Aanleiding: Jensen Huang’s Vision for Europe Runs on Plug-In Data Centers
J.Konstapel,Leiden,22-10-2026.
An essay based on “The Trunk and the Mesh: Europe’s Probable Technology Future, and What the Vacuum.Net Reads in It” (J. Konstapel, Constable Research, Leiden, October 2026).
Europe is not short of technology visions. The European Union has one; each member state has one; every large technology firm active on the continent has one. Read separately, they compete. Read together, they converge on a single direction: Europe will build more central computing capacity, in larger units, on a longer supply chain — and it will call the result sovereignty.
The purpose of this essay is to subject that direction to a disciplined business assessment. Three questions structure the analysis. First, what do the visions actually amount to in money, commitments and physical constraints? Second, what future is probable — not desired, not announced, but likely — when those constraints are applied? Third, what does an alternative theoretical lens, the Vacuum.Net theory of closure, suggest that the prevailing visions miss?
The assessment yields a sobering picture. The headline numbers are large — roughly €136 billion in private AI, data-centre and cloud commitments, a €200 billion EU mobilisation target — but the firm money behind them is a fraction of the rhetoric. Meanwhile the binding constraints are physical and slow-moving: grid connections that take seven to thirteen years, a semiconductor base stuck near 9 per cent of global output with no leading-edge production, critical raw material dependence concentrated in China, and industrial electricity at roughly twice the American price. The probable 2030 outcome is concentration without independence: ten to twenty large sites, a sovereign cloud that is largely American technology under European operation, and chips still fabricated in Taiwan. The essay closes with an alternative reading — the mesh rather than the trunk — and a set of practical directions that hold whether or not one accepts the theory behind them.
2. The Announcements: Generous Words, Cautious Money
2.1 Corporate commitments
Deduplicated across twenty-one announcements made largely in 2025 and 2026, corporate commitments for AI, data-centre and cloud capacity in Europe come to approximately €136 billion (an honest range of €120–150 billion depending on inclusion rules). The largest items are substantial: SAP’s €20 billion sovereign-cloud programme over ten years; Amazon’s €18 billion of new money in Spain within a €33.7 billion envelope, plus €7.8 billion through 2040 for the AWS European Sovereign Cloud (live since January 2026); Microsoft’s roughly $15 billion in the United Kingdom, $10 billion in Sines, Portugal, and a $6.2 billion capacity purchase in Narvik, Norway; Google’s at least €13 billion in Finland, €5.5 billion in Germany, €5 billion in Belgium and £5 billion in the United Kingdom; and Schwarz Group’s €11 billion for a 200-megawatt site at Lübbenau. Mistral, the European champion, is building a 44-megawatt site near Paris, €1.2 billion of capacity in Sweden, and a planned 1.4-gigawatt campus with MGX and Nvidia.
Two facts about this money deserve board-level attention. First, its provenance: 63 per cent comes from American firms, 35 per cent from European firms, and the remainder from ByteDance. The capital building “European sovereignty” is predominantly foreign. Second, its nature: announced private capital is mostly non-binding intent, convertible or cancellable on strategy changes, exchange rates and demand cycles.
Every firm uses the same word — sovereignty. Microsoft promises “a European cloud for Europe”; Amazon promises no operational control outside EU borders; Google says a sovereign future must be built in Europe, for Europe. Three things are conspicuously absent from all of these visions: any efficiency figure beyond Nvidia’s claimed 30 per cent power saving for its own systems; any quantification of jobs displaced (every announcement quantifies jobs created); and any proposal to place less computation, or local computation, at the centre of the design.
2.2 The EU’s plan and its arithmetic
The EU’s answer is the AI Continent Action Plan of April 2025, with the InvestAI financing facility targeting €200 billion. Of this, €150 billion is private intent from an industry initiative, and the €50 billion public element draws on existing programmes, largely in the next budget period.
The flagship — up to seven AI gigafactories, each needing 100–250 megawatts and costing €4–5 billion — reveals the gap between headline and commitment with unusual clarity. The formal call of 30 July 2026 allows €100 million plus €400 million of EU money for each of four smaller sites and €200 million plus €800 million for each of three larger ones: an EU maximum of €5.0 billion, of which only the first €1 billion is firm. The remaining €4 billion depends on a 2028–2034 budget that is still a proposal. The €20 billion gigafactory headline thus reduces, at present, to at most €5 billion in the actual call — and no site has yet been chosen.
