Singapore's 200 MW Data-Centre Award: DC-CFA2, Jurong Island and the Shift from Megawatts to Green Molecules

The number invites the wrong argument. Critics read 200 MW as evidence that Singapore has fallen behind in the AI infrastructure race. Supporters read it as disciplined restraint. Both miss what actually changed. DC-CFA2 is a quality-and-location decision layered onto a hard green-power mandate; it was never meant to be the ceiling of national compute. The ceiling that matters is a 700 MW park on Jurong Island, and a third Call for Application the agencies have already signalled they will review within 18 to 24 months.

This analysis works through what was actually awarded, the economics underneath it, and the single constraint that now governs the entire build — not floor space, not even raw grid megawatts, but green megawatts that depend on fuels Singapore does not yet make at scale. It sits beside our buyer's guide to Singapore AI data-centre design and closes on the measurable signals that will tell executives whether this bet is working, and the evidence that would prove it wrong.

What Actually Changed on 21 August

The award is a permitting-and-power decision, not a technology launch — and its most consequential features are the conditions attached to each megawatt, not the megawatts themselves. DC-CFA2 opened on 1 December 2025, closed to applications on 13 April 2026, and drew more than 20 local and global bids. Four incumbents won. None is a new market entrant, and that is deliberate: the state wanted operators who already run multi-site Singapore platforms, can finance tropical high-density design, and can plug into Jurong Island's energy-and-chemicals ecosystem.

The word provisional carries weight. Finalisation still depends on land, grid, environmental and sustainability delivery, none of which is settled by the announcement. Data-centre construction typically takes two to four years, so these halls are a 2028–2030 story, not a 2026 compute story. Anyone reading the award as capacity that lands next year has misread the calendar.

This follows the 2023 pilot, DC-CFA1, which released about 80 MW to Equinix, Microsoft, GDS and an AirTrunk–ByteDance partnership. Taken together, the two controlled releases add roughly 280 MW of new permitted capacity since the 2019 moratorium — small against a live market industry trackers put near 1.4 gigawatts, and smaller still against regional AI demand. The scale is the point, not an accident. Singapore is releasing few megawatts on purpose, and raising the specification of every one it does.

Why Fewer, Greener Megawatts — and Why Now

Singapore is rationing data-centre power because the sector already draws about 7% of national electricity and cannot keep growing on imported gas without breaking the country's climate arithmetic. That 7% figure was the trigger for the 2019 moratorium on new data centres. The state lifted the freeze in 2022, but replaced open growth with the Call for Application framework — a way to admit capacity selectively rather than by market demand alone.

The physical context is unforgiving. Around 95% of Singapore's electricity is generated from imported natural gas, and the data-centre share of national demand is widely expected to head toward 12% by 2030. A city-state cannot grow gigawatts of AI load on more imported gas and call it an AI strategy. The 2024 Green Data Centre Roadmap set the direction: at least 300 MW of additional capacity, but tied explicitly to green energy. DC-CFA2 is the roadmap made concrete, and its requirements tightened sharply against the pilot.

The eligible green pathways are specific: biomethane, low-carbon ammonia, low-carbon hydrogen, novel fuel cells with carbon capture, or vertical and building-applied solar. That list is the policy, not the packaging. Singapore is land-poor, so domestic solar is a rounding error at 50 MW of IT load, and building-integrated panels cannot run GPU clusters. The 50% rule is therefore a bet on imported low-carbon molecules and a few island-scale plants — pathways, not a pretence that rooftops will do the work.

Is 200 MW Enough for the AI Era?

No, if the question is whether Singapore can host frontier-model training onshore at the scale of the United States or even Johor. Yes, if the question is whether it can keep a premium, regulated slice of inference, enterprise AI, capital-markets compute and sovereign workloads without overloading its power system. Both answers are true at once, and conflating them is what produces the sterile "200 MW is a trickle" argument.

