What this topic means

For most of the current AI cycle, the question "can we get more compute" meant "can we get more chips." That framing is now wrong, or at least incomplete. Across much of Asia the binding constraint has shifted to something less glamorous and far harder to fix: whether there is a place to plug the chips in.

Data centres need land, electricity, grid connection, water, and network connectivity. Chips can be manufactured in eighteen months. A new transmission line takes five to ten years, and a generating station longer. When demand outruns that, the shortage stops being a procurement problem and becomes an infrastructure problem — and infrastructure problems are governed by planning law, utility regulation, and local politics rather than by capital.

This guide covers where AI compute is actually being built in Asia, why it is being built there, and what stops it.

Why it matters

• Power availability now determines geography. The map of Asian AI compute is being drawn by electricity, not by where the customers or engineers are. That is why capacity is appearing in Johor rather than Singapore, in Hokkaido rather than Tokyo, and in Inner Mongolia rather than Shanghai.
• AI racks broke the assumptions data centres were built on. A conventional server rack draws something in the range of 5–10 kW. Dense AI training racks draw an order of magnitude more. Existing facilities frequently cannot host AI workloads at all — not because they lack space, but because they cannot deliver or remove that much energy per square metre. Much of the "data centre shortage" is a shortage of the right kind.
• It is where AI collides with domestic politics. Electricity prices, water allocation, and land use are the subjects citizens notice. Several Asian governments now face the question of whether foreign-owned compute should receive scarce power, and that argument is only beginning.
• The constraint is durable. Chip supply responds to capital within a couple of years. Grid capacity does not respond to capital on any timescale that matters to an AI roadmap.

The single most useful frame: training versus inference

Almost everything in this guide follows from one distinction.

Training is latency-insensitive. A model trained on a cluster three thousand kilometres from its users is no worse for it. Training therefore migrates toward whatever is cheapest and most available: remote land, stranded renewables, cold climates, and jurisdictions willing to permit large loads quickly.

Inference is latency-sensitive. Serving a model to users requires being reasonably close to them. Inference capacity is therefore stuck near population and business centres — which is exactly where land is scarce, grids are congested, and power costs the most.

Once you hold this distinction, the regional pattern becomes legible. Hokkaido and Kyushu make sense for Japanese training capacity and not for Japanese inference. Western China makes sense for training and not for serving Shanghai. Johor makes sense for both, because it is an hour from Singapore. And the reason inference capacity is the harder problem everywhere is that it cannot be relocated to where the power is.

Watch this frame weaken over time. As AI shifts from training runs toward serving huge inference volumes, the fraction of demand that can be sited remotely falls. The industry's siting flexibility is a property of the training era, and it is likely to be temporary.

Singapore: the moratorium that reshaped a region

Singapore was Southeast Asia's data centre hub by default — political stability, rule of law, subsea cable landings, and financial-sector demand. By the late 2010s data centres had grown into a meaningful share of national electricity consumption, in a country with no domestic energy resources, essentially no land, and binding carbon commitments.

Data centres came to account for roughly 7–9% of Singapore's total electricity consumption, against a live IT load exceeding 1.4 GW — a striking share for any national grid, and an untenable one for a country with no domestic energy resources.

In 2019 Singapore effectively stopped approving new data centres. The pause was imposed administratively by the Economic Development Board and the Infocomm Media Development Authority rather than through legislation — a very Singaporean way to do it — and it held for around three years while energy and carbon implications were assessed.

What replaced it is more interesting than the pause. Singapore reopened in 2022 through a Data Centre Call for Application process: capacity awarded competitively to operators meeting efficiency and sustainability criteria, rather than granted on request. The pilot round allocated roughly 80 MW across four operators. In May 2024 the IMDA followed with a Green Data Centre Roadmap targeting at least 300 MW of additional near-term capacity plus a further 200 MW reserved for operators deploying green energy, conditioned on a power usage effectiveness target of 1.3 or better at full IT load, equipment efficiency standards, and low-carbon energy adoption.

