A Second Bite at the Golden Apple

Australia's data centre opportunity — the NOOPS white paper
12 August 2026 John Allsopp & Mark Pesce analysis FREE
data-centresaustraliaenergypolicyinfrastructure

Introduction

Data centres have become the great divisive cultural issue of the mid 2020s.

On one hand, they represent the most concrete manifestation of technological progress, bringing together enormous amounts of resources to produce effectively unlimited cognition on demand. Electrons go in one side cognition comes out the other.

On the other hand, data centres typify all of the most arrogant and excessive demands that technology would place upon us. Often in the hands of actors whose motives are - if we can put it succinctly - impure.

Between these two extremes there lies a middle path. This document represents our own view of that path.

In part one, we lay out the already deep integration of data centres into our work practices.

Next we address directly the most significant objections to data centres.

Finally, we take a look at the enormous opportunity open to Australia - if it can walk the middle path and get this right.

As with all things NOOPS, we have a strong point of view. We sincerely believe that artificial intelligence is the biggest technological transformation of our lifetime.

For that reason, it should be treated seriously and forthrightly.

We do not claim to have all of the answers, nor do we possess a crystal ball to foresee all of the potential problems that might come in the future. We can only offer what we believe are well-grounded observations, and sensible suggestions.

We believe in Australia, and want to see the nation benefit from a generational opportunity.

We look forward to your comments and critiques.

John Allsopp & Mark Pesce

Part One: Where the Work Happens

One way to understand data centres is to recall the day the world should have ended.

On the 25th of March 2020, Prime Minister Scott Morrison came onto every television screen in Australia - and began sending everyone home, as the only known effective method for preventing the spread of COVID-19.1

Frightened, Australians did as they were told. But leaving desks and workplaces behind carried another fear: suddenly 'dead in the water', would we watch our economy collapse beyond repair?

If people can't work, they can't produce. If they can't produce, everything falls apart. That was the logic and it made so much sense that everyone believed it unquestioningly.

The right-of-centre government of Australia pivoted into socialism, creating JobKeeper - a payment of $1,500 a fortnight for every worker stood down - to keep money and commerce coursing through a foundering economy.2

Everyone predicted financial wreckage. Everyone expected it.

It never came.

Not only did the world not end, it found its footing - after a few rough weeks. People who had done office work all of their careers found that office work could happen anywhere they had a 'good enough' computer and internet connection.

Australia's National Broadband Network justified its $51 billion investment almost overnight, saving the nation that amount and more in lost productivity.3

Businesses, freed from the infrastructure costs associated with running business office spaces, found another way to boost profits.

Australia spent the better part of two years popping in and out of 'lockdown', Melbourne and Sydney being particularly affected.4

At the end of 2021, as restrictions lifted, work had changed completely, becoming 'hybrid', 'work from anywhere' and 'flexible'. Workers loved their newfound capacity to prioritise their activities. Bosses had mixed feelings, but overall output held up - and in many sectors rose.5

How can work happen without a workplace, without an office?

Pandemic restrictions did not change the nature of work; they simply forced us into an awareness of how much work had already changed.

Over the previous decades, all of the tools of work had migrated from physical objects - pen, paper, typewriters and filing cabinets - to their nearest digital equivalents: PCs, networks, printers, and servers. The organisation's output became a set of digital artefacts.

As the Web matured, it became the ideal interface for these new tools. From the late 1990s a rapid development in 'web-based applications' meant that almost all of these digital artefacts could be created, read, updated and deleted through a web browser, which could be running on almost any computer anywhere in the world.

That's the beginnings of 'Software as a Service', or SaaS. Companies like Google, SAP and Salesforce built their businesses on SaaS product offerings; incumbents like Microsoft and Adobe pivoted to offer their own SaaS services.6

By March 2020, most businesses used subscription-based software, delivered as a service over the web. Working from a corner of the dinner table during a COVID-19 lockdown meant using exactly the same applications used in the office, in exactly the same way. Work didn't have to change at all.

Business coordination came through these SaaS applications: Zoom, Slack, Teams, Google Meet, and so forth. Every one of these applications found their universal use cases during the pandemic. They - plus the plethora of other SaaS applications used by businesses - kept businesses running smoothly despite long-term closure of the 'place of business'.

That was only possible because the place of business wasn't the office.

The place of business had moved to the data centre.

Every SaaS app, and the vast majority of mobile apps, runs inside a data centre. Without a data centre there is no Microsoft 365, no Zoom, no Salesforce, and no Google Drive. All of these applications have been designed to run in data centres.7

The data centre is so fundamental to the design of these apps that there is no way to bring them 'in house', unless the business itself builds its own data centre to host those apps.

Work is no longer a story about offices. Work happens almost entirely within and because of data centres.

The final proof of this nearly invisible shift came in early 2026, when US Census Bureau construction data showed that spending on data centre buildouts had surpassed office buildouts for the first time in history.8

Australia follows a similar trajectory: Sydney and Melbourne have seen delayed office building projects, while data centre construction has grown at 65 per cent a year for three years running - reaching $2.6 billion in 2024-25, one dollar in every eight of non-residential building investment.9

The place of work has shifted to the data centre; investment reflects that shift.

Part Two: Do It Right or Do It Elsewhere

In July 2026, Prime Minister Anthony Albanese delivered his government's policy statement about AI and data centres.10

It made news around the world.

Against the populist tendency to rail against data centres as the condensation of all of the ills of technological society, Albanese presented a series of concrete proposals - on their way to becoming laws and regulations - that would effectively answer the most common concerns about their rapid proliferation, and, in the process, deprive critics of their most persuasive arguments.

