Where training-grade compute is made, held, and fought over — the third axis of bloc-level competition, as of August 2026

A companion essay in this volume argues that there is not one global defence-technology market but three, structured differently and diverging. Compute is the axis along which that divergence is now most visible, most measurable, and most contested. The ability to train a frontier artificial-intelligence model is not evenly distributed across the world; it is concentrated in a small number of physical chokepoints, governed by a shifting and internally contradictory set of export controls, and pursued by a widening circle of states that have concluded compute is strategic infrastructure rather than commercial plumbing. This Signal maps where training-grade compute is made, where it is being built, and where the contest over it currently stands.

A caution before the map, because it governs how this piece should be read. Compute writing dates faster than almost any other subject in this field: export rules change by the month, valuations by the week, and buildout figures are obsolete on arrival. Every figure below is stated as of August 2026 and should be treated as a snapshot, not a fixed value. Where we expect the picture to move materially before this volume closes, we say so. The map is accurate to the day it was drawn; the territory is moving underneath it.

The chokepoints: where compute is physically made

The first thing to understand about the compute map is that its most important features are not countries but companies, and that the concentration is more extreme than the headlines about Taiwan convey. The supply chain for training-grade chips narrows, as it ascends toward the leading edge, to a series of single points of failure.

At the apex sits one company that most readers have never had reason to name. ASML, based in the Netherlands, is the sole manufacturer on earth of the extreme-ultraviolet lithography machines required to print transistors below roughly seven nanometres — the machines without which no advanced AI chip can be made. There is no second supplier and none close; the company holds effectively one hundred per cent of the EUV market. Its scanners are objects of almost absurd complexity — hundreds of thousands of components, holding optical alignment to picometre tolerances while firing tens of thousands of tin droplets a second into a laser — and its annual production rate of roughly fifty to sixty standard EUV systems is, in a real sense, the physical ceiling on how much leading-edge fabrication capacity the world can add in any given year. A foundry can decide to spend twenty-five billion dollars on a new fab; it cannot install more EUV scanners than ASML can build.

One tier down sits the manufacturing chokepoint. Taiwan Semiconductor Manufacturing Company controls on the order of ninety per cent of advanced logic production below seven nanometres, and Taiwan-based capacity accounts for the overwhelming majority of the advanced packaging — the CoWoS process that stitches logic chips and high-bandwidth memory into a single AI module — on which current-generation accelerators depend. Below that again are chokepoints most discussions never reach: South Korea’s near-total dominance of high-bandwidth memory, and a set of Japanese material monopolies in EUV photoresists and mask blanks, plus single-supplier positions in EUV optics and drive lasers, none of which has a viable alternative at scale and all of which carry lead times measured in one to two years.

Epistemic status: Confirmed, as of August 2026. The concentration figures are drawn from company filings and supply-chain analyses current to mid-2026. The strategic reading — that these are single points of failure with no near-term substitutes — is well supported. The one moving element to watch is diversification: TSMC’s roadmap adds 3nm lines in Arizona and Japan across 2027–2028, which will spread the geography somewhat while leaving Taiwan the dominant hub for the foreseeable future.

The strategic implication of this geometry is the one the companion essay draws about boomerangs and chokepoints: the leading edge of compute rests on a supply chain whose critical nodes are monopolies, several of them concentrated on a single island within reach of Chinese power. Money does not solve this. The chokepoints represent capabilities accumulated over decades — ASML spent thirty years and more than ten billion euros turning EUV physics into engineering — and they cannot be reconstituted on the timescale of a political crisis. This is the hard floor beneath every compute-policy debate: the physical means of making training-grade chips is not, at present, something any single bloc controls end to end, and the bloc that comes closest is the Western one.

The controls: the contest over who may buy

If fabrication is where compute is made, export control is where the contest over who may possess it is being fought — and the striking feature of that contest, as of August 2026, is its incoherence. The United States has spent three years building, tightening, loosening, and re-tightening a regime of controls on advanced AI chips destined for China, and the regime now points in two directions at once.

