Where it is billed is not where it lands.
The first distortion in every compute map is accounting. Vendors book revenue to the entity that signs the contract, not the campus that racks the chips. NVIDIA's FY2025 geography reads like a map of where AI is being built. It is mostly a map of where invoices are addressed.
Read the other way, the United States line is the opposite trap. US-billed revenue near $61bn overstates US landings, because US-headquartered hyperscalers buy through American entities for campuses they are building in Ireland, Sweden and Singapore.
So a country's vendor revenue is a ceiling on what physically arrives, not a count of it. For the read that follows, billing tells you who is paying. Customs flow, grid filings and construction tell you where the silicon actually lands.
Source: NVIDIA FY2025 Form 10-K revenue by geography; company disclosure on Singapore invoicing. Total revenue $130bn; data-centre segment $115.2bn. The run rate has since roughly doubled: Q1 FY2027, the quarter ended 26 April 2026, booked $81.6bn total and a record $75.2bn data centre, split $60.4bn compute and $14.8bn networking, with Q2 guided near $91bn on the explicit assumption of no China data-centre compute revenue. NVIDIA has also replaced the geography-only breakout with two platform segments, Data Center and Edge Computing, which changes what this billing view can show from Q2 onward.
Announced gigawatts, against the hardware that can fill them.
Across ten markets the pattern repeats: a headline power figure, a far smaller pool of accelerators that can actually land and energise, and a single constraint that decides the gap. This is the hardware ceiling, not the power ceiling: even where the grid can deliver, the chips may not be landing, and Equilibrium maps the power and queue side in full. Pick a market. The bar is announced capacity; the fill is what can plausibly run at AI density in the near term.
The gap closes. It does not disappear.
Absorbable capacity is rising fast as chips land, substations energise and permits clear. On current build rates the eight-fold over-promise of 2026 narrows to roughly two-and-a-half by 2030, and the announced ceiling stays out of reach the whole way. Drag the year, or press play, to watch the tide come in. Tap a market to follow its own climb. Figures are Entelligencia estimates.
The gap is real everywhere. What causes it is not.
Strip out the announcements and five different constraints are doing the work. Scroll through them, one at a time, to see where each one binds.
If the silicon concentrates, the authorities follow it.
The divergence this page tracks, capital routed to one place and chips landing in another, has put competition authorities in the European Union, the United Kingdom and the United States on alert. That alert has since become enforcement: in March 2026 Competition Commissioner Teresa Ribera named Nvidia, Meta and Google directly and confirmed a formal review spanning the full value chain, from chips through cloud and models to deployment, while the FTC, the Commission’s gatekeeper process and the UK CMA run parallel inquiries into a cloud market where three providers hold about 68% of it, with obligations capable of biting from mid-2027. Their concern is not the build but the bottleneck: whether early movers holding exclusive, long-term cloud agreements with the dominant GPU supplier can hoard physical capacity and pull the ladder up behind them. The question they are testing is whether chip scarcity converts into durable market power downstream, in the hourly price of renting a cluster.
The proposed toolkit is unusually concrete. Rather than backward-looking revenue, regulators are exploring a compute-adjusted concentration measure built from the same customs-destination data this page uses, with obligations that trigger on cloud-rental price spikes rather than years of litigation. The remedies under discussion are non-discriminatory capacity auctions, with a guaranteed slice reserved for independent labs and startups, and a compliance bond sized to exceed the gain from hoarding, scaled by the odds of being caught. None of it is law yet. It is the clearest sign that the physical layer of AI has become a competition question, not only an engineering one.
Proposed frameworks · EU, UK and US competition-authority consultations and academic mechanism-design work, 2025 to 2026. Indicative, not yet enacted.
In July, the map this page tracks was redrawn.
Updated 17 July 2026Everything above describes flows under a control regime: silicon following licences, and licences following alliances. On 10 July, the regime moved. The US Commerce Department upgraded the United Arab Emirates to Country Group A:5, the tier reserved for NATO-grade partners, clearing licence-free exports of advanced accelerators to G42, Core42 and MGX. A market this page could only show as capital-rich and silicon-constrained became, in one filing, a legal landfall zone.
The redirection it unlocks is not marginal. Stargate UAE’s Abu Dhabi campus targets a first 200 MW cluster in 2026 on the way to 5 GW, and the $1.4tn ten-year Emirati investment framework into US technology was priced on exactly this access. Watch the customs-destination data over the next two quarters: if the thesis of this page holds, the billed-versus-landed gap for the Gulf should begin to close, and the chokepoint should migrate from the licence desk back to the substation, where it already lives everywhere else.
This reading is maintained as the regime moves. Graded under the report’s Verified / Announced / Contested / Estimated taxonomy; announced figures are commitments, not delivered capacity.
The over-promise is in their numbers too.
The eight-fold gap above is an Entelligencia estimate, so it is worth asking whether the market’s own reference dataset shows the same shape. It does. DC Byte’s Global Data Centre Index 2026 publishes total tracked capacity by stage for the metros it ranks, and in every large market the majority of that capacity sits at early stage: land, announcement or intent, not steel.
Figures as published in the DC Byte Global Data Centre Index 2026, Q1 2026 cut, graded Verified as published figures rather than as capacity on the ground. Metro definitions are theirs and are not always metropolitan: the Atlanta entry is charted at state level as Georgia, and the Greater Helsinki entry is charted at national level as Finland. Read the shares, not the boundaries.
Then read what their method does with it. Their index scores each market on demand, delivery and depth, where delivery is defined as under construction plus committed only. Early stage is collected, charted, quoted in the commentary, and excluded from every scoring dimension. The reference index for this industry already knows that most announced capacity is not deliverable. It simply does not count it, and does not say so.
The accelerators arrive. Then they have to be reached.
This piece tracks where compute physically lands against where it is billed. Dr Cristina Crucini of EXA Infrastructure makes the corresponding argument about the wire: site selection runs on four fundamentals, land, power, cooling and connectivity, and the fourth is “usually the one left furthest down the list, even though it’s just as often what actually breaks a site later”.
Her sharpest point is on resilience. Route diversity is counted in cables, but cables share landing stations, backhaul and chokepoints, and repair depends on a global fleet of only just over 60 ships, all in high demand. A cluster that lands somewhere unreachable is a different kind of stranded asset. Contributed
Read the full file in Europe →On the data. No customs regime anywhere publishes an AI-accelerator line. Chips and AI servers are recorded inside broad codes (HS 8542.31 for processors, 8473.30 for boards, 8471 for servers), so every landed-compute figure here is a proxy, built by reading customs flow, vendor disclosure, grid-connection filings and named-project construction against each other, and tagged Estimated or Contested where it cannot be verified. Announced-capacity and absorbable-share figures are Entelligencia estimates drawn from the ten chapter reconciliations. Allocations and licences are tagged Verified; ambitions are Announced. Sources include NVIDIA and AMD disclosures, Eurostat Comext, US Census and USITC, NESO, JLL and Knight Frank market data, US Commerce and BIS, and the national customs and energy authorities cited per market. See the full method →
