The Case for Data Centres: What the Critics Miss

The protesters outside a Sydney planning meeting last August carried signs that said the same thing signs outside data centre hearings have said from Virginia to Johor: too much water, too few jobs, too little say. They are not wrong to be angry. They may be wrong about what they are angry at.

The debate over data centres — across Australia, across APAC, and increasingly across the legislative chambers of a dozen US states — is being conducted with bad information on both sides. The industry publishes press releases. The critics respond with worst-case projections. Communities make decisions in the middle, based on neither. That gap is worth closing. Not because data centres are innocent, but because the honest version of this argument is more useful than the one currently circulating.

Start with the strongest version of the case against.

In the United States, more than 230 environmental and public-interest organisations have signed a letter calling on Congress to impose a national moratorium on new data centre construction. The coalition, led by Food & Water Watch, argues that facilities are driving up electricity bills, increasing fossil fuel dependence, straining water supplies, and delivering far fewer permanent jobs than promised — all while receiving billions in public subsidies and tax breaks that communities will be paying off for decades.

Good Jobs First, which tracks corporate subsidies, has documented the gap between job promises and job delivery across US markets. The pattern is consistent: a developer announces thousands of jobs at a press conference beside a state governor. Construction workers arrive — several thousand of them, for 18 to 24 months. Then they leave. What remains is typically 200 to 500 permanent operations roles, many requiring specialist qualifications not available locally, in a building that is largely automated and does not look anything like the economic engine that was promised.

The jobs argument applies differently in APAC, as we will get to. But the underlying concern — that the public subsidy calculus is opaque, that communities are accepting risk without understanding what they are accepting — is legitimate and should not be dismissed as technophobia.

On 3 April 2026, Microsoft President Brad Smith sat down with Japanese Prime Minister Sanae Takaichi in Tokyo and announced a $10 billion investment in Japan’s AI infrastructure between 2026 and 2029. It is Microsoft’s largest-ever financial commitment to Japan, more than three times the $2.9 billion package announced two years earlier.

The announcement has three pillars. Technology: expanding Microsoft’s own data centre footprint and partnering with SoftBank and Sakura Internet to provide GPU-based AI compute while keeping data inside Japan’s borders. Trust: deepening cybersecurity cooperation with Japan’s national institutions. Talent: training one million engineers, developers, and workers by 2030, in partnership with NTT, NEC, Fujitsu, and Hitachi.

Sakura Internet’s shares jumped 20% on the day — their largest single-day gain since September 2025. Sakura had been selected as Japan’s government cloud provider just one week earlier. The timing was not coincidental. Japan is building a sovereign AI stack, and Microsoft is positioning itself as the spine of it.

This is not an extraction model. It is not a company landing in a passive market to capture cheap land and government incentives before moving on. It is a country with a projected shortage of 3.26 million AI and robotics workers by 2040 actively constructing the infrastructure it needs to compete — and insisting that its data stays home while it does so.

The critics’ frame — thin jobs, opaque subsidy, extraction — does not fit this picture. That does not mean the frame is wrong everywhere. It means it was built for somewhere else.

Here is the claim you will have seen: data centres will consume 25% of Sydney’s drinking water by 2035. Some versions say half. The figure has appeared in council meetings, parliamentary submissions, and dozens of news articles. It is worth examining carefully.

Sydney’s data centres currently use less than 1% of the city’s water supply — approximately 3.5 billion litres per year. The 25% figure comes from a Sydney Water worst-case projection that assumed no efficiency improvements and was based on peak demand numbers that operators themselves have contested. AirTrunk, whose facilities are among those cited, told regulators that early planning documents used peak demand figures and that actual modelled usage would be significantly lower. The company is transitioning its Western Sydney campus to run almost entirely on recycled water.

Amazon states that its Australian data centres avoid using water for cooling for 95.5% of the year, relying instead on fans. Most modern hyperscale facilities use closed-loop cooling systems — water circulates internally and is not consumed from municipal supply the way evaporative cooling towers are. The distinction between these cooling types is almost never explained in the coverage.