The mechanism is the point. Singapore converted data centre capacity from something you apply for into something you compete for on efficiency. That is a fundamentally different instrument from a moratorium, and it gives the state a lever it can keep pulling.

The strategic logic is worth stating plainly, because it is the most sophisticated position any Asian government has taken here. Singapore concluded it could not win a volume competition and chose not to enter one. Instead it positioned itself to host the high-value, latency-critical, regulated workloads — financial services, government, regional headquarters — while allowing commodity capacity to locate across the strait. Singapore keeps the control functions and exports the electricity problem.

Johor: what the overflow built

Johor, the Malaysian state directly across the strait, absorbed what Singapore declined. Cheap land, cheap power, available water, and a road journey from Singapore measured in tens of minutes made it the obvious destination. Development concentrated around Sedenak and the surrounding technology parks.

The Johor–Singapore Special Economic Zone, signed on 7 January 2025, then formalised what was already happening commercially — and the terms repay close reading, because they codify the division of labour rather than merely encouraging it. The agreement spans nine flagship zones across eleven sectors, provides for cross-border renewable energy trading, offers concessionary corporate tax treatment for qualifying AI and technology supply chain investment, and includes explicit co-location provisions allowing firms to pair Singapore headquarters and financial operations with land- and power-intensive facilities in Johor.

That last provision is the striking one. The arrangement this guide describes — Singapore keeping the control functions and exporting the electricity problem — is no longer an emergent commercial pattern. It is written into a bilateral agreement as a feature.

Johor is now the clearest example in Asia of a cross-border compute region — a metropolitan economy split across an international boundary, with the expensive functions on one side and the energy-intensive ones on the other.

It also demonstrates that overflow destinations acquire the origin's problems on a delay. Johor's growth ran into its own power adequacy questions, tariff arrangements, and water allocation debates, and Malaysia has had to confront the same question Singapore did: how much national electricity should be committed to serving compute demand that is largely foreign-owned. The advantage was never permanent. It was a several-year arbitrage on planning and generation headroom, and arbitrages close.

Malaysia acquired a second complication that Singapore did not have. As US controls on advanced AI chips extended their attention to transshipment, Malaysia's trade ministry introduced, from July 2025, a requirement for 30 days' advance notification before exporting or re-exporting high-performance US-origin AI accelerators. The measure followed US pressure over potential diversion of Nvidia hardware to Chinese entities through facilities in Johor and the Klang Valley.

A country that attracted compute investment on infrastructure grounds found itself administering another government's export control policy — a role it did not seek and is not obviously equipped for. This is the clearest instance in the region of a state discovering that hosting compute carries obligations it did not price in.

Japan: two grids, and the periphery

Japan has a structural electricity problem that has nothing to do with generation capacity and everything to do with an accident of nineteenth-century equipment procurement.

Eastern Japan runs at 50 Hz. Western Japan runs at 60 Hz. The two halves are joined only by three frequency conversion stations — Shin-Shinano, Sakuma, and Higashi-Shimizu — which between them have historically capped east–west transfer at around 1.2 GW.

That figure deserves a moment. It is roughly one large power station, and it is the total amount of electricity that can move between the two halves of the world's fourth-largest economy. Japan does not have one national grid; it has two, joined by a bridge narrower than a single modern AI data centre campus. Surplus generation in the west cannot meaningfully serve demand around Tokyo.

Expansion is under way — the transmission coordination body OCCTO has committed on the order of ¥300 billion to raise transfer capacity to about 3.0 GW by 2027 — but even tripled, the interconnection stays small relative to regional demand.