Critics tend to both dramatically oversimplify and catastrophise the issues of data centres. Dealt with reasonably and rationally - the way forward, according to the Federal Government of Australia - they offer a path toward a 'best of both worlds' scenario; where data centres bring both growth and resilience.

The critiques break down, broadly, into two classes: environmental impacts and social impacts. Let's cover the environmental impacts first.

Energy

Data centres consume huge amounts of electricity; effectively, a data centre is a factory where electrons enter on one side, and computation exits the other. Electricity is the raw material; without it, the data centre has nothing to work with.

Until about a decade ago, electricity utilisation had broadly plateaued in the developed economies, largely due to efficiencies in lighting - LEDs are an order of magnitude more energy efficient than incandescents - and improved designs of electrical appliances.11

Since then, the developed world has embarked on a decades-long project of 'electrification', where electricity is supplanting other fuel sources, such as petroleum and natural gas, as the main source of energy consumed at home and in businesses.12

Though still in a relatively early stage, vehicle electrification has significantly moderated the demand for petroleum, paired with a greater demand for electricity for vehicle charging. In addition, home appliances such as stoves, ovens and furnaces are being replaced with their electric equivalents - induction ranges, electric ovens and heat pumps.13

All of this adds to the demand on the grid, although this growth in demand could be confidently forecast and planned for by electricity generators and distributors.

No one had planned for the explosive growth in data centres.

Capital spending on data centres has roughly tripled since 2022, with the hyperscale operators alone budgeting close to two-thirds of a trillion US dollars for 2026 - and no slowdown in sight.14 Global data centre electricity consumption has grown from roughly 290 Terawatt-hours (TWh) in 2022 to around 565 TWh projected for 202615 - growing at nearly a fifth a year, a pace the International Energy Agency expects to continue, reaching close to 950 TWh by 2030: a doubling in six years.16

It takes years to build out electricity infrastructure; the data centre explosion has unfolded on a timescale of months. That mismatch creates a massive overdemand situation for electricity, where consumers, businesses and data centres all compete for the same limited supply of electrons.17

Data centres are primarily gated by access to electricity, so any proposal to build a data centre has to factor in the ability of the grid to deliver that electricity reliably, sustainably, and without placing undue burdens on other consumers of electricity.

The United States of America represents a case study in how to do this incorrectly.

Throughout that nation, electricity bills are skyrocketing, because consumers are shouldering new and substantial buildout and maintenance costs for generating and delivering electricity. In the PJM market - thirteen states, from Illinois to North Carolina - capacity pricing jumped more than eight times over the year before, with data centres identified as the primary driver, feeding through to household bill increases of fifteen dollars a month and more in parts of Ohio and Maryland.18

Consumers consider themselves as being publicly taxed for private gain, and that has produced a very strong political response: some jurisdictions seek to shut down public debate over new data centre proposals, while in others, incumbent office-holders have lost pre-selection for being seen as insufficiently sensitive to these concerns.19

Australia seeks to step around this issue before it becomes a problem, with one requirement at the heart of the announced standards: that every data centre proposal come with a fully costed plan to generate the electricity to be used by that data centre through renewable means. Those standards are due to become law in early 2027, after passing through National Cabinet.10 Roughly ninety data centre projects currently sit in Australia's approvals pipeline;20 the standards should be applied to them now, as conditions of approval, rather than waving them through ahead of the law they'll be grandfathered out of.

More than a third of Australia's single-dwelling homes have solar PV systems installed; the grid generally has a surfeit of electrons on bright, sunny days.21 The technologies of battery-grid storage are well developed and deployed throughout the nation.22

The infrastructure is, in short, well understood and already available.

Given that a data centre is relatively price-inelastic when it comes to the cost of electricity - they will pay any reasonable amount for their input electrons - the Australian government has outlined how a data centre buildout can provide the capital inputs needed not just for the renewable electrification of those data centres, but for the rest of the Australian economy.23 This isn't a thought experiment: Ireland's energy regulator now requires new data centres to source at least 80% of their demand from additional renewable generation,24 and Amazon has already signed agreements with nine Australian renewable projects specifically to power its facilities here.25

That's framed as the 'price to play' in Australia. Given the cheap-and-dropping cost of renewable generation, the numbers work out well for a data centre operator who gets an inexpensive, reliable and inexhaustible supply of electrons, while never having to worry about a war in the Middle East or a sudden supply shock jeopardising operations.26

It also means that data centres are better than 'non-polluting'; the data centre buildout leaves Australia cleaner and more resilient than beforehand.

Water

The conversion of electrons into computation generates waste heat. Every time every transistor on every chip turns on or off, a tiny bit of heat gets released. Multiplied across the quadrillions of transistors within a single data centre, a tiny bit of heat becomes a blast furnace's worth of heat energy that has to be transported away from the chips quickly, before they begin to melt.

Data centres have historically cooled themselves using air passed through a heat exchanger with cool water running through it. The warm air is passed through another heat exchanger; the air gets cooled down while the water - at both entrance and exit - gets warmer.27

The water then gets passed through a cooling tower where the water loses its heat to the air in evaporation.27 In a 100 megawatt data centre - a large facility of the type proposed to be built in Australia - keeping the facility at safe operating temperatures would evaporate somewhere between 5 and 20 million litres of water, every single day - roughly enough to meet the demands of a city of 40,000 residents.28 (There are only 42 municipalities in Australia with populations larger than 40,000.)

Those figures describe the worst case - but it's also common enough that communities in the United States have experienced competition for water resources because of data centre buildouts.29

As the driest continent, Australia does not have the capacity to site facilities that consume the equivalent water supply of a large regional centre.30 Evaporative cooling cannot work in Australia.

