The Physics of AI Power: Chips, Energy, and an EU–Japan Bargain
Yujiro Masaki is a researcher at the External Affairs Institute and an entrepreneur specializing in artificial intelligence, digital policy, and international cooperation. He studied for an M.S. in European Studies at UCLouvain and previously served at Japan’s Ministry of Foreign Affairs (Middle East Bureau). He is the Founder and CEO of Jexen. Views are the author’s.
1. Introduction
European AI policy still watches the wrong screen. Brussels debates model releases, benchmark scores and the next implementing act of the AI Act as if the contest were a software race with a regulatory coda. That is a misreading. The binding constraints of the mid-2020s are physical: graphics processors, leading-edge foundry slots, extreme-ultraviolet lithography, high-bandwidth memory, undersea cables, and electricity delivered on a timetable a developer will sign. NVIDIA, TSMC and ASML are not software stories. Whoever books the transformers, the wafer starts and the cable landings sets the conditions under which models are trained. Communiqués about “shared values” do not.
I spent seven years in Japan’s Ministry of Foreign Affairs, including the Middle East Bureau, watching energy security as a problem of straits and spare capacity, and I now run a Tokyo firm that puts AI into workplaces. A resource you cannot move or substitute on your own calendar is a diplomatic object. A model without a rack and a grid connection is a press release.
Japan’s leverage is upstream: wafers, photoresists and process tools that leading-edge fabs cannot replace on a political timetable. Europe’s is equipment — lithography and the German optics without which those tools do not exist — plus market access. A bargain worth the name would be specific: European lithography and predictable access to the European market against Japanese materials, robotics and grid-aware industrial practice. Data-centre siting is energy policy. Trade directorates, energy ministries and EEAS Asia desks should treat it as one file.
2. The stack Europe pretends is software
The internet produced the same illusion. In the 1990s governments talked about content, PCs and handsets and under-designed landing stations, exchanges, data halls and process tools as national infrastructure. Platforms and clouds accrued elsewhere. AI punishes that error more severely. A thin last mile still carried email. Without accelerators and megawatts, training does not start.
The value chain is a stack. At the top sit research, models and developer environments. Below them sit GPU and custom-accelerator design, and the software lock-in that makes a purchased chip usable the week it arrives. Then leading-edge logic, high-bandwidth memory, and the packaging that bonds the two. Then lithography, materials and process tools. Then the data centre as a factory — power, water, cooling, transformers — and the fibre and submarine cables that move what it produces. Cloud operators sit on that base as wholesalers of compute.
European debate has clustered on the top: model risk, fundamental rights, the Brussels Effect applied to general-purpose AI. Those layers matter. They do not train a model, site a hall, qualify a Japanese resist on a European line, or decide whether a Dutch scanner can be serviced in Kumamoto without a six-month licence argument. The AI Act (Regulation (EU) 2024/1689) is a governance instrument. It is not a supply strategy.
Two physical facts should re-order the file. First, international connectivity is a cable business. The International Telecommunication Union’s public estimate is that submarine systems carry about 99 percent of international internet traffic. Japan’s Ministry of Internal Affairs and Communications recorded in 2025 that roughly 90 percent of the country’s data centres sit in the Tokyo and Osaka belts, and that international landing stations cluster on the Bōsō and Shima peninsulas. Europe has the same shape: a handful of metro markets, a handful of landings, and a political fight over who may connect how many megawatts to which substation.
Second, cloud is not a neutral utility. Amazon Web Services, Microsoft Azure and Google Cloud are the effective wholesalers of large-scale compute — and an extension of another jurisdiction’s industrial and export-control policy. When a licence condition changes, a reservation disappears. That is the recent history of advanced GPUs.
If “AI”, “semiconductors”, “energy” and “connectivity” remain four directorates, the binding constraint will always be the slowest permit. The unit of analysis should be the path from a qualified material to a powered, connected rack. Europe already thinks this way about gas. It has not yet applied the habit to compute.
3. Chokepoints: NVIDIA, TSMC, ASML — and Japan’s upstream
Public estimates of NVIDIA’s share of merchant data-centre AI accelerators still cluster around four-fifths of revenue. AMD and hyperscaler custom silicon are growing; neither dissolves the training-class lock-in. CUDA and a decade of developer habit are why a ministry that “buys a national GPU” discovers it has bought a circuit diagram and a recruitment crisis.
