourth is an audited accounting of the circular transactions. A regulator or standard-setter could require that vendor investments tied to purchase commitments be presented as such, with revenue recognition, cancellation rights and collateral disclosed. The precedent from telecommunications vendor financing is that this disclosure arrives after the write-downs rather than before.

Established A frontier training cluster is a power and network problem wearing a computing costume. Current-generation rack designs draw on the order of 100 kW and above per rack, which forces liquid cooling, new power distribution and a building designed around the electrical plant rather than the floor space. Retrofitting an existing facility is usually more expensive than building a new one, which favours large greenfield sites and therefore large balance sheets.

Established Interconnect is the second scaling wall. Training at scale requires every accelerator to exchange gradients with every other at each step, so the network is part of the computer. This is why co-packaged optics and proprietary high-bandwidth fabrics have moved to the centre of accelerator vendors’ roadmaps, and it is a further source of integration between the parts of the stack that competition analysis prefers to treat separately.

Frontier Geographic distribution of a single run is the engineering question with the largest structural consequence. If a frontier run can be split across sites hundreds of kilometres apart with acceptable efficiency loss, the power constraint loosens and the minimum viable site shrinks — which lowers the entry barrier. Multi-site training has been reported at increasing scale; the efficiency penalty at frontier scale has not been published in a form outsiders can evaluate.

Established Platform competition doctrine is the adjacent field, and it does not fit this case. Digital Economies shows that the tipping story fails where users multihome and that standard market-definition tests may overstate network-good power. Compute is different in the way that matters: it has a real and substantial marginal cost, so the relevant analogy is a capital-intensive utility industry rather than a zero-marginal-cost software market, and the doctrinal toolkit built for the latter transfers badly.

Frontier Capital-market structure decides who can finance a frontier run. Future Capital Markets documents the migration of scale finance into private credit and out of disclosure regimes, and the concentration of index weight in a handful of firms. Both bear directly here: the marginal dollar funding a data centre increasingly arrives through private vehicles, and household exposure to the outcome runs through index funds rather than through choices anyone made about artificial intelligence.

Frontier Governance uses compute as a proxy and inherits its measurement problem. AI Governance shows the leading statute triggering on a quantity of arithmetic because capability is not countable. That choice makes the compute-estimation weakness described in section 5 a regulatory weakness, not merely an analytical one.

Frontier Capability forecasting sets the value of the whole argument. Artificial General Intelligence carries the scaling-returns evidence. If returns flatten, the concentration question shrinks to an ordinary industrial-organisation problem about utilities; if they do not, the same question becomes a question about who holds a strategic capability.

Established The instruments in use were not designed for this market. Merger control asks whether a transaction creates a relevant merger situation, and the partnership structures were built in part to avoid that question. Market-investigation powers of the kind used in the United Kingdom cloud inquiry reach conduct and structure without a transaction, and they are rare: most jurisdictions do not have them.

Frontier Compute reporting is the missing institutional capability. Regulators currently learn about training runs from developer self-reports and from academic estimators. A reporting regime with audit rights — chip counts, cluster locations, run durations — is technically straightforward, and it is the precondition for any enforceable threshold. Proposals exist in the technical governance literature; no jurisdiction had implemented an auditable version by early 2026.

Frontier Industrial policy and competition policy are pulling in opposite directions. Subsidy programmes for fabs, packaging and sovereign clusters are justified on resilience grounds and are mostly awarded to the largest incumbents, because they are the parties who can build. The resulting structure is more national and no less concentrated, and no major jurisdiction has published a framework reconciling the two objectives.

Established Energy regulators have become de facto AI regulators. Interconnection priority, tariff design for large loads and cost allocation between data centres and other ratepayers now determine which projects proceed. These decisions are made by state and provincial commissions under statutes written for a different industry, with the competition consequences nowhere in their remit.

Frontier Cost allocation is the near-term distributional question. When a large load arrives on a constrained grid, someone pays for the reinforcement. Whether that cost falls on the load through special tariffs or on the general ratepayer base is being decided case by case, and the documented rise in wholesale capacity prices attributed to data-centre demand shows the amounts are not trivial.

Established Access to compute is a research-equity question with measurable consequences. Academic groups have been priced out of frontier-scale training, which shifts the composition of published work towards evaluation and analysis and away from independent replication of frontier results. The national research-resource programmes exist for this reason and are, at current scale, one to two orders of magnitude short of a frontier run.

Speculative Open weights trade concentration risk against misuse risk, and the trade is not empirically resolved. Releasing weights reduces dependence on a few providers and removes the release decision from the developer. Both effects are real; the quantitative balance depends on the marginal uplift an open model gives a bad actor over what is already available, which has not been measured in a way that survives scrutiny.

Frontier A concentrated substrate is a concentrated point of failure. If advanced packaging, high-bandwidth memory and leading-edge logic remain geographically concentrated, a single regional disruption removes a large fraction of world capacity for the years it takes to rebuild. This is the same argument made for semiconductor resilience generally, applied to the specific parts that gate AI capacity.

Speculative Compute may become a sovereign capability class. States that treat compute as strategic will build or buy national capacity and attach conditions to its use, which turns a commercial market into an allocated resource with quotas and priority access. The early programme announcements are consistent with this and do not yet demonstrate it.