The Commission’s own scoreboard, the State of the Digital Decade 2026, is candid about the distance to the 2030 targets: 20 per cent of firms use AI against a 75 per cent target; 46.7 per cent use cloud against 75 per cent; 60.4 per cent of adults have basic digital skills against 80 per cent; 10.5 million ICT specialists against 20 million; and an 8.8 per cent world chip share against a 20 per cent target that the European Court of Auditors calls “very unlikely” (its own projection is 11.7 per cent). On current trends, the digital-skills target is reached only in 2037.
2.3 Member states and the counter-voices
Twenty-one national strategies examined for this assessment all place AI and computing capacity at their core, yet committed public money ranges only from tens to hundreds of millions of euros — Germany’s roughly €1 billion for a gigafactory and the Netherlands’ €200 million AI factory in Groningen being typical. Announced private figures such as France’s €109 billion (February 2025) remain largely non-binding; of the €93 billion announced at the 2026 Choose France summit, €45 billion is firm.
Data-centre policy also divides the continent: France, Germany, Italy, the United Kingdom and Norway court data centres, while the Netherlands permits hyperscale sites in only two locations (with Amsterdam admitting none after 2035) and Ireland now requires new data centres to bring their own generation, sourcing at least 80 per cent from additional Irish renewables.
A quieter counter-current exists. The EuroStack initiative (Bertelsmann Stiftung and CEPS) argues for a full European technology stack at an estimated €300 billion to 2035 — money that does not exist. More than ninety European firms, including Airbus and Dassault Systèmes, have asked for European preference in public procurement. And some public bodies have already moved: Schleswig-Holstein migrated more than 40,000 mailboxes to open-source software, expecting over €15 million of licence savings in 2026 against a one-time €9 million cost; the Austrian armed forces run LibreOffice on some 16,000 workstations; the International Criminal Court adopted the German openDesk suite. The money gap, however, is stark: committed funding for open software, efficient computing (ternary models, neuromorphic chips such as Innatera’s 0.5-milliwatt processor, integrated photonics) and alternatives totals about €4.2 billion — 3.1 per cent of the corporate commitments, or 1.2 per cent without the €2.6 billion IPCEI-CIS state-aid project.
3. The Physical World Sets the Pace
Visions are drafted in years; infrastructure is built in decades. Five physical constraints discipline every scenario.
Electricity. EU data centres consumed 68 TWh in 2024; the Commission projects 114 TWh in 2030. Against Eurostat’s measured 2024 final consumption of 2,378 TWh, that is 4.8 per cent — the Commission’s own 3.2 per cent figure requires an implausibly larger denominator. The realistic 2030 share is 4 to 5 per cent of EU electricity: significant, not decisive at Union level. Locally it is decisive. Irish data centres already used 23 per cent of national electricity in 2025, heading for a projected 31 per cent by 2034; they consumed roughly 80 per cent of Dublin’s electricity in 2023.
Scale, made tangible. A single Nvidia DSX POD — a factory-built AI data centre in a shipping container holding seven Vera Rubin NVL72 racks — draws roughly 1.9 megawatts including cooling, consuming about 16.7 GWh per year: the continuous electricity of roughly 6,900 Dutch households. A one-gigawatt AI factory equals about 524 such PODs.
The grid. The binding limit is the connection, not the generation. Grid connections in Europe’s established hubs take seven to thirteen years; Dublin has a de facto moratorium until 2028; 4.6 GW of requests queue in Noord-Holland alone; an Italian queue holds more than 300 projects requesting over 50 GW. Europe needs an estimated €584 billion of grid investment this decade, and 40 per cent of distribution grids are more than forty years old. A container that installs in fourteen days does not shorten a ten-year queue.
Price. European industrial electricity at roughly €0.18 per kilowatt-hour costs about 2.2 times the American industrial price. Every compute investment in Europe carries this premium.
Chips and materials. Europe produces about 9 per cent of the world’s chips, with no leading-edge production: TSMC’s Dresden plant will make 28- to 12-nanometre chips for cars and industry, not AI accelerators; Intel cancelled its Magdeburg and Polish plants in 2025. Every GPU in every gigafactory will be imported — mostly from one island. Upstream, 71 per cent of the EU’s gallium imports and 45 per cent of its germanium come from China, whose export-control suspensions expire in November 2026.