Layer onto that the post-2011 reduction in nuclear output, the concentration of demand in the Kanto region, and land costs around Tokyo, and Japanese AI siting strategy follows:

• Hokkaido, particularly around Ishikari, offers cold ambient temperatures that cut cooling loads, local wind and hydro generation, and available land. Sakura Internet's zero-carbon AI cluster there is the flagship. The cost is distance from Tokyo — acceptable for training, marginal for inference.
• Kyushu has relatively abundant generation including geothermal and solar, and capacity is concentrating around Kumamoto and Fukuoka, following the semiconductor build-out at TSMC's JASM fabs. Compute and fabrication are being sited together, which is unusual and probably deliberate.
• The Kansai region hosts the most instructive single project: Sharp's former Sakai LCD plant, a site of over 1.2 million square metres that stopped making displays in 2024, with portions acquired by KDDI and SoftBank for conversion into high-density liquid-cooled AI campuses.

The Sakai conversion is worth dwelling on. A facility built for one capital-intensive industry Japan lost is being repurposed for the one it wants to enter — and it works because a display fab already has the grid connection, the water supply, and the floor loading a dense AI facility needs. Expect more of this. The scarce asset is not land; it is an existing heavy industrial power connection, and declining manufacturing regions across Asia hold a great many.

Japanese policy actively subsidises domestic AI compute through METI's GENIAC programme, which routes support through domestic cloud providers — including the operator of the Ishikari cluster. Japan's model subsidy and its data centre siting policy are the same policy. (See Sovereign AI in Asia and Japan AI Policy and Regulation.)

Korea: generation and demand in the wrong places

Korea's problem is geographic mismatch. Generation — nuclear and coal — is concentrated on the coasts, particularly the southeast. Demand is overwhelmingly concentrated in the Seoul capital region. Moving power between them requires long-distance transmission, and transmission line construction has faced sustained local opposition.

The result is that new large loads around Seoul face grid connection as the binding constraint. Government policy has responded by encouraging data centre siting outside the capital region, using incentives and connection priority — which runs directly against operator preference, because customers and engineers are in Seoul.

Korea also has a specific institutional memory that shapes its regulation. On 15 October 2022, a lithium-ion battery fire at the SK C&C data centre in Pangyo, south of Seoul, disrupted Kakao's services — messaging, payments, and ride-hailing used by tens of millions — for days, with partial effects on Naver.

The regulatory response was substantial, and is why Korean data centre rules are unusually focused on continuity: major platforms were designated critical digital infrastructure, amendments mandated multi-site active redundancy rather than backup alone, and battery storage safety standards were tightened.

The episode is a useful corrective to a common assumption. The risk that materialised in Korea was not grid failure or cyberattack. It was energy storage inside the building — precisely the equipment that dense AI facilities install more of.

Korea's compute story is inseparable from its sovereign AI programme, which treats public GPU capacity as national infrastructure. (See Korea AI Industry Map.)

China: relocation as national policy

China is the only country in Asia to have addressed the siting problem with an explicit national programme rather than through permitting and incentives.

The "Eastern Data, Western Computing" initiative (东数西算), launched in February 2022 by the NDRC alongside MIIT, the Cyberspace Administration, and the energy administration, establishes eight national computing hubs and ten data centre clusters. Three hubs sit in the east — Beijing-Tianjin-Hebei, the Yangtze River Delta, the Greater Bay Area — and five in the west and centre: Inner Mongolia, Guizhou, Gansu, Ningxia, and Chengdu-Chongqing. The intent is to move compute-intensive, latency-tolerant workloads west, where power is cheap, land abundant, and cool climates cut cooling loads, while latency-sensitive workloads stay east.

This is the training/inference split implemented as industrial policy, with efficiency mandates attached, and it is the most coherent policy response to the siting problem anywhere in the region.

It has also struggled — in a way that validates the frame rather than undermining it. Against national targets above 60% utilisation, reporting has indicated utilisation as low as 20–30% at some western facilities, particularly in Gansu and Ningxia. The causes are exactly what the training/inference distinction predicts: latency for real-time inference, high cross-provincial bandwidth costs, and eastern enterprise customers preferring compute near their users.