The sequence is worth stating plainly because it is easy to lose. Controls imposed from 2022 restricted the most capable chips; Nvidia designed successive China-specific parts — the A800, then the H20 — to fall just within the permitted thresholds; the H20 became the chip on which a striking Chinese model was optimised in early 2025; the United States then reclassified the H20 as non-compliant in April 2025, reversed itself months later, and in December 2025 moved further still, permitting the far more capable H200 to be sold to China subject to inter-agency approval and a twenty-five per cent levy on the sales. A leading US think tank described the resulting framework, accurately, as strategically incoherent: tight enough to alarm, loose enough to fail, and depending entirely on how it is implemented.

The genuinely surprising development is what happened next, and it inverts the usual story. Having won the right to buy more capable American chips, Beijing declined to use it. Chinese authorities discouraged domestic firms from purchasing the approved chips on security grounds, and as of early 2026 Nvidia had reported essentially no revenue from the newly permitted H200 sales into China. The Chinese state chose to forgo available American compute in order to force demand toward its domestic alternatives — Huawei’s Ascend line chief among them — accepting a capability cost now in exchange for indigenous capacity later. Nvidia’s share of the Chinese market, once a fifth of its data-centre revenue, has fallen below sixty per cent and is being driven down by policy rather than by technology.

Epistemic status: Probable and fast-moving, as of August 2026. The sequence of control changes and the December 2025 H200 decision are Confirmed from primary and authoritative reporting; the interpretation — that Beijing is deliberately forgoing American compute to force indigenisation — is well supported but is a reading of intent, and the policy could shift again. This is the single element of the compute map most likely to have changed by the time this volume closes; the managing editor should refresh it at publication and flag it as a live situation.

This is, in miniature, the boomerang dynamic examined elsewhere in this volume. A denial regime aimed at slowing Chinese compute has become an accelerant of Chinese self-sufficiency, because above a certain threshold of pre-existing capability, denial induces indigenisation rather than stagnation. The high-bandwidth-memory chokepoint is the pressure point to watch here: it is the component China most lacks for its domestic Ascend chips, the technology Beijing has reportedly pressed hardest to have decontrolled, and the place where the contest over compute will most likely next concentrate.

The buildout: where compute is being massed

The third feature of the map is scale, and the scale is difficult to convey without numbers that look like errors. The four largest American hyperscalers are committing on the order of seven hundred billion dollars to infrastructure in 2026 alone — roughly triple what they spent two years earlier — and when the Stargate build-out led by OpenAI, SoftBank, and Oracle is added, combined AI-infrastructure spending has, for the first time, cleared a trillion dollars. Longer-range estimates from major banks put cumulative buildout over the second half of the decade in the multiple trillions. Whatever else is true of the AI economy, the compute layer is being financed at a scale with few peacetime precedents.

The constraint on that buildout is no longer chips alone; it is power. A training rack draws more than an order of magnitude more electricity than a traditional enterprise rack, and the binding limit on new capacity has shifted from silicon to the substation. This is why the buildout has become entangled with energy policy, nuclear baseload, and grid capacity, and why the geography of compute is increasingly the geography of available power. It is also why a facility’s location has become a strategic choice rather than a commercial convenience: compute massed in one place is a concentration of capability, and — as a US envoy pointedly noted of Gulf data centres this year — a concentration of capability is also a target.

The third axis: sovereign compute

The final feature is the one that most complicates the two-bloc framing, and it is the reason compute belongs in a discussion of three markets rather than two. A widening circle of states has concluded that sovereign compute — training capacity owned and controlled within national borders, insulated from another jurisdiction’s export decisions — is strategic infrastructure worth acquiring at almost any cost. The Gulf states are the sharpest example. The United Arab Emirates, through G42 and the Stargate UAE campus, is building toward a five-gigawatt AI complex in Abu Dhabi, its first phase powered by a blend of nuclear, solar, and gas and equipped with among the largest concentrations of current-generation Nvidia systems ever deployed. Saudi Arabia, through its Public Investment Fund and the HUMAIN initiative, has committed tens of billions to a comparable ambition. The European Union’s AI Continent plan targets two hundred billion euros. India, Japan, and others are building national programmes of their own.

What distinguishes this sovereign layer from the commercial buildout is its logic. These programmes are contractually committed, politically supported, and deliberately insensitive to short-term return, because their purpose is not to earn a commercial rate but to secure a strategic capability. They add a second, non-correlated layer of demand beneath the hyperscaler market — and, more consequentially for this volume’s argument, they turn compute into an instrument of statecraft. A Gulf state that hosts a hyperscale campus is not merely renting servers; it is buying a position in the compute geography, and often buying it with American chips under American licences, which makes the data-centre campus a node where the export-control contest and the sovereign-capability contest intersect.