Two anchors set the scene. Singapore's live capacity sits near 1.46 GW in 2026, yet only about 20 MW was under construction in mid-2026 against a development pipeline of roughly 980 MW. The binding constraint is permission and power, not developer appetite — a market where operators want to build far more than the state will energise. The second anchor is regional. Wood Mackenzie's base case has Southeast Asia's data-centre power demand quadrupling from about 2.6 GW in 2025 to 10.7 GW by 2035, with a high scenario reaching 13.7 GW.

The geography of that growth tells the story. Singapore holds about 54% of the region's data-centre load today, but its own capacity barely moves — from roughly 1.4 GW now to about 1.9 GW by 2035 — while Malaysia climbs toward 4.5 GW and Thailand toward 2.6 GW. Malaysia alone accounts for around 60% of all proposed regional projects. Singapore does not lose the region; it stops being where the volume lands.

So the honest forecast is a barbell. Onshore, Singapore keeps high-assurance, high-density, liquid-cooled capacity for latency-sensitive inference, regulated data, financial-market systems and sovereign AI that must stay under Singapore law. Nearshore — Johor, Batam and, later, the rest of Indonesia and Thailand — absorbs training clusters, batch analytics and anything power-hungry and legally portable. That is not a failure of industrial policy. It is the geography of electrons: capacity follows the cheapest firm power, and Singapore's is neither cheap nor abundant.

Efficiency also changes what "200 MW" means. A 50 MW liquid-cooled hall energising in 2029 will host far more useful AI compute than a 50 MW air-cooled hall approved in 2018. Comparing megawatts across generations without rack density and utilisation is exactly how the "mere 200 MW" slogan overstates the shortfall — and how the government can understate the remaining scarcity. Each new watt works harder, and there are still too few onshore watts for training-scale clusters. Both statements sit in the same paragraph.

The Economics Behind the Build

Large expenditure is not the same as attractive returns, and the CFA2 economics turn on whether green-power costs and utilisation hold up across a 2028-to-2040 asset life. The capital intensity is real: Keppel's awarded facility, described as its Singapore data centre number 11, is a build exceeding S$1 billion targeting a PUE below 1.2 and a water usage effectiveness around 1.6. High-density AI halls cost more per megawatt than legacy colocation — the electrical distribution, the liquid-cooling plant and the structural loading all scale up before a single accelerator is installed.

Follow the rent and the barbell reappears. Land and permission scarcity favour incumbents that already operate Singapore platforms, which is why no new entrant won. The green-power mandate then transfers value a second time — toward whoever can actually deliver low-carbon molecules. Keppel is the structural standout because it sits on both sides: an operator that also owns roughly 1.9 GW of Jurong Island generation. Operators carry the utilisation risk and the green-premium risk; if the molecules stay expensive, either margins compress or the 50% clause quietly becomes a certificate exercise.

The depreciation mismatch is the CFO's problem. The AI accelerators inside these halls turn over on a three-to-four-year cycle as each silicon generation lands; the shell, the substation and the cooling plant depreciate over fifteen to twenty years or more. The building outlives several generations of the chips that justify it, which means the return case depends on sustained, re-fillable demand — not one wave of GPUs. Financing structures reflect that: Temasek-linked ownership at ST Telemedia, Keppel's balance sheet and generation, and the green bonds Equinix and others have issued in Singapore. Whoever holds the asset when AI demand normalises carries the residual risk.

Is Green Electricity, Not Chips, the Binding Constraint?

Yes — the scarce input for new Singapore AI capacity is now deliverable low-carbon power, and the award is underwritten by fuels the country does not yet produce at commercial scale. Time-to-power, the interval before a completed site can actually be energised, has become more decisive than time-to-build. A finished hall without green electrons on the meter cannot meet the mandate that let it be built in the first place.

Some of the pathway already exists. Keppel Sakra Cogen began commercial operations on 29 May 2026 — Singapore's first hydrogen-ready combined-cycle plant, 600 MW, able to co-fire up to 30% hydrogen at the outset and cut emissions by up to 220,000 tonnes of CO2 a year. It is fully contracted for 2026 and 2027. Alongside the 1,300 MW Keppel Merlimau Cogen, that puts roughly 1,900 MW of Keppel generation on or beside the island where the new halls will sit. This is what "industrial symbiosis" means in practice rather than in a brochure.