The foundry arithmetic now matches the software arithmetic. TSMC does not name customers. Reporting on its 2025 annual report, together with public comments by NVIDIA’s chief executive, indicates that NVIDIA became the foundry’s largest client that year, with the top customer at 19 percent of revenue and the second — widely identified as Apple — at 17 percent. Apple’s volumes used to set the cadence of leading-edge capacity. AI accelerators now do.
TSMC’s concentration of leading-edge logic on one island is the most important single fact in the compute economy. Earthquakes, water, power and people are as relevant as any communiqué about the Taiwan Strait. The so-called silicon shield is dependence, not insurance.
Diversification is under way and should be described without romance. Arizona is still proving yields. In Japan, Japan Advanced Semiconductor Manufacturing — TSMC’s Kumamoto joint venture with Sony, Denso and Toyota — entered mass production in December 2024 on 12/16-nanometre and 22/28-nanometre nodes. A second Kumamoto fab is planned. In Germany, European Semiconductor Manufacturing Company, the TSMC joint venture with Bosch, Infineon and NXP, is building in Dresden for production targeted in 2027 on 28/22-nanometre and 16/12-nanometre processes. These are real plants. They are insurance at mature and upper-mature nodes, not substitutes for the leading-edge capacity that still sits in Taiwan.
Rapidus, the state-backed Japanese venture in Chitose, Hokkaido, started a 2-nanometre pilot line in April 2025 and states an aim of mass production in 2027, working with IBM and with Europe’s imec in the research background. That is a public fact, not a completed card. Yield, customers and packaging will decide whether Chitose is a foundry or a monument.
An AI accelerator is unfinished until a logic die is bonded to high-bandwidth memory. TSMC’s CoWoS family and its peers have been as binding a constraint as wafer lithography. HBM volume supply is a three-firm market — SK hynix, Samsung and Micron — and through 2025–26 allocation has repeatedly gated GPU shipments.
ASML is the only firm that produces extreme-ultraviolet lithography systems in commercial volume. Nikon and Canon, which once led the industry, did not take the EUV bet. Nodes finer than the 7-nanometre class are, in practice, inaccessible without these scanners. The monopoly is the residue of decades of research, customer pre-financing by Intel, TSMC and Samsung, and a European industrial base willing to carry a technology that lost money for a very long time. Veldhoven is a reminder that Europe is a supply power when it funds a thirty-year problem.
The European fact is more precise than “the Netherlands.” ASML’s filings describe Carl Zeiss SMT as the single supplier of optical columns for its lithography systems, producible only at facilities in Oberkochen and Wetzlar. ASML has stated that a prolonged interruption of that relationship would leave it unable to conduct its business. EUV is a Dutch–German machine with a Japanese materials and track ecosystem around it. Policy that files “ASML” under Dutch trade and “resists” under a Japan desk will mis-negotiate both.
Japan lost the race to manufacture leading-edge logic at scale. It did not lose the race to make that manufacturing possible. Nikkei’s 2025 survey of major goods put Shin-Etsu Chemical first in silicon wafers at 26.3 percent and SUMCO second at 17.8 percent — together 44.1 percent of the world market. In photoresists, Tokyo Ohka Kogyo, JSR and Shin-Etsu occupied the top three positions, with a combined share of 60.5 percent. Tokyo Electron holds a commanding share of the coater/developer market that applies and develops those resists, and a near-exclusive position on EUV-compatible tracks. A leading-edge chip may be printed in Taiwan. A non-trivial share of what the printer consumes, and of the tool that prepares the wafer, is Japanese.
Two errors follow. Complacency — Japan is safe because it has materials — confuses upstream indispensability with downstream decision rights. Autarky ignores a cross-border division of labour. The design problem is which layers must remain indispensable, and which should be complementary with allies.
Robotics is the adjacent Japanese card. The International Federation of Robotics’ 2024 World Robotics report put Japan at 38 percent of global industrial-robot production. Using density inside Japan is not first in the world; making the machines — Fanuc, Yaskawa, Kawasaki, Denso — still is. If AI value accrues where models meet arms and sensors, Europe and Japan are closer to a joint industrial offer than either is to a frontier-model championship.