Fortunately, there is a strong alternative.

At the 2026 Consumer Electronics Show, NVIDIA CEO Jensen Huang introduced the new 'Vera Rubin' series of AI accelerators. The full 'rack', the size of a room, contained hundreds of components that had been fully redesigned around a technology now sweeping the industry: Direct Liquid Cooling (DLC).31

The system runs a continuous, closed-loop supply of warm water - around 43ºC - over the components, carrying the heat away, where it is ejected from the data centre through banks of radiators. Because the returning water runs hotter than the outside air on all but the most extreme days, the radiators alone can shed the heat, and the same water returns to the chips to collect more. No water is lost to evaporation.

In fully dry configurations (a design Microsoft announced in late 2024 as its standard for new builds)32 water consumption after the initial fill approaches zero: top-up amounts measured in thousands of litres a year, about the same as a small garden pond. One caveat for our climate: radiators shed heat less readily on 45-degree days, so Australian facilities must be specified for summer extremes - a solvable design parameter, and a regulatable one.

Albanese promised that regulation would 'minimise water use' by data centres.10

Direct Liquid Cooling with dry heat rejection represents the current state-of-the-art for minimal water use.

Regulation should require new data centre designs to employ these systems - not merely encourage them. Today only around a fifth of operators use direct liquid cooling;33 most facilities described as 'closed-loop' still include an evaporative stage. The technology standard is exactly the difference between the two columns below - and because dry cooling replaces evaporative towers and chillers with slow-turning radiator fans, the noisiest cooling plant disappears along with the water.34

This chart lays out the stark difference between doing it, and doing it right:

Metric Worst-Case (Evaporative Towers) Best-Case (Fully Dry Closed-Loop DLC)
Cooling Method Evaporative cooling towers (water continuously evaporates into the air to dissipate heat).27 Direct-to-chip closed loops paired with dry air coolers (radiators).31
Annual Water Consumption ~2.6 million gallons per MW (~9.8 million litres/MW).28 Near zero after the initial charge.
Daily Footprint (100 MW Facility) 1.5 million to 5 million gallons/day (~5.7M to 19M litres/day).28 Negligible (top-up measured in thousands of litres per year).32
Equivalent Impact Equal to the daily water demand of a city of 30,000 to 50,000 residents.28 Equivalent to a household filling a small garden pond once a year.

Best-case column describes fully dry direct-to-chip designs (per Microsoft's zero-water design for new builds, announced December 2024);32 most currently deployed "closed-loop" systems retain a hybrid evaporative stage — which is precisely why the standard matters.

Property

The first and most substantial of the social impacts of data centres - and, specifically data centres used either to train or inference generative AI - concerns intellectual property rights. Training of these AI models requires feeding them trillions of words. While publicly available sources on the Internet comprise a significant amount of that data, training has extended to anything and everything that can be 'scraped' up by an AI model trainer looking for content, any content. Some of this content is protected by copyright.35

The question, so far not cleanly settled by any court anywhere in the world, is whether using material under copyright for training represents a violation of that copyright - by definition, a material economic damage to the copyright holder. The closest any court has come is a split: a US federal court found that training on lawfully acquired books could be fair use, while building a library of pirated copies was not - a distinction that produced the largest copyright settlement in history rather than a clean precedent.36

DISCLOSURE: Both Mark Pesce and John Allsopp are registered claimants in the Bartz v. Anthropic class action settlement (referenced above), under which Anthropic will compensate authors whose written works were used for training purposes.

The Australian government's position, as outlined in Anthony Albanese's policy statement, is that access to Australian intellectual property for model training must be based on explicit consent and fair compensation.37 The first of those is burdensome, but conceivable. The second - what constitutes 'fair compensation' - will be contested throughout the global legal system for many years to come.

As this is an active and strongly contested area of both policy and case law, NOOPS can only point to the open issues around property rights.38 We believe this is a solvable problem, though any solution will almost inevitably produce a WIPO-class international framework. Individual nations (with the possible exception of the United States) lack sufficient standing to force their own solutions on other parties.

Can Australia insist on a 'good enough' solution that will in all likelihood be 'routed around' by other jurisdictions? That's less a question of whether this is a hill worth dying on, than a question of whether it's actually any sort of hill at all.

Our next white paper will explore this topic in greater detail.

Work

The final issue, the one that feels most existential, concerns the future of work, jobs and the economy. The data here can be read multiple ways, and an unnerving string of statements from both AI industrialists and academic economists has produced broad concern that AI-in-data-centres will simply absorb all available work, resulting in mass unemployment.

Here are the basic facts:

2025 may be remembered as 'the year copywriters disappeared'. Not all copywriters, not in every instance, but in cases where 'good enough' copy would suffice, that work has been automated by language models.41

Similar impacts have been repeatedly predicted for a range of professional practices, including call centres, software engineering and radiology. Call centres will likely move to nearly full automation. There is some indication of softening in software engineering, particularly in wage growth. Demand for radiologists, though, remains high42 and may be subject to a Jevons paradox: the advent of AI diagnostic tools has increased demand for radiologists to consult on the most difficult scans.43

This indicates that employment impacts may be very uneven. Some jobs will vanish, never to return: call centres. Some will be transformed: software engineering. Others will be amplified: radiologists.

However, even if a job or role survives its encounter with artificial intelligence, it will be transformed by that encounter. As AI enters organisational workflows, it changes the why as much as (often more than) the how.

Nearly everyone's job will change. Many will change significantly. A few will become almost unrecognisable to their antecedents.

That will feel existentially terrifying.