The harder half does not yet exist. The Keppel–Sumitomo–Advario ammonia project, appointed by the Energy Market Authority and the Maritime and Port Authority in October 2025, is a front-end engineering study for a 55–65 MW ammonia power plant with terminal and bunkering — and it has not reached a final investment decision. Green ammonia and certified low-carbon hydrogen remain expensive and supply-constrained globally. Biomethane volumes in the region are modest: the EMA–EDB sandbox only launched in October 2025, and the first physical bio-LNG contract in Asia was signed weeks later. Singapore's broader answer is imports — a target of about 6 GW of low-carbon electricity by 2035, with 13 projects conditionally approved to import 9.25 GW — but those are approvals, not energised cables.

This is where the award is risky in a useful way. If the molecules arrive at scale by the time the halls energise, DC-CFA2 will look like the start of a genuinely distinctive tropical, industrial-symbiosis AI park. If they do not, the 50% clause either slips commissioning or hardens into a paper construct met with certificates. Operators can lean on existing tools — Digital Realty already matches its Singapore load with 100% renewable coverage through Tuas Power — but the CFA2 language points at pathways, not only certificates, which is a higher bar than the pilot ever set.

How AI Is Rewriting the Data-Centre Hall

AI racks have broken the thermal and electrical assumptions of the conventional data centre, which is why a PUE of 1.25 in the tropics forces liquid cooling as a default rather than a design choice. A general-purpose server rack draws around 12 kW; a dense air-cooled rack of previous-generation AI GPUs tops out near 40 kW. NVIDIA's GB200 NVL72, the current rack-scale AI system, draws roughly 120 kW nominal and up to 132 kW under load — six to thirteen times a typical air-cooled rack — and ships only as a liquid-cooled unit. Megawatt-class racks are already discussed for around 2030.

Cooling is where the tropics bite. In a conventional Singapore hall, cooling can account for 30–40% of total energy, so direct-to-chip and immersion cooling are now design defaults rather than pilots — the only way to hold PUE near 1.25 at these densities. The tropical data-centre standard also raises chilled-water temperatures: Digital Realty's own pilot lifted operating temperature by 2°C across two halls and cut energy roughly 2–3%. Small per-hall gains compound across a fleet, and they are cheaper than new generation.

This is the mechanism behind the earlier point about efficiency. Higher accelerator density raises rack-level power, which changes cooling requirements, which increases electrical distribution complexity, which shifts site-selection economics toward markets where large blocks of firm, clean power can be delivered quickly. The chip, the rack, the cooling loop and the substation are one coupled system now. That coupling is exactly why the location decision came out the way it did.

Why Jurong Island, Not Another Industrial Estate?

Because power, land and green-fuel optionality now dominate site selection more than last-mile fibre — and only Jurong Island offers all three next to the plants that can eventually green the load. JTC's low-carbon data centre park, announced in October 2025, is about 20 hectares — some 25 football fields — with a published power envelope of up to 700 MW. That is not a conventional colocation campus. It is an industrial co-location bet, placed where Singapore concentrates its energy and chemical infrastructure.

Jurong Island is the country's energy-and-chemicals core: pipelines, cogeneration, an LNG import terminal running since 2013 at about 11 million tonnes a year, and a planned new-energies belt of close to 300 hectares — roughly 10% of the island — reserved for hydrogen-ready gas, ammonia, batteries and related fuels. Putting AI halls next to that kit is the entire rationale. The plausible synergies are concrete rather than aspirational.

