Data Centers Are Efficient. That Doesn't Mean They're Fair.

In our first editorial — describing what a data center actually is and how it works — we introduced a metric called Water Usage Effectiveness. WUE measures how many litres of water a facility uses per kilowatt-hour of computing power delivered. Lower is better. The best modern facilities, using direct-to-chip liquid cooling, score close to zero. The industry average has fallen steadily over the past decade. The efficiency journey is real and the gains are documented.

WUE cannot tell you where the water comes from.

In Devanahalli, outside Bengaluru, India, at least eight data centers have been approved between 2013 and 2025. The area has no perennial water source. It depends entirely on groundwater. The stage of extraction — the percentage of water drawn relative to what the aquifer can replenish each year — sits at 169 per cent. Not approaching stress. Not pushing limits. One hundred and sixty-nine per cent of what the system can sustain, already being drawn.

A single 100 MW data center draws water at a rate equivalent to the annual needs of roughly 13,000 Indian households. In March 2024, as Bengaluru recorded its worst water crisis in decades, the city’s water board cut supply by 20 per cent to 38 bulk users, including major tech parks. By May, the city logged its driest April in 41 years. The investment approvals continued.

The WUE scores for these facilities are real. They appear in sustainability reports. They say nothing about the supply source below Devanahalli — whether that source is an aquifer, a dam, a river under agricultural pressure, or a municipal system already running at deficit. WUE is a ratio. It measures consumption against output. It does not measure whether the source has any slack left to absorb the draw, or who else depends on it.

That is the distributional problem this piece is about. Not whether data centers are efficient — they increasingly are. Whether the water they use competes with the people who live where it is drawn. And whether any government in APAC is asking that question before signing an approval.

Three pairs. Same metric on both sides. Completely different watershed reality below it. Expand each pair to see what WUE cannot see.

Japan contains one of the clearest natural experiments in data center location policy available anywhere in APAC — because the same operators, running the same equipment, made different choices within the same country and measured the difference directly.

Tokyo, Japan holds 132 operational data centers. Grid connection wait times in inner Tokyo now stretch up to ten years. Land prices rose 69 per cent in 2024. Summer temperatures regularly require full mechanical cooling — every degree of heat rejected goes into a system drawing from a municipal supply network that serves one of the world’s most densely populated cities. The pressure has produced engineering creativity that borders on the absurd: a consortium recently announced plans to test installing modular server racks beneath active railway overpasses.

Hokkaido, Japan — the country’s northernmost island, 800 kilometres north — offers a different arithmetic entirely. Temperatures in Ishikari City stay below 18 degrees Celsius from October through May, allowing facilities to run without any backup air conditioning for seven months of the year. The Ishikari Bay New Port industrial zone requires all tenants to operate on 100 per cent renewable energy. One operator measured the gap against its own Tokyo facility: locating in Hokkaido reduced power consumption by 40 per cent and eliminated freshwater draw almost entirely.

In Bibai, Hokkaido, the White Data Center uses snow as its primary cooling medium — the same snow the municipality was already paying ¥400 million per year to clear from roads. Heat from the servers slowly melts an insulated mound outside the building. The cold water circulates through the facility via closed loop. Water warmed by the cooling process, exiting at around 33 degrees Celsius, is used commercially to farm eels. The waste output of a data center becomes an input for local aquaculture. No supply source drawn down. No community competing for the same water.

Japan’s Environment Ministry subsidises the Hokkaido choice at 50 per cent of capital cost. The instrument is well-designed. But it is an invitation, not a mandate. AI training workloads — which have no latency requirement, for which proximity to users is irrelevant — continue flowing into Tokyo. The right geography exists and the government is trying to signal it. The policy instrument to direct the right workloads toward it does not exist.

In Issue 003, we described how Singapore’s planning controls pushed data center investment across the Causeway into Johor, Malaysia. We called it adjacency rather than sovereignty. The water version of that story is more direct.

Singapore imposed a moratorium on new data center construction in 2019, citing energy and water concerns. When it lifted the moratorium in 2022, it came back with hard conditions: new facilities must achieve a Power Usage Effectiveness of 1.25, source at least 50 per cent of power from green energy, and comply with mandatory water efficiency management requirements — metered monitoring at key points, and a Water Efficiency Plan submitted to the national water authority. These are access conditions, not voluntary commitments.

The problem is that Singapore’s framework stops at its border.

As Singapore’s door closed, investment flooded into Johor. The Malaysian state grew from 10 MW of data center capacity in early 2021 to over 1.3 GW by 2024, becoming the fastest-growing data center market in Southeast Asia. By late 2025, Johor’s state government told operators using evaporative cooling to defer new water connections until mid-2027. A sand mining accident that closed four treatment plants cut supply to more than half of Johor’s 1.7 million residents for up to twelve hours.

The Diplomat named the dynamic precisely: Singapore’s environmental cost did not disappear. It moved. The benefits of compute flow globally. The water stress lands locally, across a border that policy cannot cross.

The geography makes it visible. Johor’s water resources sit in the state’s central and eastern regions. Hyperscale investment concentrates in the south — closest to Singapore’s connectivity infrastructure, furthest from Johor’s supply. Singapore’s demand geography is being physically stamped onto Johor’s water stress map. There is no bilateral framework to account for this transfer. No shared supply assessment. No mechanism by which Singapore’s hard conditions follow its overflow investment across the Causeway.

In October 2025, the government of Andhra Pradesh, India, approved a $15 billion, 1 GW AI hub for Visakhapatnam — Vizag — on India’s eastern coastline. In processing the approval, the state did something that deserves attention. It issued Google a distribution company licence — a DISCOM licence — making the facility the first private entity in Andhra Pradesh to procure and distribute its own electricity independently of the state grid. Google will not compete with Vizag households for power. A dedicated, separated supply arrangement was created as a condition of approval. The facility is structurally removed from the domestic electricity system before a single server is installed.

That is exactly the right instrument. Applied to power, it eliminates the competition between industrial demand and household supply at the point of approval — not managed after the fact, not mitigated through efficiency commitments, but structurally separated before ground is broken.

Vizag sits on the Bay of Bengal. Seawater cooling — which Google already operates at its Hamina, Finland facility — is structurally available. A water equivalent of the DISCOM licence would designate a non-freshwater source as the approved cooling supply: a seawater intake, a treated wastewater connection, recycled industrial effluent. The facility connects to that supply. It never draws from the network that households depend on.

The Andhra Pradesh government demonstrated it knows how to build this instrument. The DISCOM licence is proof of concept. The approval documentation for water says arrangements will be coordinated “in consultation with the developer.” The Bay of Bengal is metres from the campus boundary. The right model was invented for one resource and not extended to the one that needed it most.

The three cases share a structural gap, not a failure of intent. No government in APAC has yet arrived at a data center approval with a water stress map in one hand and a designated supply arrangement in the other — and made the second a condition of the first. In Hokkaido, the subsidy exists but the mandate does not. In Johor, the conditions exist on one side of a border with no framework to account for what crosses between them. In Vizag, the instrument was created for power and set aside for water.

Efficiency is the right answer to the wrong question. The question the industry answers is: how much water per kilowatt-hour? The question nobody is asking at the approval stage is: which water, from where, and who else depends on it? Until that question has a mandatory answer, the scores will keep improving and the supply sources will keep being drawn down — and both things will be true simultaneously.

Issue 006 examines the energy side of this argument — who absorbs the grid costs of hyperscale AI, and whether those costs are being socialised in the same way water costs are.