Prime Minister Albanese directly rejected the framing of AI as a 'jobs apocalypse'. "We should not treat AI as a threat to good jobs - we must use it as an instrument to help create them."44

Flipping the script points Australia toward a future where disruptions are both expected and welcomed as opportunities to create new roles in an increasingly automated business sector.

Albanese made the argument that technology shifts work rather than erasing it with a story from his own life: his first job out of school was standing in a Commonwealth Bank branch, convincing customers to abandon their passbooks and trust a strange new hole in the wall called an ATM.44 The tellers' work changed beyond recognition; the banking remained.

There's an example that is even more instructive: the disappearance of the entire class of secretarial labour following the introduction of the personal computer.

In 1970 roughly one in six workers did secretarial and clerical labour: typing, filing, communications and scheduling. Today it's closer to one in ten, while the role has been rebadged as 'executive assistant', 'administrative assistant' and so forth.45

The how of that role has completely changed in the last five decades, because of technology, but the need for the role remains - the why - even as we get 'autonomous agents' that will handle a lot of the time-consuming detail work.

This pattern will be repeated throughout the workforce as AI automation takes hold.

All of our roles will change - including analysts. ;)

Doing It Right

In each of these areas of central concern: power, water, property and work - we have complete agency. We can decide to power data centres with electrons generated renewably - or with gas turbines. We can mandate data centres use closed-loop cooling systems, or deplete our limited aquifers. We can make sensible decisions about the permission, compensation and notification rights for intellectual property. And we can use 'social licence' to police businesses that would use AI as a weapon against labour.

We can abandon neither our agency nor our responsibilities when it comes to the construction of data centres. They represent, in microcosm, the choices we are making for our future.

Australia will be the first country in the world to bring these issues into a single, national framework.10

We can decide to 'do it right' - or prepare ourselves to pay a different price.

Part Three: Another Bite of the Golden Apple

Australia has been here before.

A generation ago the greatest resources boom in our history arrived on our doorstep. China industrialised, and it did so out of Australian iron ore and Australian gas.46 We dug it up, we shipped it out, and we let the proceeds slip through our fingers.

Norway faced the same moment with North Sea oil and made a different choice. It taxed petroleum profits at seventy-eight per cent and banked the money, by law, into a sovereign fund.47 That fund is now worth about NOK 21.3 trillion — around two trillion US dollars, for five and a half million people.48

Qatar is a closer comparison, because Qatar's boom was gas, like ours. Australia and Qatar each export roughly 80 million tonnes of LNG a year. In 2023, Qatar's government collected about A$56 billion from its share. Australian governments, all levels combined, collected about A$11 billion — and more than half of our LNG production pays no royalties at all.49 Qatar's gas money went into the Qatar Investment Authority, now managing somewhere between US$500 and 600 billion.50 Ours went — where, exactly? The Grattan Institute did the accounting: the Commonwealth earned about $200 billion from the boom and spent more than ninety per cent of it as it came in.51 There is a fair defence of some of this — Norway owns its petroleum industry outright, we chose royalties and private investment decades earlier — but the fund balances speak for themselves.

We got the century's easiest economic decision and turned it into a decade of tax cuts. Everyone agreed nothing like it would come around again.

It just did.

This time the resource isn't under the ground. It's the sun, the wind, the land, the stable institutions — and the timing.

Parts One and Two dealt with what data centres are and what we have to get right if we want to enjoy our great luck, instead of being burned by it.

This part is about the money: where it flows now, where it flows if we build, and where it flows if we don't.

The case against, put properly, is an accounting argument. In the March quarter of 2026, data centre equipment pushed Australian private capital spending to records — about 17 per cent of all private investment.52 Nearly all of that equipment is imported. The server racks were enough to tip Australia into its first trade deficit in ten years.53 HSBC's chief economist reckons imports offset close to 85 per cent of the GDP contribution of the investment surge;54 Westpac puts the leakage nearer half.55 From there the conclusion seems to write itself: we're buying hardware we don't make, the money leaves the moment we spend it, why are we cheering?

The numbers are right. Roughly 60 per cent of what an AI data centre costs is the servers,56 they come from overseas, and they'll keep coming from overseas every four to six years as the hardware is refreshed.57 Nobody should pretend otherwise, and this paper doesn't.

But an import bill only means something next to the alternative. So look at the other column in the ledger.

Australia already imports its digital life. The ABS has tracked it: digital services imports grew from $119 million in 2009–10 to $18.5 billion in 2023–24. Over those fifteen years we imported $87 billion of digital services and exported $5 billion. Seventeen dollars out for every dollar in — and that was before the AI era began in earnest.58 Gartner expects Australian organisations to spend more than A$33.6 billion on public cloud in 2026, up 17.9 per cent in a year, having grown 18.9 per cent the year before.59 Nobody publishes an Australian forecast for 2030, but the arithmetic isn't hard: hold that growth rate and the cloud bill alone passes $60 billion a year within four years. Add AI, which every business in the country is now being sold, and the that growth accelerates.

The GPUs are foreign either way, but if we build, while we import the hardware, we keep the construction, the land, the energy system, the operations, the grid assets and the option to sell. If we rent, we import the hardware anyway: embedded in someone else's hourly rate, alongside their electricity, their staff, their buildings and their margin. We keep nothing. One path has a large import bill attached to a domestic asset. The other is a subscription, compounding at close to twenty per cent a year, forever. Trade deficits for server racks will run for a few quarters. The rent runs for decades.

And the rent is only the defensive half of the argument.

