Who Pays for the Cloud?

Artificial intelligence, data centres, and the stewardship Australia owes its own people

There's been a lot of news lately about artificial intelligence and the data centres built to run it, so it's worth pausing to ask why, because most of what makes the headlines, chatbots, government AI offices, corporate deals with tech giants, ultimately comes back to the same physical thing: warehouses full of computers.

A data centre is, at its simplest, a very large shed of servers, wired together and kept running around the clock. It generates enormous amounts of heat, which has to be removed continuously. Depending on the facility's design and climate, that may be done through air cooling, evaporative systems, chilled water, direct liquid cooling, or some combination of them, each carrying a different balance of electricity demand, water consumption and cost.

Older data centres, the kind that stored files or served ordinary websites, drew comparatively modest amounts of power. What's changed is the hardware inside the newer ones: training and running large AI models need dense clusters of specialised chips that draw several times more power per rack than before, and the cooling load that comes with them, whichever technology handles it, has grown to match. That is the actual reason this wave of construction looks so different in scale to anything that came before it.

Governments and companies are racing to build this capacity because it has become a genuine strategic asset, the computing power behind AI systems now spreading into finance, health, defence and everyday consumer products. Australia's government has spent much of 2026 courting that investment, brokering agreements with major American technology companies in the hope of securing jobs, sovereign capability, and some influence over how the technology develops. That is a defensible ambition. But it also means Australia is absorbing the physical footprint of a global race, substations, water licences, concrete, largely without the kind of public conversation that infrastructure this size would normally receive.

In Sydney's north, in the business parks of Lane Cove West and the neighbouring Ryde council area, a cluster of hyperscale data centres is going up. Two projects reported locally, one proposed for Mars Road and another at Artarmon, are estimated to need an average of around 1,405 and 2,300 kilolitres of water a day respectively, though those figures are the developers' projected average demand rather than measured consumption.

Ryde Council has estimated that the eleven data centres operating or proposed within its area could collectively require around 15,000 megalitres of water a year. Published reports place that at approximately three per cent of the relevant Sydney water-supply measure, although the precise percentage varies according to the denominator used.

Sydney Water has told the state pricing regulator, IPART, that citywide data centre demand could reach as much as 250 megalitres a day by 2035; compared with projected average daily water sales of around 1,327 megalitres in 2029–30, as per their 23 September 2025 final report. That is close to one-fifth of projected system sales, though the two figures are drawn from different forecast years and are best read as an indication of scale rather than a precise share. Residents in these areas report finding out about approvals after the fact rather than before it, which is the same complaint that runs through most of what follows.

The energy story is larger again, and better documented, because the Australian Energy Market Operator now tracks it separately from other industrial demand for the first time. AEMO estimates that Australia's operational data centres currently account for around two per cent of grid-supplied electricity use, with its planning assumptions projecting that share rising to around six per cent, 12 terawatt hours, by 2029–30, and to around ten to twelve per cent, 34 terawatt hours, by 2049–50. In its most recent Integrated System Plan, AEMO named data centres as one of the defining drivers of the near-doubling of electricity demand it expects across the whole grid over the next twenty-five years.

In New South Wales alone, industry figures from Data Centres Australia and DC Byte counted 44 data centres in the development pipeline as of 31 March 2026, representing 11.4 gigawatts of prospective demand, nearly four times the output of Eraring, Australia's largest coal station. Pipeline capacity is not the same thing as likely consumption: it includes active projects, early proposals and speculative connection requests, and AEMO itself filters out a share of this as phantom demand, projects that will never proceed, and independent analysis of early connection queues elsewhere has found similarly high phantom-demand ratios, a reason to treat the headline pipeline figures as an upper bound rather than an inevitability, not a reason to disregard them.

But even after that filtering, AEMO's central forecast still represents a rapid, structural change in how the grid is used, and the risk runs in one direction: if the extra renewable generation and storage does not arrive as fast as the demand does, Australia leans longer on coal and gas to keep the lights on, which is precisely the outcome the energy transition was meant to avoid.

Water, put honestly, is not yet a national crisis; current data-centre water consumption remains a very small proportion of Australia's total use. But national averages can disguise local pressure. A nationally modest industry can still become a material customer where multiple facilities are concentrated within the same drinking-water network, and Sydney Water's own modelling above points to a real and rapid local step-change over the next decade, even if the precise share depends on which projects proceed, which forecast year is used, and which measurement basis is applied.