• Shared power and microgrids. JTC, EMA and Keppel are studying island microgrids that mix gas, storage and emerging fuels, with a testbed study slated to begin in 2026 — turning data centres into controllable loads rather than passive sinks.
• Low-carbon molecules on the same island. Sakra Cogen is already running on up to 30% hydrogen, and the Keppel–Sumitomo ammonia project is in engineering study. Neither greens 200 MW of IT load on day one, but together they define the pathway the mandate is underwriting.
• Thermal reuse. LNG arrives at about −161°C, and studies have floated using that "cold energy" from the SLNG terminal as a cooling source; an NUS–Keppel–SLNG project estimated a potential 20% PUE improvement. In a climate where cooling is a third of a hall's energy, any real waste-cold scheme changes the maths.
• A research on-ramp. JTC and NUS are advancing a Sustainable Tropical Data Centre Testbed, phase two, inside the park — real IT load, liquid cooling, and integration with hydrogen-ready plants and microgrids that commercial halls are meant to inherit.

The trade-off is obvious and was accepted knowingly. Jurong Island is a high-security industrial island with hazardous-industry neighbours and constrained access, not a latency-free extension of the central business district or of Changi's cable-landing stations. Operators took that deal because the constraint hierarchy has inverted: power, land and green-fuel optionality now outweigh last-mile fibre for the workloads this park is built to host. If the 700 MW envelope fills over the next decade, this becomes Singapore's main new AI campus, and everything else is retrofit, densification of existing halls, and overflow across the causeway.

Who Controls the Bottlenecks?

The scarce positions in this build are not the servers; they are grid interconnection, low-carbon fuel supply, and the engineering capacity to deliver tropical high-density halls on schedule. Grouping the value chain by bargaining power rather than by logo makes the pinch points visible. The bottleneck owners are the ones with pricing power over the next phase.

• Winners. Incumbent operators with existing Singapore platforms captured the award by design; Keppel captures twice, as operator and as power supplier.
• Bottleneck owners. The Energy Market Authority controls grid connection and import approvals; ammonia, hydrogen and biomethane suppliers control whether the green mandate is met; specialist contractors and liquid-cooling vendors control whether halls energise on time.
• Commoditising layers. Merchant GPUs and generic colocation space compete on price and are easiest to substitute; margin there compresses as supply broadens.
• Concentration risk. A single awardee that both operates and generates power is efficient, but it also concentrates dependence — a point regulators will watch as the park fills.

The operators themselves illustrate the pattern. Digital Realty runs three Singapore sites at roughly 70–84 MW and adds a fourth here; Equinix operates five live sites plus a sixth from the pilot; ST Telemedia GDC runs six sites above 110 MW under Temasek-linked ownership. Keppel Data Centres runs eight facilities near 200 MW and, uniquely, controls its own generation on the island. That dual role is the structural tell of the whole exercise: the awardee best positioned to satisfy the green mandate is the one that can supply its own low-carbon molecules.

What Could Break the Investment Thesis?

The thesis breaks if the green molecules stay expensive, if efficiency gains outrun demand, or if the talent-and-compute gap pushes even inference across the causeway. A credible analyst has to hold all three against the consensus that AI capacity only compounds upward.

The molecules may not arrive. The anchor ammonia project is at engineering study, not final investment decision; low-carbon hydrogen and certified biomethane remain supply-constrained and costly. If they are not commercial by the time the halls need them, the 50% mandate either slows energisation or softens into paperwork. That is the single most important risk in the entire structure, and it is external to the operators — they cannot manufacture ammonia supply by wanting it.

An efficiency shock is the second risk. If model optimisation, custom silicon and falling cost per token mean materially less compute per unit of useful output, a 700 MW park sized against 2030 demand could overshoot. The counterweight is Jevons' paradox: cheaper inference has historically expanded usage rather than shrinking total demand. The evidence that would settle it is utilisation — whether the first Jurong halls run full or half-empty once energised.