Those flows can run the other way. Compute is a tradeable service - on the way to becoming a commodity, and Australia sits next to the largest concentration of demand on earth. Austrade is already pitching Australia as the data centre and AI services hub for Asia: the cables land here, the time zones work, and Australia holds Tier-1 status under US chip export rules — clearance most of the region doesn't have.60 This isn't a theoretical capability. AirTrunk was built from Sydney into the Asia-Pacific's hyperscale platform and sold for A$24 billion, the largest data centre transaction anywhere.61 Fifteen years of digital services exports totalling $5 billion isn't a ceiling. It's a measure of how much room the flow has to reverse.

The domestic dividend follows the infrastructure too. The e61 Institute, no cheer squad for the sector, makes the point that matters: the direct GDP effect of the build is modest, and the real prize is whether cheap local compute triggers the follow-on investment in software, R&D, and AI adoption across every industry - where productivity gains actually live.62 That dividend lands in whichever country's firms sit closest to the infrastructure. At the moment it is landing offshore. The same honesty applies to jobs: a hyperscale campus employs hundreds in construction and dozens in operation, and developer projections overstate the local effect by 300%.63 The employment case is the electricians, the construction trades, and the industries that grow up around cheap rewewable power — not the badge count in the server hall.

Then there is the question of owning any of it. Canada has put CAD$2 billion into sovereign compute.64 Britain has a national research resource and a six-gigawatt target.65 France has committed a €100 billion envelope.66 Australia, the second-largest destination for data centre investment in the world,67 has an Office of AI.68 Norway's deepest lesson was never the tax rate; it was that the state held equity in the boom.69 A public stake in training capacity for Australian research, Australian government, and Australian models describes the difference between hosting the boom and owning a piece of it.

All of which would matter less if the window were open indefinitely. It isn't.

The buyers are at the door now, with deadlines attached. OpenAI has anchored NEXTDC's $7 billion, 612-megawatt campus at Eastern Creek, first phase targeted for 2027.70 Anthropic, reported to be after 1.4 gigawatts of Australian training capacity71 (about the size of the country's entire existing data centre fleet72) is, per the AFR, telling developers it will buy any capacity deliverable by mid-2027, ahead of its listing. The hyperscalers will spend over two-thirds of a trillion US dollars this year14 and are deciding now where the next decade of capacity goes. When they can't get grid connections fast enough they don't wait: in Texas, Anthropic's newest campus comes with a 1.6-gigawatt gas plant built into the financing.73 Capital this impatient goes wherever it can plug in. It is choosing between Australia, the American gas belt and its Gulf of Mexico in roughly a two-year window. It will not reopen the question because we weren't ready.

A seller's market is when you set terms, and Australia has already shown it can: when we introduced mandatory power-consumption reporting for data centres, the United States copied us.74 The terms worth setting are the ones outlined in Part Two, and the Australian Federal Government has announced most of them.23 What we need is speed, a bit of nerve - and the discipline not to fumble a windfall for the second time in living memory. The one certainty is that the AI transformation of Australian work arrives either way. Not building doesn't stop it. It just means we go through the same upheaval as everyone else while paying rent on the infrastructure it runs on.

Same land. Same sun. Same wind. Same stable, slightly boring, trustworthy country. Except this resource doesn't run out, the customer can't ship the mine home, and what we build to serve those data centres powers everything else we do for a century.

Last time, we ate the apple and spat out the core. The tree is fruiting again.