The precise, scoped version of the claim, this will strain specific water systems in specific growth corridors well before it strains the country, is the one the evidence actually supports, and it is also the more useful one for anyone trying to plan around it.

Forecasts like these deserve neither automatic belief nor automatic dismissal. They need to be treated for what they actually are: scenarios built from stated assumptions, not settled facts.

Actual current consumption, a developer's approved project estimate, and the maximum development pipeline are three different categories of evidence, and public debate gets distorted whenever a maximum scenario is quietly presented as an accomplished fact. But it is equally distorted when the mere existence of uncertainty is used as a reason to do no planning at all.

That is the standard this piece has tried to apply throughout: state plainly what is measured, what is projected, and what is merely proposed, and don't let any of the three borrow the others' certainty.

The second track running through this story is not resources but governance, and it is the one the Senate's own inquiry into artificial intelligence and data centres, referred on 13 May 2026 and reporting by 16 November 2026, was set up to examine. The time for making a submission to the inquiry has been extended to 1 September 2026. Its terms of reference ask, among other things, whether existing regulatory frameworks are keeping pace with the sector's growth, including in relation to deals between government and global AI companies, and what the impacts on communities, industries, water and energy actually are. That framing matters, because it treats the regulatory question and the resource question as one inquiry rather than two, which is the right instinct.

On 23 March 2026 the federal government released its Expectations of Data Centres and AI Infrastructure Developers, at the time exactly that: expectations intended to shape a developer's social licence and influence which projects received government priority, operating alongside existing law rather than creating new legal duties.

That position shifted materially on 15 July, when the Prime Minister announced the government would legislate national standards for the next generation of large data centres, requiring them to underwrite new power supply, pay their full share of grid-connection costs, reduce demand when the grid is under stress, minimise water use, and fund the additional water infrastructure their development creates, alongside more coordinated input from local communities on where facilities are built.

A new Office of AI was established immediately within the Department of the Prime Minister and Cabinet to draft the standards, which are due to go to National Cabinet in August and to Parliament as legislation in early 2027. The announcement also drew immediate criticism from the Greens, whose member Sarah Hanson-Young chairs the Senate inquiry itself; the party has called for a moratorium on new data centres until the regulations are finalised, warning they risk draining Australia's energy and water resources, a position the government has so far resisted in favour of finalising the framework first.

Whichever side of that argument is right, the proposed standards directly address the cost-allocation question this piece keeps returning to, promising to convert what were voluntary expectations into legal obligations. This is a welcome recognition of the problem, and it shows the government now accepts that the cost-allocation and resource questions raised throughout this piece are serious enough to require legislation rather than voluntary expectations.

But a promise to legislate is not yet a law, and the open questions are exactly the ones the Senate inquiry itself is positioned to test: what capacity threshold will trigger the new rules, whether facilities already approved or under construction will be covered, how the promised new electricity supply will actually be verified, who will measure real water consumption once a facility is operating, and what enforcement follows if a developer's own projections turn out to be wrong.

The same speech also touched directly on the copyright question. Officials have been lobbied by at least one AI company, Anthropic, to change Australian copyright law to ease the training of AI models on Australian material, a request Anthropic has made openly. The Prime Minister has now stated publicly that Australian writers, artists and journalists should retain control over whether their work is used to train AI systems. That is a welcome commitment, and I mention the lobbying not to single out one company; it is simply the clearest illustration of the pattern this piece keeps returning to, that deals and lobbying can run ahead of the standards meant to govern them. The sharper question now is no longer whether copyright remains unresolved, but how a promise made in a speech becomes a right enforceable in law, and against whom.

South Australia is a clean illustration of the same tension playing out at state level. In June 2026, Premier Peter Malinauskas announced a Data Centre Strategy proposing a new Data Centre and AI Infrastructure Act and a more coordinated, state-led approval pathway involving the state's Coordinator-General. The bill has not yet been introduced in final form, so it would be premature to say precisely how far it will displace ordinary council assessment, consultation and appeal rights, but the strategy itself asks the public to welcome a new approval pathway before the operative legislation exists, which is its own kind of accountability question.

IREN has since signed for an 800-megawatt facility at Bundey in the state's mid-north, and Firmus Technologies has flagged sites near Tailem Bend and Port Augusta. When water concerns were raised, the Premier characterised them, as reported by InDaily, as “outdated and in some cases exaggerated,” pointing to newer, less water-intensive cooling technology. There is real substance to that: the newest facilities increasingly use direct-liquid and hybrid cooling systems that can substantially reduce on-site water consumption in some configurations, though measured consumption once a facility is actually operating still matters more than a developer's efficiency promise made before construction. Even so, it sits awkwardly next to South Australia's own description of itself, in its own government's water security documents, as the driest state on the driest inhabited continent.