4. Electricity as industrial policy
Fatih Birol, the IEA’s executive director, has said in public that there is no AI without energy — specifically electricity. By 2026 that sentence is how site-selection committees talk.
The Agency’s 2026 update, Key Questions on Energy and AI, keeps its central projection close to the April 2025 Energy and AI report. Data-centre electricity consumption is estimated at 485 terawatt-hours in 2025 and is projected to roughly double to 950 terawatt-hours in 2030, or around 3 percent of global electricity demand. Consumption from AI-focused facilities is projected to triple over the same period. In 2025, data-centre electricity demand rose 17 percent, against about 3 percent growth in global electricity demand. Efficiency per task is improving at a rate the IEA calls unprecedented; simple text queries can now use less electricity than a television left on for the same interval. The fallacy is to conclude that the problem dissolves. Cheaper inference creates new uses — video, long-running agents, continuous tool-use — that eat the gain.
Capital is running ahead of electrons. The IEA records that capital expenditure by five large technology companies exceeded $400 billion in 2025 and is set to rise a further 75 percent in 2026, driven by data-centre investment. The same report notes that the technology sector accounted for around 40 percent of corporate renewable power-purchase agreements signed in 2025, and that the pipeline of conditional offtake agreements between data-centre operators and small modular reactor projects grew from 25 gigawatts at the end of 2024 to 45 gigawatts by publication. Microsoft’s support for an existing-reactor restart, and nuclear and geothermal contracts signed by Amazon, Google and others, are a search for firm, low-carbon power by buyers whose training runs do not follow the wind.
Bottlenecks have moved into equipment energy ministries already understand. The IEA is explicit that supply chains for gas turbines and transformers have tightened, and that planning systems now constrain connections. On-site gas is appearing, especially in the United States, as a response to interconnection queues — a bridge with fuel-price, emissions and consent costs.
Japan has begun to write this into energy policy rather than a digital sidebar. The Cabinet approved the Seventh Strategic Energy Plan on 18 February 2025, together with the GX2040 Vision. The 2040 electricity mix is set at 40–50 percent renewables, around 20 percent nuclear, and 30–40 percent thermal. In fiscal 2023, renewables were 22.9 percent and nuclear 8.5 percent. The plan dropped the post-Fukushima formula of reducing nuclear dependence “as much as possible” and adopted “maximum use” of both nuclear and renewables as carbon-free sources. Those are official targets, not forecasts: almost the entire existing reactor fleet would have to run for the nuclear number to be reached.
One plant moved. On 16 April 2026, TEPCO returned Kashiwazaki-Kariwa Unit 6 to commercial operation — the company’s first commercial restart since the Fukushima Daiichi accident, and the unit’s first in fourteen years. Local consent and technical interruptions are part of the public record. A restart announced is not a restart that lasts; an AI strategy that leaves eastern Japan on a thinner firm-power margin is not a strategy either.
The more transferable Japanese instrument is institutional. In March 2025 METI and the Ministry of Internal Affairs and Communications launched a Public-Private Advisory Council on Watt-Bit Collaboration — the pairing of watts and bits so that data-centre growth is planned against electricity and telecommunications at the same time. Report 1.0 was published on 12 June 2025. Days earlier, MIC issued the Digital Infrastructure Development Plan for 2030, which treats data centres and submarine cables as a single build and argues for dispersing both away from the Tokyo–Osaka concentration. Europe needs the habit, not the brand name: connection queues, transformer lead times and landing-station diversity on the same map as model policy.
Europe already knows the political cost of getting this wrong. Irish official statistics have shown data centres accounting for more than a fifth of national metered electricity. A de facto connection pause, followed by a tighter framework that asks large new loads to bring dispatchable capacity and additional renewables, is what a small advanced grid does when compute arrives faster than substations. The Commission’s June 2026 Strategic Roadmap for Digitalisation and AI in the Energy Sector names the same integration problem. Naming is not interconnection.
Data centres do not automatically raise household bills. The IEA is clear that spare generation can absorb them, and that in tight systems a large, fast load will raise prices unless connection order and cost recovery are designed. The quarrel is who pays for the substation. Cooling is also water policy; cities that pipe waste heat into district networks are making opposition harder to organise.