Footnotes

  1. Department of the Prime Minister and Cabinet, transcript of Prime Minister Scott Morrison's press conference, 25 March 2020; and ABC News, “Read Scott Morrison's full statement on the new national coronavirus restrictions”, 25 March 2020. The national restrictions directed Australians to stay home except for specified purposes and to work from home where possible.
  2. Australian Treasury, JobKeeper Payment: Supporting businesses to retain jobs, 2020. The programme paid eligible employers $1,500 per fortnight for each eligible employee.
  3. NBN Co, Corporate Plan 2021, 2020, records the network's peak-funding estimate and its role in supporting work, education and telehealth during the pandemic. NBN Co, “Coronavirus (COVID-19) and nbn: latest updates, working from home tips and FAQs”, documents the March 2020 provision of up to 40 per cent additional capacity as home data demand rose.
  4. Reserve Bank of Australia, The COVID-19 Pandemic: 2020 to 2021, traces the repeated imposition and easing of restrictions. ABC News records Melbourne reaching 200 cumulative lockdown days by August 2021 in “Melbourne marks 200 days of COVID-19 lockdowns since the pandemic began”, 19 August 2021, and the end of its final 2021 lockdown in “Melbourne bids farewell to COVID-19 lockdown and the controversial curfew”, 21 October 2021.
  5. Australian Productivity Commission, Working from home, 2021. The Commission documents the pandemic-driven shift to home working, the likely persistence of hybrid work, and evidence that many employees can work effectively outside the office.
  6. Peter Mell and Timothy Grance, US National Institute of Standards and Technology, The NIST Definition of Cloud Computing, Special Publication 800-145, 2011. NIST defines software as a service as using a provider's applications running on cloud infrastructure, accessed through client interfaces such as a web browser.
  7. NIST describes cloud computing as on-demand access to a shared pool of configurable computing resources in The NIST Definition of Cloud Computing. Microsoft provides an example of the physical global infrastructure behind such services in Azure global infrastructure.
  8. US Census Bureau, Value of Construction Put in Place at a Glance; and Building Design+Construction, “Data center construction spending outpaces office market for the first time”, 3 February 2026. The latter reports that the seasonally adjusted annual rate for data-centre construction first exceeded office construction in December 2025.
  9. Infrastructure NSW, Data centres in NSW: Consultation paper, 2026. It reports that data-centre construction averaged 65 per cent annual growth over the preceding three years, reached $2.6 billion in 2024–25, and accounted for about 12 per cent of NSW non-residential construction.
  10. Prime Minister of Australia, “AI in Australia's interests”, 15 July 2026; and Department of the Prime Minister and Cabinet, transcript of the Prime Minister's National Press Club address, 15 July 2026. The announcement covers legislation requiring new data centres to secure new generation, pay their grid-connection costs and meet water-efficiency standards; a national AI framework; an Office of AI; and protections for Australian creative work.
  11. International Energy Agency, Electricity 2026: Demand, 2026, attributes the long period of flat electricity demand in advanced economies partly to efficiency improvements and economic restructuring. For lighting specifically, the US Department of Energy reports that LED products use at least 75 per cent less energy than incandescent lighting in Energy Savings Forecast of Solid-State Lighting in General Illumination Applications, 2021.
  12. International Energy Agency, Electricity 2026: Demand, 2026. The IEA identifies electrification as a structural driver of renewed electricity-demand growth in advanced economies.
  13. International Energy Agency, Electricity 2026: Demand, 2026. The IEA identifies electric vehicles, air conditioning and heat pumps among the technologies increasing electricity consumption.
  14. Apollo Chief Economist, The Daily Spark: Hyperscaler capex, 22 February 2026. The note gives a 2026 hyperscaler capital-expenditure estimate of US$646 billion.
  15. Gartner, “Data Center Electricity Demand to Grow 26% in 2026”, 10 June 2026. Gartner forecasts 565 TWh of data-centre electricity demand in 2026.
  16. International Energy Agency, Energy demand from AI, 2025. The IEA estimates that data centres used about 415 TWh in 2024 and projects about 945 TWh by 2030. Its earlier sector overview put 2022 data-centre electricity use at 240–340 TWh: Data centres and data transmission networks.
  17. International Energy Agency, Energy demand from AI, 2025. The IEA notes that data centres can be operational within two to three years while the energy infrastructure needed to supply them often has longer lead times.
  18. PJM Interconnection, “PJM Auction Procures 134,311 MW of Generation Resources; Supply Responds to Price Signal”, 22 July 2025. PJM reports a record auction clearing price of US$329.17/MW-day and describes data-centre growth as a contributor to tightening supply and demand.
  19. Maryland Office of People's Counsel, Summer 2025 Electric Rates Factsheet, 2025, reports capacity-price increases of more than 800 per cent and estimated monthly residential impacts of $14–$18 for several utilities. Ohio Consumers' Counsel, Quick facts: Data centers in Ohio, discusses projected data-centre-driven bill impacts. Utah reporting connects data-centre disputes to incumbent election losses: Utah Public Radio, “Box Elder data center campus claims two casualties: a pair of incumbent commissioners”, 24 June 2026; KSL, “How a data center helped topple one of Utah's most powerful lawmakers”, 26 June 2026.
  20. Data Centres Australia and DC Byte, Australian Data Centre Forecast Report, Issue 1, April 2026. The report's pipeline inventory counts 90 proposed data centres as at 31 March 2026.
  21. Clean Energy Regulator, “Australia reaches 4 million small-scale renewable energy installations”, 10 December 2024. The regulator reports rooftop solar on one in three suitable Australian homes. AEMO, “Minimum operational demand records tumble”, 21 October 2025, describes operational-demand lows associated with high rooftop-solar output.
  22. Australian Government, Department of Climate Change, Energy, the Environment and Water, “Renewables surge as batteries reshape Australia's grid”, 2026. The department reports rapid growth in grid-scale and household battery capacity.
  23. Australian Government, Department of Industry, Science and Resources, Expectations for data centres and AI infrastructure developers, 23 March 2026; and Minister for Industry and Innovation and Minister for Science, “Australian approach to AI: Expectations for data centres to deliver for Australians”, 23 March 2026. The policy sets expectations covering new renewable supply, grid costs and flexibility, water efficiency, jobs, skills and Australian research.
  24. Commission for Regulation of Utilities, Ireland, “The CRU publishes its decision on new electricity connection policy for data centres”, 12 December 2025. The connection policy requires at least 80 per cent of a data centre's annual demand to be matched by additional Irish renewable generation.