Frontier The most likely long-run structure is a regulated-utility shape without the regulation. High fixed costs, real marginal costs, essential-input status and a handful of suppliers is the classic description of an industry that ends up either regulated or nationalised. The current arrangement has the economics of a utility, the disclosure of a private partnership and the oversight of neither.

These horizons track market structure and the physical substrate, not model capability, which the corpus carries elsewhere.

• 10 yr: Frontier The depreciation cycle reports, and the capital-expenditure-versus-revenue gap resolves one way or the other. Packaging and memory capacity additions land, easing the narrowest physical constraint while power remains rationed by queue. Expect at least one jurisdiction to impose auditable compute reporting, and at least one significant write-down of vendor-financed commitments.
• 25 yr: Speculative Either the substrate broadens — multiple packaging sources, several accelerator families with portable software, distributed training across sites — and the market resolves into an ordinary capital-intensive industry with three to six credible suppliers; or it does not, and compute is administered through allocation regimes resembling spectrum or export licensing.
• 50 yr: Speculative If compute remains an essential input to economic activity at current relative cost, some form of common-carrier or utility regulation is the historically normal endpoint. The alternative is that efficiency gains make frontier capability cheap enough that the question dissolves, which is the outcome the efficiency literature treats as plausible and unproven.
• 100 / 250+ yr: Handwave Projections of planetary compute budgets, or of compute as the organising unit of economic accounting, extrapolate a fifteen-year cost curve past every institution that currently exists. The physics of computation permits enormous headroom; nothing in the economic record supports a schedule.

• Depends on This brief depends on results other briefs on the map already carry. Digital Economies supplies the measured finding that digital markets do not reliably tip when users can multihome, which is why this brief locates durable concentration in the substrate rather than in the models. Future Capital Markets supplies the migration of scale finance out of disclosure regimes, without which the circular-financing argument in section 2 would be speculation. AI Governance supplies the compute-threshold and evaluation-access record. Artificial General Intelligence supplies the scaling-returns evidence that sets how much any of this matters, and Ultra-Efficient Computing Energy Systems supplies the efficiency trajectory that determines whether the power constraint tightens or loosens.
• Requires (not on this map) Packaging capacity at commercial volume outside the single foundry that currently holds it, since that line is the narrowest physical point in the chain. Memory stacks qualified across more than three suppliers and across accelerator programmes, because the qualification pairing makes a three-supplier market behave like a narrower one. Interconnection and firm power delivered on a two-year rather than a five-to-seven-year horizon, which is an institutional and construction problem rather than a generation-technology problem. An auditable compute reporting regime, without which both the competition analysis and the statutory thresholds rest on academic reconstruction. An accounting treatment that presents vendor investments tied to purchase commitments as what they are, before rather than after any write-down. And a depreciation schedule validated against realised revenue, since the sector’s reported profitability and the true cost of a frontier run both hang on that assumption.
• Enables If these hold, the enabled outcomes are unremarkable and valuable: contestable inference markets with portable workloads, research access to frontier-scale training outside the largest firms, and a competition analysis conducted on reported rather than estimated data. If they do not hold, the enabled outcome is an administered market in an essential input.
• Adjacent Digital Economies for platform doctrine and market definition; Future Capital Markets for how the build is financed and who bears the exposure; AI Governance for compute as a regulatory trigger; Ultra-Efficient Computing Energy Systems for the energy denominator.

Handwave “Three companies control artificial intelligence.” Three companies hold roughly two-thirds of cloud infrastructure revenue and a comparable concentration exists in accelerators, and neither fact supports the stronger claim. Capable models have been trained outside those firms, including under an export-control ceiling, and open-weight releases track the frontier by months. The defensible statement is about the substrate: control of packaging, memory and grid connections is more concentrated and harder to duplicate than control of models.

Handwave “A frontier model costs a hundred million dollars, so only five firms can build one.” The cost figure is an estimate of the final run, and the inference skips the part that matters: near-frontier capability has repeatedly been reached at an order of magnitude less. The binding requirement is not the cheque but the fleet, the engineering team and the data pipeline, and those are concentrated for reasons that include but are not limited to capital.

Frontier “Open weights solve concentration.” Open weights reduce dependence at the model layer, which this brief argues is the layer least likely to stay concentrated anyway. They do nothing about packaging allocation, memory qualification or interconnection queues, and serving a large open model at scale still requires the same substrate. Open weights are worth having for other reasons; as a competition remedy they aim at the wrong layer.

Frontier “Export controls failed.” They did not achieve the strong objective of preventing near-frontier models under the ceiling, and they did change the cost, timing and hardware mix of that work. Judging them against the strong objective alone is the same error as judging them against no objective; the honest reading is a partial, expensive and still-running experiment whose result is the most informative evidence in this brief.

Speculative “Sovereign compute delivers independence.” A national cluster built from the same accelerators, the same packaging line and the same memory suppliers relocates ownership without relocating dependence. Independence would require domestic capability at the substrate layer, which the announced programmes do not fund and could not deliver within their stated timeframes.

Frontier “The capital expenditure proves a bubble.” It proves a gap between spending and disclosed revenue, which is normal when infrastructure is built ahead of demand and abnormal at this ratio. The diagnostic is not the gap but the financing: demand funded by the suppliers of the thing being demanded is the pattern that turned a telecommunications build-out into write-downs, and it is the pattern currently least visible in public accounts.