Southeast Asia's data-centre corridor: Johor, Singapore, and the cost of overflow

Singapore and Johor are usually compared as competing locations. Their relationship is closer to a pressure valve. Singapore is the region's connectivity hub and has the strongest legal and financial reputation, but it imposed a moratorium on new data-centre builds in 2019 over energy, land, and water concerns, relaxed it through a pilot call in July 2022, and in 2023 awarded just 80 megawatts to four proposals (Morgan Lewis 2026). Demand that could not be placed in Singapore went to the nearest place that could supply land, power, and speed.

That place was Johor. Cushman & Wakefield's first-half 2026 survey reports 1,110 megawatts operational in the state, up 24 percent over six months, with 602 megawatts under construction, up 91 percent, and a combined development pipeline of 3,088 megawatts. Malaysia as a whole has 1,039 megawatts under construction, the largest figure in Southeast Asia (Cushman & Wakefield 2026). Knight Frank's atlas puts Johor's colocation vacancy at 0.7 percent (Malaysia Property Trend 2026). Johor is not an emerging market. It is a full one.

The useful frame is a system of overflow. When a constraint binds in one node, demand moves to the next node in the cost curve, and the constraint reappears there in a different form. Singapore's constraint was policy, a deliberate rationing of megawatts. Johor's constraints are physical and political: transmission, generation, water, and tolerance among residents and regulators. Beyond Johor, the same movement has begun toward Batam, where Firmus has reportedly announced a 360-megawatt AI campus (Digital in Asia 2026).

Singapore's rationing is explicit and conditional. The second Data Centre Call for Application, announced on 1 December 2025, offered at least 200 megawatts and required best-in-class energy and IT efficiency, at least 50 percent of power from green sources such as biomethane, low-carbon ammonia, hydrogen, or on-site solar, and a power usage effectiveness below 1.3 for new builds (Morgan Lewis 2026). On 21 August 2026 the authorities provisionally allocated 200 megawatts, 50 each to Digital Realty, Equinix, Keppel Data Centres, and ST Telemedia Global Data Centres, all on Jurong Island, after a competition that drew more than 20 proposals (Telecompaper 2026). Capacity is awarded by the megawatt, in chosen locations, to chosen operators, and at a pace set by the grid and policy.

That design sends the latency-tolerant and capacity-hungry load, in particular AI training and large cloud regions, to Johor, a short drive from Singapore's fibre. The arrangement is attractive to hyperscalers, which commonly pair a small Singapore footprint with a larger Johor campus. It also moves the externalities. Singapore keeps its power and water for other uses while Johor supplies raw water to Singapore under a 1962 agreement and now hosts the load that Singapore declined. The policy logic is rational for each government and creates a distribution of costs that neither fully owns.

For Malaysia, this is both an opportunity and a trap. It attracts capital, jobs in construction, and a place in the AI supply chain. But when a host's growth is a function of another jurisdiction's restrictions, its bargaining position depends on those restrictions staying in place, and its infrastructure bears the load of a decision it did not make.

Cushman & Wakefield describes the region's AI and cloud build-out as power-constrained execution, in which the challenge increasingly lies in securing the power infrastructure needed for the next generation of workloads (Cushman & Wakefield 2026). In Malaysia the numbers show a system moving fast but with considerable uncertainty about how much of what is announced is real. The Energy Commission reports that data centres reached 9.28 percent of total electricity consumption in September 2026, that peak demand has risen above 21 gigawatts, and that data centres are projected to account for 31 percent of Peninsular Malaysia's demand by 2035 (Eco-Business 2026). The Energy Ministry told Parliament that TNB supplies 36 operating data centres with planned supply capacity of around 4.5 gigawatts and has met the demand of 23 projects under construction with a maximum demand of 3.8 gigawatts (Malay Mail 2026).

The gap between declared and actual load is the central signal. Parliamentary data from late 2025 showed data centres drawing 603 megawatts against a declared maximum demand of 1,276 megawatts, roughly 47 percent, against a government target of 85 percent (Tech Wire Asia 2026). The shortfall reflects ramp-up as much as speculation, since a new campus fills over years, but it explains why the government began revising projected capacity to reduce stranded-asset risk and cost transfers to other consumers. The generation plan relies on gas and renewables: 6 to 8 gigawatts of new gas-fired capacity and up to 10 gigawatts of renewables by 2030, a 2.5-gigawatt solar-and-battery programme from 2029, and eventually power from Sarawak and the ASEAN Power Grid. The prime minister has said current supply is sufficient for one to two years (Tech Wire Asia 2026; Eco-Business 2026).

Water is the second binding constraint. In November 2025 Johor reportedly asked investors to postpone water-cooled expansions for roughly 18 months and halted new approvals for its largest, most water-intensive facilities, classed as Tier 1 and Tier 2 (W.Media 2025). Data centres must use reclaimed water supplied by Syarikat Air Johor, and in January 2026 the state said it would enforce water-efficiency benchmarks at Singapore's level before approving projects (New Straits Times 2026).

The investment record explains the speed. Reported commitments include Microsoft's 2.2 billion dollars announced in May 2024 and a new Johor cloud region, Google at roughly 2 billion dollars, Amazon Web Services at about 6 billion dollars through 2038, and ByteDance at around 2.1 billion dollars for an AI hub in Johor (Light Reading n.d.). YTL Power reportedly completed Malaysia's first Nvidia-powered AI data centre in Kulai in October 2025, a project of about 10 billion ringgit on a campus planned for up to 500 megawatts (Light Reading n.d.).