  25. Amazon Australia, “Amazon Australia signs nine new renewable energy deals, taking Australian capacity to nearly 1GW”, 19 May 2026.
  26. CSIRO, “GenCost 2025–26 confirms renewables backed by storage remain the lowest-cost new-build electricity”, 1 July 2026. CSIRO reports that renewables backed by storage remain the lowest-cost pathway for new electricity generation and that battery costs continue to fall.
  27. Landon Marston and others, “The environmental footprint of data centers in the United States”, npj Clean Water 4, article 11 (2021). The study describes direct water consumption through cooling-tower evaporation and estimates the water demands of conventional data-centre cooling.
  28. Marston and others, “The environmental footprint of data centers in the United States”, npj Clean Water 4, article 11 (2021), estimates approximately 25.5 million litres of annual direct water consumption for a small 1 MW data centre using traditional cooling—about 7 million litres a day when scaled to 100 MW, within the stated order of magnitude.
  29. Marston and others, “The environmental footprint of data centers in the United States”, npj Clean Water 4, article 11 (2021), discusses the local competition and environmental consequences associated with data-centre water use.
  30. Australian Government, Australia State of the Environment 2021, Water sources. The assessment describes Australia as the driest inhabited continent.
  31. NVIDIA, “Inside the NVIDIA Vera Rubin Platform”, 5 January 2026; and “NVIDIA 2026 CES Special Presentation”, 5 January 2026. NVIDIA describes a high-temperature liquid-cooling design that can reject heat through dry coolers without chillers and recirculates water in a closed loop.
  32. Microsoft, “Sustainable by design: Next-generation datacenters consume zero water for cooling”, 9 December 2024. Microsoft announced a chip-level closed-loop cooling design that, once filled, requires no fresh water for cooling and can avoid more than 125 million litres of water per data centre each year.
  33. IEA 4E Efficient, Demand Flexible Networked Appliances, Liquid Cooling in Data Centres, February 2026. The report estimates that around 22 per cent of data centres use liquid cooling and describes expanding adoption for high-density AI hardware.
  34. IEA 4E Efficient, Demand Flexible Networked Appliances, Liquid Cooling in Data Centres, February 2026. It explains that warm-water liquid-cooling systems operating around 40–50°C can reject heat through dry coolers and eliminate both chillers and evaporative cooling.
  35. Meta, “Introducing Meta Llama 3”, 18 April 2024, reports training on more than 15 trillion tokens collected from publicly available sources. OpenAI, Training data summary pursuant to California Civil Code section 3111, describes models trained on trillions of tokens drawn from public internet content, partner data, user-provided data and synthetic data, including material that may be protected by copyright.
  36. US District Court for the Northern District of California, Bartz et al. v. Anthropic PBC: Order on Fair Use, 23 June 2025, held that training the models at issue was a transformative fair use while Anthropic's retention of pirated books in a central library was not. Associated litigation produced a proposed US$1.5 billion settlement: Associated Press, “Anthropic agrees to pay $1.5 billion to settle authors' copyright lawsuit”, 5 September 2025.
  37. Department of the Prime Minister and Cabinet, transcript of the Prime Minister's National Press Club address, 15 July 2026; and Prime Minister of Australia, “AI in Australia's interests”, 15 July 2026. The Prime Minister said Australian artists should control whether and how their work is used, including its price and value, and announced an enforceable framework to protect Australian intellectual property.
  38. US Copyright Office, Copyright and Artificial Intelligence, documents its multi-part initiative on AI-related copyright law and policy. World Intellectual Property Organization, Artificial Intelligence and Intellectual Property, describes its international AI-and-IP policy work.
  39. Australian Government, Department of Employment and Workplace Relations, “AI and employment in Australia report”, 15 July 2026; Minister for Employment and Workplace Relations, “Australia's labour market remains resilient amid the rise of AI”, 15 July 2026. The government describes the report as its first comprehensive research into AI and employment and reports strong graduate employment and growth in technology and software roles.
  40. Stanford Digital Economy Lab, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, 2025, finds a 16 per cent relative employment decline for 22–25-year-olds in highly AI-exposed occupations. US Census Bureau, Artificial Intelligence and the Labor Market: Early Evidence from Linked Employer-Employee Data, 2026, provides additional linked employer–employee evidence on early-career workers and AI exposure.
  41. Brookings Institution, “Is generative AI a job killer? Evidence from the freelance market”, 2024, reports declines in freelance contracts and earnings after the release of generative-AI tools. A 2026 evidence review reports substantial falls in postings for writing-intensive freelance work: Frontiers, “Generative AI and labor-market outcomes: an evidence review”, 2026.
  42. American College of Radiology, “Radiologist Shortage: Workforce Update”, 2026; and npj Digital Medicine, “Artificial intelligence and the radiology workforce”, 2025. Both discuss growing imaging demand alongside radiology-workforce constraints.
  43. “The impact of artificial intelligence on the radiology workforce: a Jevons paradox?”, 2026, reviews evidence that AI can raise radiology productivity while increased use and demand sustain or increase employment.
  44. Department of the Prime Minister and Cabinet, transcript of the Prime Minister's National Press Club address, 15 July 2026. The Prime Minister used ATMs and bank tellers as an example of technology changing jobs without simply eliminating employment and argued that Australia should shape AI adoption around jobs and national interests.
  45. US Bureau of Labor Statistics, Assessing the Impact of New Technologies on the Labor Market, charts clerical employment as a share of total US employment from 1850 to 2017 and records its sustained decline after its 1980 peak as information technology spread. BLS counted 18.5 million office and administrative-support jobs, or 12.2 per cent of US employment, in May 2023: “18.5 million office and administrative support jobs in May 2023”, 24 April 2024.
  46. Reserve Bank of Australia, Australia and the Global Economy – The Terms of Trade Boom. The RBA links China's industrialisation and urbanisation to higher demand and prices for Australian resource exports, including iron ore, coal and gas.
  47. Norwegian Petroleum, The petroleum tax system, updated 2026. Norway taxes net petroleum-company profits at a combined marginal rate of 78 per cent. The Norwegian Government explains that the state's net petroleum cash flow is transferred to the Government Pension Fund Global in Economic policy.
