Nations are building their own AI compute.
For decades, computing power was something companies bought from a catalogue. Three decisions taken between 30 June and 30 July 2026 turned it into infrastructure that governments plan, finance and allocate.
On 30 June 2026, Japan’s Ministry of Economy, Trade and Industry announced the start of a programme to build a domestic multimodal foundation model for robotics and physical AI. Its implementing agency, NEDO, reviewed fifteen applications and selected a joint proposal from Noetra — a company founded by SoftBank, Sony, NEC and Honda — and the National Institute of Advanced Industrial Science and Technology. The project runs from fiscal 2026 to fiscal 2030. Asia Times put first-year subsidies at ¥387.3 billion and total support at roughly ¥1 trillion over five years.
On 24 July, NAVER, NVIDIA and Brookfield announced a proposed expansion of the sovereign AI factory at NAVER’s GAK Sejong data centre, from 55 to 200 megawatts by 2028. NVIDIA plans to invest USD 1 billion in NAVER, subject to closing conditions and to NAVER securing at least USD 9 billion of committed financing; Brookfield has entered a non-binding term sheet for up to that amount. On 30 July, the European High Performance Computing Joint Undertaking opened its call for AI gigafactories, EUROHPC-2026-CEI-AIGF-01, seeking up to seven facilities, expected to be located in at least seven member states.
None of this involves new technology. What is new is who pays, and on what terms the resulting capacity is handed out.
The stated reason is continuity of supply rather than national prestige
The Japanese file is unusually explicit. The summary of the selected proposal, published by NEDO, states that Japan’s domain-specific model development currently rests on foreign models, and identifies two consequences: a widening digital deficit, and doubts about sustained supply. Noetra’s president, a SoftBank executive, made the operational version of the argument in remarks reported by Asia Times: if an AI system controls a production line and its use is suddenly restricted by a change in foreign law, the line stops.
METI’s own announcement adds a second constraint. Because Japan’s energy self-sufficiency is low, it argues, reducing the power consumption of AI use matters more there than elsewhere. Compute is being treated as an energy question, not only a semiconductor one.
The Korean and European framings converge. NAVER describes the expansion as building a sovereign AI ecosystem; a Brookfield executive called access to trusted, sovereign and scalable AI infrastructure a growing strategic priority for companies and countries. The EuroHPC call states that the gigafactories will serve researchers, public bodies and businesses “with a particular focus on start-ups, scale-ups and SMEs”.
Public capacity arrives with allocation rules, and those rules are the usable part
For a company, the building is not the deliverable. The entitlement is.
NAVER’s version of the Korean announcement specifies that the 200-megawatt facility will establish a dedicated resource pool for emerging AI companies, bundling compute, software and support. The Japanese project commits to publishing trained model weights domestically during the project period, so that other developers and service providers can build on them. The European programme is the most codified of the three: nineteen AI Factories have been selected since December 2024, together with thirteen AI Factory Antennas, and the Commission puts combined EU, member state and associated country investment in supercomputing infrastructure and AI Factories at €10 billion over 2021–2027.
Access runs through published calls, not sales contracts. EuroHPC operates five access modes, applications are made through a single portal, and access is currently free of charge. Eligibility extends to industrial users and SMEs established in a member state or participating state — with the caveat that the main modes are built around open science and EU-funded research projects, so a proposal has to be framed accordingly.
JUPITER, Europe’s first exascale system at Jülich, runs 5,884 Booster nodes of four NVIDIA GH200 superchips each — roughly 23,500 GPUs. To qualify, each gigafactory must integrate at least three to four times the number of the most advanced AI processors currently available in Europe’s most powerful AI factories: over 100,000 per site, as the Commission puts it. The programme has grown: from up to five gigafactories under InvestAI in February 2025 to up to seven in the July call.
Switzerland sits inside this system, and only rejoined it in November
On 11 November 2025, Switzerland became the thirty-eighth participating state of EuroHPC, having originally joined in 2019, following association to Horizon Europe and the Digital Europe Programme with retroactive effect from 1 January 2025. A month earlier, on 13 October 2025, EuroHPC selected HEARTS — Helvetic AI Resources, Technologies and Services — as the Swiss AI Factory Antenna. It is to work with the LUMI AI Factory and collaborate with the BSC, MIMER and IT4LIA AI Factories, opening AI-optimised computing and data services through the Alps research infrastructure to academic, start-up, SME and public-sector users.
Switzerland also has an operating example of what national compute produces. Apertus, released on 2 September 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre, was trained on more than ten million GPU hours on Alps, in 8- and 70-billion-parameter versions, across more than a thousand languages including Swiss German and Romansh. It carries a permissive open-source licence, was built with Swiss data-protection and copyright law and the EU AI Act’s transparency obligations in view, and reaches business customers through Swisscom’s sovereign Swiss AI platform.
The limit is clear. The gigafactory call places the facilities themselves in EU countries, on one or several sites. Switzerland’s route into this build-out is participation and access, not hosting — which makes the access calls and the HEARTS service catalogue the instruments to watch, rather than a Swiss site announcement.
Power capacity is a selection criterion, not an operating detail
A 200-megawatt site running continuously would draw on the order of 1.75 terawatt-hours a year. For scale: the May 2026 short report on Swiss data centres, prepared for EnergieSchweiz at the Swiss Federal Office of Energy, puts Swiss data centre consumption at just under 2.1 TWh in 2024, and expects 2.5 to 3.2 TWh by 2030 — 4.3% to 5.6% of national electricity consumption. The remaining efficiency potential is put at around 0.78 TWh, or 38%.
The same report notes that artificial intelligence plays a subordinate role in Swiss energy terms today, because large-scale model training barely happens domestically. The compute behind Swiss AI use is, for now, someone else’s infrastructure. Hence the European call’s requirement that bidders demonstrate adequate power capacity infrastructure alongside best practice in energy efficiency, water efficiency and circularity. Grid connection, not only processor supply, is part of the bid.
What a management team can settle this quarter
Little of this is within a single company’s control. Several adjacent decisions are.
- Classify workloads by data sensitivity before renewing any AI contract, separating what may run anywhere from what must remain under a known jurisdiction. That distinction, not the choice of vendor, decides which of these facilities is relevant.
- Establish a portability benchmark. Running an open-weight model — Apertus through a Swiss-hosted platform — against the current provider on a real task yields a price and a fallback, both useful at renewal.
- Check eligibility for a EuroHPC access call. Switzerland has been on the eligible list since November 2025, and HEARTS is the point of first enquiry.
- Write location and exit terms into AI contracts: where inference runs, which law governs it, what happens if export rules change. The Japanese continuity argument is a contractual question before it is a geopolitical one.
- Treat power and cooling as the binding constraint in any on-premises plan.
The dates that will carry information
Four markers belong in the corporate calendar. Bids for the European gigafactories close on 12 November 2026, with selection expected in early 2027 and operations within eighteen months. Memoranda of understanding between the antennas and their linked factories were due by mid-2026, so the HEARTS service offer is the Swiss item to watch. Japan’s funding is contracted only for fiscal 2026 and 2027, with annual stage-gate reviews deciding on continuation thereafter, so the five-year figure reported by Asia Times depends on yearly confirmation. Korea’s 200 megawatts are due in 2028.
Compute is becoming a location factor in the same sense as industrial electricity or transport links: planned publicly, financed across budget cycles, allocated under published rules. The question for a management team is not whether the capacity is impressive. It is which of its own processes would justify an application.