The figures need care. Announced pipelines differ by an order of magnitude depending on definition: Cushman & Wakefield counts a 3,088-megawatt development pipeline for Johor, while Knight Frank reports an incoming pipeline of 8,542 megawatts (Cushman & Wakefield 2026; Malaysia Property Trend 2026). The difference is the line between projects with land, permits, and power contracts and projects with an announcement. A buyer needing capacity within 24 months can rely only on the first. Each announced gigawatt that cannot get a grid connection is a headline, not capacity.

Malaysia has moved from welcoming all capital to selecting it. On 24 February 2026 Prime Minister Anwar Ibrahim confirmed in Parliament that applications for data centres unrelated to high technology and AI had already been stopped, formalising an informal moratorium that dated to mid-2024 (Tech Wire Asia 2026). The official reasons are electricity, water, and the risk of higher tariffs for households. A related rule, stated in December 2025, requires developers to bear the cost of grid upgrades in full so that they are not spread across the wider grid. Johor reportedly rejected about 30 percent of new applications in 2024 on utility and planning grounds (Tech Wire Asia 2026).

The pattern is a screening regime that favours hyperscale AI projects with long-term commitments over speculative colocation. It reflects a recognition that a data centre is a large, fixed, inflexible customer whose benefits, mainly capital expenditure and construction activity, accrue early and whose costs, electricity, water, and land, accrue for decades. Projects that pay their own grid costs, use recycled water, and bring anchor tenants fit the regime. Projects that queue for capacity on the strength of a declaration do not.

The regime is also a negotiating position. The government has said it takes a neutral approach to securing AI chips from both the United States and China (Malay Mail 2025). Its leverage as a host is highest while demand exceeds its willingness to supply.

The corridor's second structural risk is who uses the compute. Malaysia's trade ministry stipulated on 14 July 2025, under the Strategic Trade Act 2010, that export, transit, and transshipment of AI chips require a permit (Malay Mail 2025). That addresses the physical movement of chips. The harder question is access. The Bureau of Industry and Security clarified on 31 May 2026 that licences are required for advanced chips destined for any entity whose ultimate parent is headquartered in China or Macau, wherever it is located, a rule reported to have technically applied since November 2023 (Asia Times 2026).

Remote access is the gap. US controls govern the physical transfer of chips and not who uses them through the cloud, so Chinese firms can rent compute hosted in Southeast Asia. A Remote Access Security Act that would extend controls to cloud use passed the House of Representatives in January 2026 but has not become law, and Washington is reported to be weighing a crackdown (CNBC 2026). For Malaysia, which hosts both US hyperscalers and Chinese-owned operators within a short drive of each other in Johor, a tightening of remote-access rules could affect a meaningful share of tenants and the revenue they bring.

The risk is asymmetric. If Washington extends controls to cloud use, hosts in third countries become enforcement points and must verify customers. If it does not, hosts risk reputational and tariff retaliation for appearing to enable circumvention. Malaysia's neutral stance is rational but fragile: it depends on the continuing ability to serve both US and Chinese customers without being forced to choose.

The strongest argument against the thesis is that much of the pressure is overstated. Declared demand is nearly double drawn demand, announced pipelines are several times the delivery pipeline, and the government has the tools to ration. Water efficiency can be engineered, grid capacity can be built, and the plan for 6 to 8 gigawatts of gas and up to 10 gigawatts of renewables by 2030 is a credible response. A host that sets conditions early can still capture the benefits. The counter-case does not remove the constraints but changes their timing: they bite on the margin of speculative load, not on credible hyperscale projects that bring capital and anchor tenants.

Three paths follow. In managed hosting, Malaysia screens for AI-linked, self-funded, water-efficient projects, the grid delivers on a 12-to-24-month clock, and verified load converges on declared load. In overflow capture, announced capacity outruns the grid, tariffs rise, and the government is pushed to relax conditions or accept stranded assets. In a compliance shock, remote-access controls or tariff retaliation reduce Chinese-linked demand and expose over-reliance on tenants whose access is contested. The signposts are observable: the ratio of drawn to declared megawatts, new gas and solar additions against the 2029 schedule, a third Singapore call, the passage of remote-access controls, and the energisation of Batam campuses as the next overflow node.

The record supports a selective conclusion. Hosting AI compute pays where a jurisdiction can deliver power and water on the buyer's clock, screen demand so that declared load becomes real load, and keep access to both US and Chinese supply chains without breaching either side's rules. It disappoints where growth is a function of a neighbour's restrictions and the host's infrastructure absorbs the cost. Singapore will keep rationing, Johor will keep absorbing, and the next constraint will arrive wherever the overflow settles.

For governments, the priority is delivery of grid and water capacity on the buyer's timeline, strict verification of load, and a clear and stable position on compliance. For operators, siting should follow verified power rather than announced pipelines, and cooling design should assume water limits. For investors, the scarcest asset in the corridor is energised capacity, not land or announcements, and the value lies with owners of grid connections, anchor tenants, and permits already in hand.

The corridor is a lesson in how constraints move. Singapore solved its own problem by exporting it, and Malaysia now holds the problem and the opportunity. Whether it keeps the second depends less on demand, which is abundant, than on delivering electricity, water, and credibility at the speed of a technology that does not wait.