Adelaide's ten reservoirs, which hold roughly a year's supply when full, fell to 44 per cent capacity in early 2025 after the city's driest year since 2006, forcing the Adelaide Desalination Plant up to near-full production to cover the shortfall. A state whose own water security rests on that kind of margin is not an obviously safe place to be told, without independent verification, that a large new category of industrial water user is nothing to worry about.

The state's energy picture resists a simple verdict in either direction. In the final quarter of 2025, South Australia recorded exceptionally low average wholesale prices, a real demonstration of the value abundant renewable generation and storage can provide. In the first quarter of 2026, though, it was the only region in the National Electricity Market to record a year-on-year price increase, averaging $88 a megawatt hour, up 33 per cent, with a single weather-related event on 26 January contributing around $26 of that average on its own, according to AEMO's own quarterly reporting. The lesson isn't that renewable generation has either solved or caused every price problem in the state. It's that a grid carrying large amounts of variable generation can swing between long stretches of very low prices and short, concentrated periods of real scarcity.

Adding hundreds or thousands of new megawatts of steady industrial demand to that kind of system is not necessarily a problem, but it is a genuine planning question, one about firming capacity and peak conditions rather than annual averages, and one more reason the same standard of evidence should apply to reassurance as to alarm.

Victoria offers another illustration of how consultation can formally exist while arriving late in the life of a project. At Barry Road in Campbellfield, on part of the former Ford Broadmeadows Assembly Plant, where the last Australian-built Falcon rolled off the line in October 2016, the Singapore-based developer Zerra DC has lodged an application for a six-building data-centre campus. Victoria's own planning register confirms the application was received on 25 April 2026 and placed on public notice on 8 July 2026, under the state's Development Facilitation Program, an expedited pathway in which the Minister for Planning, not the local council, is the responsible authority.

Greens MLC Anasina Gray-Barberio has raised the proposal in State Parliament, according to reporting, arguing that communities in Melbourne's north and west should not become dumping grounds for developments primarily benefiting large technology corporations, and calling on the Planning Minister to meet with the local community group that has organised against the project. The project has not yet been decided, so it would be unfair to say residents were consulted only after a final decision was made. But by the time formal public notice actually began, more than two months after the application was received, the site had already been selected, a detailed six-building proposal designed, and the application placed into that expedited, state-led process. Consultation, in other words, can be technically available while the decision that matters most, whether this is the right use for this land, has already substantially hardened before the public is invited to comment on it.

None of what is happening here is unique to Australia, and it is worth being honest about that too, because the alternative, treating this as some peculiarly Australian failure of foresight, would not survive contact with what has already happened overseas.

Ireland's data centres consumed 23 per cent of the country's metered electricity in 2025, up from 22 per cent in 2024 and just 5 per cent a decade earlier, concentrated overwhelmingly around Dublin. From 2021 the country's grid operator effectively stopped connecting new data centres in the Dublin area, a de facto moratorium that lasted until the regulator replaced it, with a decision paper published in December 2025, with a conditional connection policy: new facilities must now provide additional dispatchable generation or storage of their own, and, according to trade reporting on the same decision, source at least eighty per cent of their annual demand from new renewable generation. It is a concrete example of a country moving from an emergency freeze to a workable ongoing rule, a sequence Australia could still choose to skip by legislating the rule before a freeze becomes necessary.

The Netherlands took a blunter path. Amsterdam banned new data centres and expansions in the municipality outright at the end of 2023, extended that ban again in 2025 to run until at least 2030, and the national government has separately barred new hyperscale facilities above seventy megawatts across most of the country since 2022.

And in the United States, a Gallup poll fielded in March 2026 found that seventy-one per cent of Americans oppose building an AI data centre in their own area, more opposition than Americans express toward a local nuclear power plant, with water and energy use the two most common reasons opponents give. These are three different legal systems reaching for three different tools, moratoriums, conditional connection policies, planning bans, but the underlying shape is the same everywhere: infrastructure arriving faster than the public conversation about who bears its cost.

None of these overseas comparisons prove that Ireland, the Netherlands and Australia face identical conditions, or that each should adopt identical laws. Their value is more modest than that: they show what happens when concentrated demand grows faster than the rules governing its connection. Australia still has the chance to learn from those experiences before its own local constraints harden into a national fait accompli.