None of this means the water question is resolved. A Reuters investigation published in September 2025 found that the New South Wales government approved all ten data centre applications it reviewed without requiring measurable water reduction plans. Fewer than half of approved applications included projections at all. The NSW Data Centre Consultation Paper, released in March 2026, acknowledged that Sydney Water’s demand forecasts "differ drastically" from industry figures — and that precise water demand data is simply not available at the planning phase.

That is the real story. Not that data centres are drinking Sydney dry. The disclosure framework is broken. The answer is better data, not the wrong data repeated more loudly.

The energy picture is similarly complicated. Data centres globally consumed an estimated 415 TWh of electricity in 2024 — roughly 1.5% of worldwide demand. In Australia, the sector currently accounts for around 2% of grid-supplied electricity, a figure the government expects to triple by 2030.

Against this, the industry cites efficiency gains that are genuine and measurable. Google’s fleet-wide Power Usage Effectiveness — the ratio of total facility energy to IT equipment energy — stands at 1.09, compared to an industry average of 1.56. A PUE of 1.09 means 9% of power goes to overhead; 1.56 means 56%. The gap between a hyperscale facility and an ageing enterprise data centre is not marginal. Migrating workloads to a modern hyperscale cloud facility can reduce the energy required to run the same computation by a factor of three to five.

The commitments are substantial: AWS has matched 100% of its global electricity consumption with renewable energy. Microsoft has pledged to be carbon negative and water positive by 2030. AirTrunk secured Japan’s first green loan for a data centre and targets 24/7 carbon-free operations.

The honest caveat is that matching is not the same as additionality. Buying renewable energy certificates does not always mean new renewable capacity was built. And water positive by 2030 is a global pledge that says nothing about what happens in Western Sydney during a drought in 2027. Efficiency at the aggregate level and impact at the local level are different measurements, and the industry has been too comfortable letting one serve as the answer to the other.

The debate about data centres and communities is almost entirely framed around jobs — how many, what type, how permanent. That frame misses most of what is actually being built.

Take women in technology. In November 2024, more than 1,000 students from 44 schools across Greater Western Sydney gathered at Blacktown Leisure Centre for AWS Girls’ Tech Day — now in its sixth year, having reached over 3,200 students and 230 teachers from 95 local schools since inception. The event features robotics workshops, autonomous race car demonstrations through AWS DeepRacer, and panels of working female engineers from AWS’s own data centre operations teams. AWS has also run Girls’ Tech Day in Melbourne in partnership with Wyndham Tech School and Victoria University, co-developing classroom resources on AI in health science that reach schools unable to attend. These are curriculum-linked, teacher-trained programs embedded into the same Western Sydney suburbs where councils are calling for moratoriums on the facilities that fund them.

On skills more broadly: AWS has trained more than 400,000 Australians in cloud and AI skills since 2017. Its Women in STEM Cadetship program with Swinburne University produced a 100% job placement rate for its inaugural cohort of 19 women — all offered permanent AWS roles after a two-year earn-while-you-learn apprenticeship. The re/Start program specifically targets unemployed and underemployed people with no prior technology background. In India, AWS has trained more than four million people since 2017. In Indonesia, the Terampil di Awan program reached over 11,000 students across 26 provinces, and the Jabar Digital Academy has benefited more than 5,000 Indonesians including over 1,400 women and ten participants with disabilities.

The community infrastructure dimension is almost entirely unreported. In Rangareddy district, Telangana — the villages surrounding AWS’s Hyderabad data centre region — the InCommunities program has since 2020 renovated government schools and pre-primary centres, restored a 65-acre lake at Machanpally village, installed solar street lighting and water purification systems, built an oxygen generator producing 500 litres per minute at the local community health centre, and renovated a primary health centre now serving more than 35,000 people annually. AWS also runs a breakfast program for government school students in the area — typically their first meal of the day. These are not skills programs. They are the physical infrastructure of a functioning community, built because a data centre was built nearby.