The lesson generalises well beyond China. You can move electricity-intensive workloads to where the electricity is, but only the ones that tolerate distance — and no state can mandate that a workload become latency-tolerant. Every remote siting strategy in this guide is subject to the same limit.

The cross-cutting constraints

Grid connection is the real queue. Operators and governments talk about generation capacity, but the practical delay is usually the interconnection process — securing a connection agreement, and the substation and transmission work behind it. Announced capacity and energised capacity differ by years, and the gap is where most data centre announcements quietly go.

Rack density changed the engineering. AI training racks draw far more power per rack than the facilities of the previous decade were designed for, which pushes operators toward liquid cooling. This is not an incremental upgrade — it changes the physical plant. Most existing Asian data centre capacity cannot host dense AI workloads without substantial rebuilding, which is why headline square-metre figures are misleading as a measure of AI readiness.

Water is a live political issue. Cooling consumes water, and in Singapore, Malaysia, and parts of India this competes directly with agricultural and municipal demand. Liquid cooling changes the water-versus-power trade-off rather than removing it, and closed-loop designs are becoming a condition of approval in several jurisdictions.

Renewables are timing-mismatched. Data centres want constant power. Solar and wind do not supply it. The gap is filled by storage, grid firming, or fossil generation, which is why "renewable-powered" claims frequently rest on annual matching rather than hourly supply. Interest in nuclear — including small modular designs — across Japan, Korea, and Southeast Asia is largely a response to this mismatch, though nothing on that path arrives soon enough to affect the current build cycle.

Export controls now shape siting. Where advanced accelerators may be deployed has become a matter of US policy as much as commercial choice, and successive changes to the rules governing chip distribution to third countries have introduced real uncertainty into Southeast Asian investment decisions. A data centre is a fifteen-year asset; export control policy has recently changed on a scale of months.

What to watch

• Energised versus announced capacity. The single most useful discipline in this topic. Announcements measure intent; energisation measures grid access. Track the second.
• Whether Johor's advantage persists. Malaysian power adequacy, tariffs, and water allocation will determine whether the overflow destination remains attractive, and the answer will shape Southeast Asian compute for a decade.
• Japan's 3.0 GW interconnection target for 2027. One of the few structural fixes available to Japanese siting, and slippage would be a genuine constraint on where Japanese AI capacity can go. Progress is slow and badly underreported in English.
• Korean transmission to the capital region. Whether new lines are actually built against local opposition determines whether Korean inference capacity can grow where it is needed.
• Whether western Chinese utilisation recovers. Utilisation of 20–30% against a 60% target is the clearest evidence anywhere that remote siting has a hard ceiling. If it improves, something changed about the workloads; if it does not, the ceiling is structural and applies everywhere.
• Industrial site conversions. The Sakai project suggests the scarce asset is an existing heavy industrial grid connection, not land. Watch for more closed fabs, smelters, and plants being converted across Japan, Korea, and Taiwan — this is probably the cheapest new capacity available anywhere in the region.
• Nuclear restarts and extensions. Taiwan reached zero nuclear generation in 2025 and immediately began reconsidering under semiconductor and data centre demand pressure. Japan and Korea face versions of the same question, and AI load growth is now a stated factor in energy policy debates that were previously settled on other grounds.
• The inference share. As inference grows relative to training, siting flexibility declines. Watch for capacity plans that no longer mention remote locations — that is the signal the industry's easiest option is closing.

Related guides

AI Chips and Hardware in Asia covers the accelerators these facilities house. Taiwan Semiconductor Ecosystem covers the power constraint in its most acute form, where fabs and data centres compete for the same grid on an island that cannot import electricity. Japan AI Industry Map and Korea AI Industry Map cover the domestic compute programmes driving national capacity build-outs, and China AI Industry Map covers the demand side of the Eastern Data, Western Computing policy.

Last updated: August 2026