Epistemic status: Probable, as of August 2026. The sovereign-compute programmes and their headline figures are Confirmed from current reporting, but announced gigawatts and committed dollars routinely exceed delivered capacity, and the gap between the two is wide; these figures describe ambition and commitment, not operational capacity. The direction — compute as sovereign strategic infrastructure — is firmly established; the specific numbers will move.

Reading the map

Three things hold across the moving detail. First, the physical means of making training-grade compute is concentrated in Western-aligned monopolies, several of them on a single vulnerable island, and that concentration cannot be reconstituted quickly by anyone — which is simultaneously the West’s strongest card and the system’s greatest fragility. Second, the export-control contest over who may buy that compute has become genuinely incoherent, and has begun to function as a boomerang, with Chinese self-sufficiency accelerating in response to denial rather than collapsing under it. Third, compute has become an object of sovereign ambition well beyond the two principal blocs, adding a third axis to a competition usually described as bipolar.

For a reader whose task is to anticipate where capability will sit, the compute map is among the most reliable instruments available, precisely because compute is physical, expensive, and slow to build, and therefore harder to fake than most indicators. But it must be read with its date attached. The chokepoint geography is stable on a timescale of years; the export-control regime and the buildout figures are stable on a timescale of weeks. We have drawn the map as of August 2026 and marked where we expect it to move — the H200-into-China question, the high-bandwidth-memory chokepoint, and the gap between announced and delivered sovereign gigawatts. Where those move, the strategic picture moves with them, and the next reading of this map will differ from this one. That is not a weakness of the instrument. It is the nature of the territory, and the reason the map is worth redrawing.

References

Manufacturing chokepoints — lithography, foundry, memory, materials

Supply-chain chokepoint analyses of ASML’s EUV monopoly (100% of EUV lithography; ~50–60 standard EUV systems produced per year), TSMC (~90% of advanced logic below 7nm; dominant share of CoWoS advanced packaging), South Korean HBM dominance, and Tier-2/Tier-3 material monopolies (EUV photoresists, mask blanks, Zeiss optics, TRUMPF drive lasers), current to 2026.

Reporting on TSMC’s diversification roadmap (CEO C.C. Wei, April 2026 investor conference): new 3nm lines in Southern Taiwan, a second Arizona fab, and a second Kumamoto (Japan) fab across 2027–2028.

Analyses of ASML’s High-NA EUV transition and TSMC’s deferral of High-NA in favour of advanced packaging (2026).

Export controls and the China contest

NVIDIA Corporation, Form 10-K and Form 10-Q (FY2026), SEC filings, on the evolving export-control regime and its business impact.

US Congressional Research Service, “U.S. Export Controls and China: Advanced Semiconductors” (R48642), on the sequence of tightening and loosening actions, entity-list additions, and the HBM chokepoint.

Council on Foreign Relations, “The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable,” January 2026 (the December 2025 H200 decision and volume cap).

Brookings Institution, “Ball game’s over — the US is out of the AI chip market in China,” July 2026 (the H20/H200 sequence and Beijing’s response).

Reporting (CNBC, Al Jazeera, Tom’s Hardware) on Nvidia’s stalled H200 sales into China, the 25% levy, Beijing’s discouragement of purchases, and Nvidia’s falling China market share, late 2025–early 2026.

The buildout and sovereign compute

Reporting on 2026 hyperscaler AI capital expenditure (~$700–800bn across the four largest hyperscalers; combined AI-infrastructure spending clearing $1tn including the Stargate build-out led by OpenAI, SoftBank, and Oracle); and multi-year cumulative estimates from major banks.

Coverage of Stargate UAE / G42 (a 5 GW Abu Dhabi campus; first ~200 MW targeted for 2026 on NVIDIA GB300-class systems), Saudi Arabia’s PIF and HUMAIN programmes, and the EU AI Continent Action Plan (~€200bn through 2030).

Analyses of power as the binding constraint on AI data-centre buildout (training-rack power density; nuclear baseload; grid capacity), and of the security exposure of concentrated compute campuses, 2025–2026.