AirTrunk’s social impact program, launched in March 2025, operates through an unusual funding mechanism: margin adjustments from the company’s sustainability-linked loans. When AirTrunk hits its environmental KPIs, the financial savings flow directly into community programs — accountability with a structure attached, not a press release. In Western Sydney, AirTrunk funds career pathways for students from low-socioeconomic high schools into STEM through Western Sydney University. AirTrunk’s founder Robin Khuda made a personal $100 million donation to the same university directed at women in STEM over 20 years. Through DeadlyScience, AirTrunk funds STEM programs for Aboriginal and Torres Strait Islander students in regional and remote schools. In the Northern Territory, it funds solar-powered hydropanels that extract clean drinking water from atmospheric humidity for remote Indigenous schools. A company under scrutiny for its water consumption in Western Sydney is funding water access for Aboriginal communities in the NT. That tension is worth sitting with, and the program is worth acknowledging.

On biodiversity: AirTrunk funds conservation of the Mandai Mangrove and Mudflat Nature Park in Singapore through NParks, and supports WWF’s Cyber Spotter Program training citizen scientists to report illegal wildlife trade. In Japan, Science Tokyo Masters scholarships for women in STEM and AI teacher training in high schools through CLACK — training teachers, not just students, with the multiplier effect that implies.

Microsoft’s Japan commitment includes a research grant program for scientists working on large-scale AI simulation and a fellowship connecting researchers with its global network, targeting workers in robotics and manufacturing — sectors facing Japan’s projected 3.26 million worker shortfall by 2040.

None of this is above scrutiny. Community programs are announced alongside investment figures and rarely tracked with equal rigour. How many students attending Girls’ Tech Day in Western Sydney go on to technology careers? How many renovated schools in Rangareddy district maintain their improvements a decade from now? The industry should be answering these questions publicly. Governments should be requiring it as a condition of approval.

The actual social contract on offer in APAC is more complicated — and in many cases more substantive — than the debate currently acknowledges.

The strongest version of the community case against data centres is not about water or energy. It is about information.

In outer Sydney, in Johor, in Melbourne’s west, residents are encountering proposals through rezoning notices, planning portal submissions, and council agenda items. Not community briefings. Not open information sessions. Not voluntary disclosure of what the facility will consume, what it will emit, or what it will look like in ten years. A rezoning notice. The infrastructure arrives before the conversation does — and by the time the conversation starts, the decisions have usually already been made.

This is not primarily an environmental failure. It is a communication failure. And it is one the industry has the power to fix without waiting for regulation.

The misinformation that circulates about data centres — water figures cited out of context, job numbers quoted at announcement and never revisited, cooling system types misrepresented in local media — does not emerge from thin air. It fills the space that transparent communication would otherwise occupy. When a company declines to hold a community information session, or submits vague assurances about water use to a planning authority rather than actual projections, it is not protecting proprietary information. It is creating the conditions in which bad information thrives.

The irony is that the record, examined honestly, is more defensible than the industry’s reticence implies. AWS’s actual water consumption in its Australian facilities — primarily fan-cooled for 95% of the year — bears little resemblance to the figures circulating in some council chambers. AirTrunk is transitioning its Western Sydney campus to recycled water. The energy efficiency gap between a modern hyperscale facility and the enterprise servers it replaces is not marginal. These are facts that communities are entitled to hear, from the companies themselves, before a project is approved — not after opposition has formed.

The framework is imperfect and still being designed in real time. Australia’s NSW Data Centre Consultation Paper, released in March 2026, acknowledged that Sydney Water’s demand forecasts "differ drastically" from industry figures — and that precise water data is simply not available at the planning phase. That is a disclosure problem, not a consumption problem. The fix is not a moratorium. It is mandatory pre-approval disclosure — specific, measurable, independently verifiable — and a genuine commitment from operators to brief communities as neighbours rather than notify them as stakeholders.

The infrastructure is arriving faster than the trust. That gap is the industry’s to close, and it is largely one of its own making.

Arc Brief will keep tracking both.