The third risk is leakage, and the talent argument is its sharpest form. Johor already carries around 3.8 GW of data-centre demand and has attracted the equivalent of roughly US$42 billion in investment; its power and land are cheaper. If Singapore's green mandate lifts onshore costs far enough, even latency-tolerant inference migrates. Attracting AI researchers without abundant onshore GPU capacity tends to produce a commute-to-Johor or rent-from-a-hyperscaler model — and DC-CFA2 does not fix that. A 700 MW park, used well, can house national and enterprise clusters. It will not, by itself, make Singapore a training superpower; the largest training runs will stay where power is cheaper and land is looser.

Three Scenarios to 2035

The next 24 to 36 months resolve which path the park takes. Three are defensible on current evidence; none should be assigned a false probability.

• Accelerated build-out. Ammonia reaches final investment decision, hydrogen blending rises above 30%, and a DC-CFA3 opens toward the 700 MW envelope. Jurong Island becomes a credible tropical AI campus; incumbent operators and Keppel's generation arm capture the upside, and utilities see firm new load.
• Constrained expansion. Green-molecule supply stays tight or expensive and grid connection lags. Energisation slips, phased 10–20 MW blocks stretch out, and more inference that firms would have kept onshore leaks to Johor. Operators hold half-filled shells; the barbell tilts harder nearshore.
• Efficiency shock. Hardware and model gains cut compute per unit of output faster than demand grows. The park still fills, but slowly, and the premium onshore case narrows to the genuinely latency- and sovereignty-bound workloads. Investors reprice data-centre assets on utilisation, not announced capacity.

What Should Executives Watch Next?

The signal is not the next 50 MW award; it is whether ammonia-to-power and hydrogen blending on Jurong Island reach commercial operation before the halls need them. That single question decides whether the greenest data-centre mandate in Asia is a strategy or a slogan. The measurable indicators below are the ones worth tracking, because each moves the thesis.

• Ammonia final investment decision. The Keppel–Sumitomo project moving from engineering study to FID is the clearest green light; continued delay is the clearest amber.
• Hydrogen blending ramp. Sakra Cogen pushing beyond its initial 30% hydrogen share signals the fuel pathway is real, not nominal.
• DC-CFA3 timing. A third Call for Application in 2028–2029 confirms the 700 MW park is being used this decade; silence suggests the constraint bit harder than planned.
• Realised PUE at energisation. Whether the first halls actually hit 1.25 — or the sub-1.2 that Keppel has floated — versus the pledge on paper.
• Grid-connection lead times and low-carbon imports. Progress on the 6 GW import target, and how fast the utility can deliver firm clean blocks, gates everything downstream.
• Johor tariffs and any power or water constraints. The nearshore alternative's cost and availability determines how much load Singapore can realistically keep.
• Rack densities crossing 120 kW as the design point. As NVL72-class racks become standard, they reset the cooling and power assumptions every new hall is planned around.

Translated into the boardroom: the CFO should read the depreciation mismatch and the green-premium risk before treating a Jurong allocation as firm capacity. The CIO and CTO weighing onshore versus nearshore should size latency and data-residency needs honestly, because most training belongs across the causeway. The COO should plan around time-to-power, not time-to-build. And corporate strategy should watch which layer holds pricing power — increasingly the fuel and the grid connection, not the servers.

The Reboot Singapore Actually Launched

The reboot is not "more megawatts at any cost." It is fewer, denser, greener megawatts, parked next to the only place on the map that can eventually make those adjectives compatible. That is a coherent industrial strategy. It is not, by itself, an AI-era capacity strategy — and both statements belong in the same paragraph. Singapore is choosing to be the region's high-assurance core while the volume grows elsewhere, and doing it deliberately.

Here is the contradiction the award has to resolve. Singapore can manufacture demand for AI compute faster than it can manufacture the molecules to power that compute cleanly. DC-CFA2 commits four operators to draw more than half of 200 MW from green pathways by 2028–2030. The single plant meant to anchor those pathways — a 55–65 MW ammonia project — has not yet reached a final investment decision. Until it does, the greenest data-centre mandate in Asia rests on a fuel Singapore does not yet make. If the next constraint on artificial intelligence is measured in tonnes of low-carbon ammonia rather than in model parameters, which industry actually controls the next phase of AI economics?