  48. Norges Bank Investment Management, The fund's value, reports a Government Pension Fund Global value of NOK 21,268 billion at the end of 2025. Statistics Norway, Population, provides Norway's population series.
  49. The Australia Institute, Government revenue from LNG exports: Australia vs Qatar, 2024. The report compares roughly 80 million tonnes of annual LNG exports by each country and estimates government revenue of A$11 billion for Australia and A$56 billion for Qatar.
  50. Sovereign Wealth Fund Institute, Qatar Investment Authority, fund profile and assets-under-management estimate.
  51. Grattan Institute, “How we spent the mining bounty: we should have saved”, 2013. Grattan estimated that tax reductions and spending consumed more than 90 per cent of the nearly A$200 billion temporary revenue windfall from the mining boom.
  52. Australian Bureau of Statistics, “Data centre investment drives new capital expenditure”, 28 May 2026. The ABS reports a record A$7.5 billion of data-processing-equipment capital expenditure in the March 2026 quarter.
  53. Minister for Industry and Innovation and Minister for Science, “Data centres: An honest accounting”, 2026. The speech attributes Australia's first trade deficit in almost a decade to a surge in imported server racks for data centres.
  54. Financial Newswire, “Data centres barely boosting Australia's growth”, reporting HSBC analysis that much of the sector's investment spending leaks into imported equipment.
  55. Westpac IQ, Australian AI Data Bulletin 2026, May 2026. Westpac estimates that imports reduce the net contribution of data-centre investment to GDP to about half its headline value.
  56. Epoch AI, AI data center cost breakdown. Epoch estimates that servers account for roughly 60 per cent of an AI data centre's capital cost.
  57. Intel, Realizing Data Center Savings with an Accelerated Server Refresh Strategy, compares four- and six-year server-refresh cycles. A 2025 equipment-life-cycle study gives an average first life of four years for servers: “Design for circularity—a data centre equipment case study”, Mineral Economics, 2025.
  58. Australian Bureau of Statistics, Introduction of digital services in the balance of payments, 2025. The ABS gives the historical import and export series for digitally delivered services, including the rise in annual imports from A$119 million in 2009–10 to A$18.5 billion in 2023–24 and cumulative flows over that period.
  59. Gartner, “Australian Organizations Will Spend More Than $33 Billion on Public Cloud in 2026”, 11 May 2026. Gartner forecasts A$33.619 billion of 2026 spending, up 17.9 per cent; its prior forecast recorded 18.9 per cent growth for 2025: “Australian Public Cloud End-User Spending to Reach Over $26 Billion in 2025”, 14 May 2025.
  60. Australian Trade and Investment Commission, AI and data centres. Austrade promotes Australia as a regional AI and data-centre hub and identifies its network of international submarine cables and proximity to Asian markets. US Bureau of Industry and Security, Export Administration Regulations, Part 740, lists Australia among the selected destinations eligible for the Artificial Intelligence Authorization provisions governing advanced computing items.
  61. Blackstone, “Blackstone Announces Agreement to Acquire AirTrunk in an A$24 Billion Transaction”, 4 September 2024. Blackstone described it as the largest data-centre transaction completed at that time.
  62. e61 Institute, Is Australia set to be Data Central?, 2026. The analysis distinguishes the modest direct GDP contribution of data centres from the larger potential productivity gains enabled by local compute, software investment, R&D and AI adoption.
  63. Brookings Institution, “New evidence on data center employment effects”, 2025. Brookings finds that developer projections can overstate local employment effects by about 300 per cent and that operational employment at large facilities is comparatively small.
  64. Government of Canada, “Canada to drive billions in investments to build domestic AI compute capacity at home”, 5 December 2024. Canada allocated up to C$2 billion under its Sovereign AI Compute Strategy.
  65. UK Government, UK Compute Roadmap, 17 July 2025. The roadmap includes the AI Research Resource, more than £2 billion of investment through 2030 and a target of at least 6 GW of AI-capable data-centre capacity by 2030.
  66. Élysée, “Faire de la France une puissance de l'IA”, 11 February 2025. France announced €109 billion of private AI investment over the following years.
  67. Australian Government, Department of Industry, Science and Resources, Australian Government response: Senate Select Committee on Adopting Artificial Intelligence (AI) report, 2026. The response reports that Australia received the second-highest capital investment in data centres globally in 2024, at US$6.7 billion.
  68. Prime Minister of Australia, “AI in Australia's interests”, 15 July 2026. The announcement establishes an Office of AI within the Department of Industry, Science and Resources.
  69. Norwegian Ministry of Energy, Norway's oil history in 5 minutes, describes the Norwegian state's direct ownership interests in petroleum fields, pipelines and onshore facilities through the State's Direct Financial Interest. The Norwegian government also owns 67 per cent of Equinor: Equinor, Our shareholders.
  70. Data Center Dynamics, “Australian Stargate data center in Sydney will offer 612MW of capacity”, 2026. It reports OpenAI as the anchor customer for NEXTDC's planned A$7 billion Eastern Creek campus, with 612 MW of capacity and an initial phase targeted for 2027.
  71. Blockspace, “Anthropic hunts for 1.4 GW in Australia, eyes IREN for colocation”, 2026, reporting on Anthropic's search for Australian training capacity. A report summarising the Australian Financial Review account says Anthropic aimed to begin using at least 1 GW by the end of 2027: “Anthropic plans to buy 1.4-GW Australian data centre capacity, AFR reports”, 6 July 2026.
  72. Data Centres Australia and DC Byte, Australian Data Centre Forecast Report, Issue 1, April 2026, estimates Australian operational data-centre capacity at approximately 1.4 GW. ABC News, “Data centre boom sparks warning to consider global location climate risks”, 18 June 2026, reports the same 2025 capacity estimate.
  73. Data Center Dynamics, “Google to provide financial backing for multi-billion-dollar Anthropic data center in Texas – report”, 2026; and Yahoo Finance, “Google reportedly backs $15B data center campus for Anthropic”, 30 July 2026. Both report that financing for the Hubbard, Texas campus included a planned 1.6 GW gas-fired power plant.
  74. Australian Government, Digital Transformation Agency, “New data centre panel”, 28 June 2024, sets mandatory reporting requirements including power consumption and the share of renewable energy. Wired, “The US Government Is Going to Ask Data Centers How Much Power They Use”, 2025, reports the subsequent US plan for a mandatory nationwide data-centre energy survey; the US Energy Information Administration had earlier run a voluntary pilot, “EIA begins collecting data on electricity use by U.S. data centers”, 8 February 2024.

This white paper is also available as a fully footnoted PDF — use the Download PDF button above.