I used to look at energy as a pattern of interruption: a strait, a pipeline, a cavern. The equivalents for compute are thin interconnectors, transformers with multi-year lead times, and landing stations on one peninsula. Interruption does not require a war. It requires a queue.
5. What a real EU–Japan bargain would contain
The Digital Partnership already exists. It should be used, not re-announced. The third Council met in Tokyo on 12 May 2025; the fourth met in Brussels on 5 May 2026. On semiconductors the two sides have a memorandum of cooperation, an early-warning mechanism, and a recorded interest in joint research on PFAS alternatives, chiplets and heterogeneous integration, and devices and processes beyond 2 nanometres. They have also named submarine cables, Arctic connectivity, AI, quantum and 6G. That is the floor, not a bargain. A bargain names what each side trades, on a clock a plant manager would recognise.
Europe’s side is equipment and market access. ASML scanners and Zeiss columns are the gate on every leading-edge logic and HBM line Japan needs for Rapidus, for the qualification of Japanese materials at the next node, and for any European fab in the same industry. Predictable licensing, servicing and spare-parts treatment for Japanese and Japanese-sited facilities — Kumamoto included — is a European deliverable, not a courtesy. The second item is the market. The AI Act, the Chips Act and public-procurement rules for cloud and high-risk systems are how European demand is allocated. Japan needs to know which materials, tools and AI-enabled machines will be qualified, certifiable and purchasable in the Union on a published timetable, including under the Machinery Regulation and the high-risk provisions that apply from 2026.
Japan’s side is materials, tools, robotics and a way of running industry against a constrained grid. Wafers and resists are not spreadsheet “inputs.” Requalification on a new line takes years. A European foundry strategy that does not lock Japanese material qualification into ESMC, Intel’s European sites, STMicroelectronics, Infineon and GlobalFoundries will rediscover, late, that the scanner arrived and the chemistry did not. Tokyo Electron’s tracks are the silent partner of every ASML tool. Robotics is the application layer Europe says it wants and does not yet buy at Japanese density. Grid-aware practice is the export energy ministries should want: watt-bit planning, non-firm connections, demand that moves to surplus-solar hours. Europe has the directives. Japan has plants that have lived with expensive power and thin interconnectors for a generation.
The swap, stated without diplomacy, is this. Europe offers lithography continuity and a predictable internal market. Japan offers materials security, process-tool depth, robots that work, and a joint method for siting compute where the grid can take it. Both offer cable diversity: new landings, Arctic and intra-Asian routes, and repair-ship access.
Three things do not belong. The first is a promise that Rapidus or ESMC will be leading-edge by a political date — targets are not collateral. The second is a values paragraph in place of a transformer order. The third is export-control theatre. Alignment against leakage of training-class compute to military end-uses is a real joint interest; extraterritorial surprise is how allies push investment into a third jurisdiction. The G7 semiconductors contact group should produce a predictable licence, jointly explained to firms.
I have sat in rooms where “like-minded partners” was the entire operational paragraph. In the Middle East that language did not move a cargo. In the compute stack it will not move a resist drum. The partner who books the transformer and the cable landing will set the terms, values included.
6. Recommendations
1. Attach a physical-stack annex to the Digital Partnership. The 2027 Tokyo Council should open on a one-page map: lithography and optics; tracks and resists; wafers; HBM and packaging; foundry slots in Taiwan, Kyushu, Hokkaido, Dresden and Arizona; landing stations; connection queues. Name owners. Use the existing early-warning mechanism for transformers and resists, not only for geopolitical adjectives.
2. Put energy ministries in the room as principals. METI’s Agency for Natural Resources and Energy, the Commission’s energy directorate, and the ministries of member states that host halls — Ireland, the Netherlands, Germany, the Nordics, and those pitching southern and eastern European sites — should hold a standing watt-bit working group with MIC and METI. The deliverable is a shared list of connection delays, transformer lead times, and contracts that pay a data centre to be flexible rather than to squat on firm capacity.
3. Trade DGs should negotiate qualification and offtake, not “resilience” in the abstract. Offer Japanese wafer, resist and track suppliers a published path to qualification at European Chips Act lines, including ESMC. Offer European lithography and optics a published path to installation and service in Japanese leading-edge and packaging projects. Write both as dated annexes.