The residents of Lane Cove West and Ryde who found out about a 1,400-kilolitre-a-day neighbour after the approval, not before, are not obstacles to progress; they are people entitled to have been asked. The public servants drafting the Expectations framework and weighing genuine, competing goods, investment, jobs, water security, energy reliability, are not villains for finding the balance hard; that is what governing well actually requires. The engineers and tradespeople building these facilities, and the researchers whose work depends on the computing capacity inside them, are not extracting resources for no reason; they are doing legitimate work in an industry the whole country stands to gain something from, if it is done honestly. None of them is reducible to a side in an argument, because none of them is only a data point, an approval, or a talking point. Every one of them bears the image of God, and that has to be true before any of the harder claims in this piece are.

There is a wider question sitting underneath all of this, and it doesn't have a tidy answer. For most of Australia's history, national investment and national pride ran through things you could see and touch: mines, farms, factories, bridges, ports. AI and the data centres that run it are not that kind of thing, and it is fair to ask whether this shift is something the country chose, weighed and decided together, or whether it is simply where global capital and computing power happen to be moving right now, with governments largely following rather than leading.

I don't think there is an obvious villain in that question. Manufacturing, agriculture and engineering haven't vanished from the national conversation, and some of the same investment appetite chasing AI capability is also flowing into critical minerals processing, defence manufacturing and renewable-energy supply chains. But it is still worth asking plainly, for the sake of the next generation rather than in the abstract: is this a proportionate share of the country's energy, water, land and government attention to be spending on running the technology, relative to what is spent helping people build, grow and make things, and who actually gets to weigh in on that balance before it is set?

Few Australian sites make that question more concrete than Broadmeadows, where land once associated with assembling Ford Falcons is now proposed for a six-building data-centre campus. Nostalgia alone cannot settle good land use, and a data centre can support legitimate and valuable work. But the location sharpens the question worth asking: was this transition something the community weighed against other possible uses, or was it invited to comment only once the preferred economic direction had already been set?

There is a principle underneath all of this that isn't complicated, even if applying it consistently is. Private benefit should not depend on undisclosed public subsidy. That doesn't mean data centres should receive no shared infrastructure at all; some of what they need genuinely serves the wider public too. IREN's own figures for its Bundey facility, for instance, point to several hundred construction jobs, a smaller number of ongoing skilled roles, and capital investment in the billions of dollars, real benefits this critique isn't asking anyone to dismiss. It means households, ratepayers and existing businesses shouldn't be quietly left to absorb project-specific costs that would never have arisen without this one development, disclosed, if at all, as annual efficiency promises made before a shovel goes into the ground rather than as measured maximum demand once a facility is actually running.

Electricity-system reliability is tested during peaks and scarcity events, not by annual averages, and water restrictions bite during droughts and heatwaves, not an ordinary week in spring; approvals should disclose peak demand and drought-stage operating arrangements, not just annual averages. It's also worth distinguishing three different things that get called public involvement. Notification tells people what is proposed. Consultation lets them comment on it. Participation lets them help shape the options before one has quietly become the default. Good stewardship should aim for the third, especially wherever a development will place unusually large, long-lived demands on shared water, electricity and public infrastructure.

So here are the questions I am left asking, rather than declaring. Who is actually being asked to bear the cost of this decade's AI build-out, the resident near the substation, the household paying a higher wholesale price, the town whose water utility now has to plan around a customer it never expected, and were any of them asked before the decision was made rather than after?

What would it look like for Australia to treat this the way a faithful steward treats an inheritance, not something to extract from as fast as possible before someone else does, but something to be handed on in at least as good a condition as it was received?

And given that the Senate's Environment and Communications References Committee is taking public submissions on exactly these questions, with its deadline already extended once to 1 September 2026, open to any resident, ratepayer or professional who wants to put something on the record, not just to industry lobbyists and government departments, what would it cost any of us to actually write one?

The committee's own guidance on preparing a submission, including an Easy English version for anyone with accessibility needs, is available through the Australian Parliament's website, and the secretariat can be contacted directly for anyone who needs support to take part.

Two questions from this piece alone would make a submission worth reading: what capacity threshold will trigger the new national standards, and will facilities already approved or under construction be covered by them; and how will real-time water and peak electricity consumption actually be verified and enforced once a facility is operating, rather than merely promised beforehand?

“Now it is required that those who have been given a trust must prove faithful.”

— 1 Corinthians 4:2

Thank you for reading.

God bless.

Mark Neugebauer