The capital arithmetic closes the argument. Nvidia’s own figures put a one-gigawatt AI factory at about $50 billion, of which $40 billion flows to Nvidia. Seven gigafactories would cost roughly $46 billion against an EU call offering about €30 billion, public and private together.
4. The Probable 2030 Picture
Combining the four voices with the physical limits yields a scenario that is neither the visions’ promise nor their critics’ nightmare, but a disciplined middle case.
By 2027–2028, up to seven gigafactories are selected, clustered where power is cheap or guaranteed — northern Germany, France’s nuclear regions, Iberia, the Baltic pairing with Poland, Czechia, Finland — with first racks running late 2028 at the earliest. Private campuses move faster than public ones.
By 2030, probable outcomes include: computing concentrated in roughly ten to twenty large sites, increasingly outside the traditional hubs (whose share of European capacity falls from 62 to 55 per cent); data centres consuming 4 to 5 per cent of EU electricity, with Ireland above 25 per cent; a sovereign cloud that is largely American technology under European operation (AWS, Microsoft and Google sovereign offerings, plus the Delos and Bleu joint ventures), with European stacks growing from a 15 per cent base; leading-edge chips still from Taiwan and accelerators still mostly from Nvidia; firm AI adoption rising but far short of 75 per cent; and open-source migration growing steadily on 1 to 4 per cent of the money. The binding limits, in order: grid connection; chips and raw materials; public money after 2027; local acceptance of water, land and electricity bills.
By 2035, the picture becomes uncertain, hinging on four open variables: the 2028–2034 budget, the fate of the Cloud and AI Development Act, China’s export controls, and real demand for AI output — the last of which is currently projected by suppliers, not customers.
5. The Human Layer
A technology future happens to people in places. The places that gain — host municipalities receiving land income, construction work and a few hundred permanent jobs (Schwarz expects about 120 direct jobs at its Rostock site) — are rarely the places that carry the load: the roughly 30,000 Amsterdam homes delayed by grid shortage; the dry regions of Aragón and Talavera defending their water; every Irish household sharing a grid that data centres increasingly fill. Local resistance has measurable effects: Meta redesigned its Talavera project after protest, cutting projected water use from up to 4.8 million cubic metres a year to about 40,600, and withdrew entirely from Zeewolde.
Work shifts without being counted. Routine office tasks migrate to AI services; new work appears in construction, maintenance and operation. No national vision calculates this shift numerically — except Denmark, which plans to free about 30,000 public-sector posts through automation. The announcements count jobs created; none counts tasks removed. A business or a state that cannot measure displacement cannot manage it.
6. A Second Reading: The Vacuum.Net
The Vacuum.Net theory describes the world as a single strand forming a net: only what closes on itself persists; every state has a unique computed address (a balanced-ternary string, proven in theorem T7); capacity at a given depth is fixed, and exceeding it forces the next depth rather than a larger copy of the same one (T7.3); and base three is the cheapest integer radix, 5.4 per cent cheaper than binary. The essay is careful to separate what is proved in the theory, what is computed, and what is reading — an epistemic discipline rare in technology commentary.
Read through the net, the prevailing visions scale the term that does not close:
- More capacity at one depth. Gigafactories multiply copies at the same depth. The theory holds that exceeding a depth’s capacity forces the next depth — the visions grow the copy instead of moving deeper. (T7.3 proved; application is reading.)
- Sovereignty as ownership. A computed address is the same for anyone who applies the rule; an assigned address depends on the assigner. Europe can own a building that runs on a foreign stack, foreign chips and a foreign vocabulary — and call it sovereignty. One earlier analysis put it sharply: tripling the capacity of the wrong architecture is not sovereignty.
- Converging points. Five hubs holding 62 per cent of capacity; one island making the leading-edge chips; one country supplying 71 per cent of the gallium; three firms holding over 70 per cent of the cloud. On the net, a converging point without a second path is an open mesh — where tension passes first.
- Held open by money. Most gigafactory money awaits the next budget; most private money is not binding. Only what closes persists; a configuration held open by outside supply disappears when the supply stops.
- The register. BitNet b1.58, a working three-valued language model, needs 0.4 GB of memory against 2.6 GB for a comparable binary model, with estimated energy per token 12.4 times lower at one benchmark point less. The ternary register is already engineering practice — and it needs no gigafactory.
The