4. EEAS Asia desks should hold one map, not three. Taiwan/TSMC, Dutch–German lithography and Japanese materials are a single interruption scenario, currently split among a China/Taiwan file, a tech-sovereignty file and a Japan partnership file. A Taiwan contingency that does not list which European and Japanese plants stop, in what order, and which licences would reroute service, is not a plan.
5. Treat submarine cables as energy-adjacent critical infrastructure. Joint work should cover landing-station diversity (Japan’s Sea of Japan gap; Europe’s north-south and Arctic routes), repair-ship access, and trusted-supplier rules that do not become a subsidy war.
6. Buy the robotics-and-safety package, not a humanoid demonstration. Japan supplies machines; Europe supplies the Machinery Regulation, AI Act high-risk duties and a single market. A qualified inspection cell in European mid-sized plants on a short clock is worth more than another humanoid fair.
7. Align export controls for predictability, and say so to firms. Share the objective — managing transfer of training-class compute and the tools that make it — and publish a joint interpretation of licence timelines.
8. Stop issuing values communiqués that do not book interconnection. If a summit cannot announce a reserved substation, a qualified resist, a cable landing or a service licence, it should not announce “AI partnership.”
7. Conclusion
AI policy that watches model releases is watching a leaderboard. The contest is in GPUs, foundry slots, EUV scanners, memory stacks, cables and electrons. NVIDIA, TSMC and ASML are the visible chokepoints. Japan’s wafers, resists and tracks, and Europe’s lithography and optics, are the less visible ones. Electricity is no longer a utility input to a digital strategy. It is the strategy.
An EU–Japan bargain is available because the two sides are already complementary, already paired, and too small to pretend to autarky against the United States and China. Complementarity does not negotiate itself. What belongs on the Digital Partnership table is European lithography and market access against Japanese materials, robotics and a method for putting compute where a grid can bear it. Shared values are why that trade is possible. They are not the trade.
I left a foreign ministry that measured energy in barrels and straits. The equivalent units now are wafer starts, scanner-service licences, landing stations and transformers. The officials who book those units will write the next decade of AI policy. The rest is commentary.
Selected references
International Energy Agency, Energy and AI (April 2025); Key Questions on Energy and AI (2026); IEA news release, “Data centre electricity use surged in 2025…” (2026).
Fatih Birol, remarks at the IEA Global Conference on Energy and AI (December 2024) and the Artificial Intelligence Action Summit, Paris (February 2025).
Ministry of Economy, Trade and Industry (Japan), “Cabinet Decision on the Seventh Strategic Energy Plan” (18 February 2025); Agency for Natural Resources and Energy, Seventh Strategic Energy Plan and 2040 electricity-mix outlook.
METI and Ministry of Internal Affairs and Communications (Japan), Public-Private Advisory Council on Watt-Bit Collaboration, Report 1.0 (12 June 2025); MIC, Digital Infrastructure Development Plan for 2030 (11 June 2025).
Tokyo Electric Power Company Holdings, “Commercial Operation of Kashiwazaki-Kariwa Nuclear Power Station Unit 6” (16 April 2026).
European Commission / METI / Digital Agency / MIC, joint statements of the third (Tokyo, 12 May 2025) and fourth (Brussels, 5 May 2026) EU–Japan Digital Partnership Council meetings; Memorandum of Cooperation on semiconductors.
European Commission, Strategic Roadmap for Digitalisation and AI in the Energy Sector (3 June 2026); Regulation (EU) 2024/1689 (AI Act).
International Telecommunication Union, backgrounder on submarine-cable resilience (c. 99 percent of international internet traffic).
International Federation of Robotics, World Robotics 2024 (Japan’s 38 percent share of industrial-robot production).
Nikkei, “Major Goods and Services Market Share Survey” (2025), silicon-wafer and photoresist rankings.
TSMC, 2025 annual report (customer-concentration figures); public reporting on JASM mass production (December 2024) and ESMC Dresden.
Rapidus Corporation, public statements on the IIM-1 2 nm pilot line (April 2025) and targeted mass production (2027).
ASML Holding, annual-report disclosures on Carl Zeiss SMT as sole supplier of lithography optical columns (Oberkochen and Wetzlar).