The Great Australian Power Squeeze: When AI Meets the Energy Transition
Australia may be about to conduct one of the more interesting infrastructure experiments in the developed world. We are attempting to retire much of the electricity generation upon which the country historically depended, electrify increasing portions of transport and industry, build a predominantly renewable replacement system, and at exactly the same time welcome an industry whose appetite for electricity is expanding at an extraordinary rate. Welcome to the age of the data centre.
Macrobusiness has drawn attention to the developing collision, particularly in Sydney, where Transgrid has received more than 20 gigawatts of data-centre connection enquiries. To put that extraordinary number into perspective, it is approximately twice the current peak electricity demand of the entire state of New South Wales. That does not mean another 20 GW of demand is about to appear tomorrow morning. Connection enquiries are not operating data centres, projects are cancelled, delayed, downsized or never financed, and even completed facilities do not necessarily operate continuously at their maximum connection capacity. Any claim that NSW therefore suddenly needs to triple its electricity supply would be nonsense, but unfortunately the less sensational numbers are alarming enough.
The Australian Energy Market Operator has substantially increased its forecast for data-centre electricity consumption. In the National Electricity Market, data centres are expected to consume approximately 5 TWh in 2025–26, rising under the central scenario to approximately 34 TWh by 2035–36. That would take data centres from around 3 per cent of electricity supplied through the NEM today to approximately 13 per cent within a decade. Thirty-four terawatt-hours is an enormous amount of electricity, approaching the present combined residential electricity consumption of New South Wales and Victoria.
More remarkable is how quickly the forecast has changed. The number of known data-centre developments under consideration has more than doubled in roughly a year, from 97 to 225. AEMO's central forecast for data-centre electricity consumption is now substantially higher than its previous modelling, demonstrating the difficulty of forecasting an industry undergoing explosive expansion. Unfortunately, electricity systems are not constructed at software speed. A technology company can announce a new AI model, raise billions of dollars and order thousands of processors remarkably quickly, while building transmission lines, power stations, pumped hydro facilities and other major electricity infrastructure can consume years through planning approvals, environmental assessments, land acquisition, financing, equipment procurement and construction.
Electricity demand can therefore materialise considerably faster than electricity infrastructure, which is exactly what Transgrid is confronting around Sydney. It says the transmission network serving Western Sydney has limited remaining capacity. Around 1.5 GW of additional data-centre load can apparently be accommodated within existing arrangements, but anything substantially beyond that will require further network investment. The company has consequently warned of supply constraints and potential reliability problems unless investment keeps pace with demand.
This brings us to the strange timing of Australia's AI revolution. Over approximately the same decade in which AEMO expects data-centre electricity consumption to increase enormously, around 15 GW of coal and gas generation is scheduled to retire from the National Electricity Market. There is nothing inherently impossible about replacing that generation. Australia possesses renewable-energy resources, solar generation is abundant during daylight hours, wind can provide quantities of electricity, batteries are being installed rapidly, transmission is expanding and household solar has transformed the structure of daytime electricity generation; or so it is said.
AEMO consequently does not predict imminent catastrophe. Its current Electricity Statement of Opportunities finds an improved near-term reliability outlook and no forecast reliability gaps before 2030, assuming committed and anticipated investments proceed as expected. AEMO identifies a large pipeline of new generation and storage, which is substantial evidence against simplistic claims that Australia's electricity system is about to collapse, but it simultaneously expects electricity consumption to increase dramatically as electrification and data centres increase demand.
We are therefore not merely replacing old generators; we are attempting to replace old generators while dramatically expanding the electricity system. Those are very different engineering tasks. The distinction tends to disappear in political discussions of the energy transition. Closing a 2 GW coal-fired power station and building 2 GW of wind and solar does not leave the system unchanged. Wind and solar produce electricity according to weather and time of day, whereas a dispatchable thermal generator can generally be instructed to produce when required. A renewable system therefore needs additional generation capacity, storage, transmission, demand management and other services to deliver reliable electricity across time rather than merely matching nameplate capacity.
Now place enormous data centres on top of that transition. Data centres have characteristics that make them particularly interesting electricity consumers. They can require hundreds of megawatts, while hyperscale facilities can approach the consumption of substantial industrial complexes. Unlike an aluminium smelter built around twentieth-century industrial technology, their demand can also expand rapidly as racks are filled with increasingly powerful processors. Artificial intelligence intensifies the problem because AI computation is extremely energy intensive.
The great AI companies are effectively building factories whose product is computation. Instead of turning iron ore into steel or bauxite into aluminium, they turn electricity into tokens, images, models and answers. Their raw material is electrons, and this creates a peculiar political contradiction. Governments want AI investment because it promises billions of dollars of capital expenditure, construction activity and participation in what may become one of the defining industries of the twenty-first century, while those same governments have made rapid electricity-sector decarbonisation a central policy objective. The two ambitions are now colliding.
Canberra's emerging solution is essentially to make new data centres bring their electricity with them. Large data centres may be required or encouraged to support additional renewable generation and firming sufficient to offset their consumption while providing some capacity to reduce demand when the grid is under severe stress. In principle, this makes considerable sense. If a hyperscale data centre adds 500 MW of enormous new demand, existing households should not have to finance all the transmission and generation required to accommodate it. Nor should an AI facility consume existing renewable generation and then announce that it operates on "100 per cent renewable electricity" because it purchased certificates while the rest of the system burns additional gas to compensate. Additional demand requires additional supply.
Even that apparently simple equation becomes complicated in the physical electricity system. Suppose a data centre contracts enough solar and wind generation to produce annually as many megawatt-hours as it consumes. On an accounting spreadsheet, everything balances beautifully, but electricity grids do not operate annually; they operate every second. The data centre may want 300 MW at 2 am during a wind drought when its contracted solar farm is producing precisely zero. Somewhere in the system, something else must supply that electricity. Batteries can bridge shorter periods, pumped hydro can provide longer-duration storage, interconnectors can transfer electricity between regions and gas generators can supply dispatchable backup, but somebody has to build and pay for all of it.
This is why the data-centre boom could expose weaknesses in Australia's energy transition that were easier to conceal while electricity demand was relatively stagnant. For years, Australia benefited from flat or declining grid demand as rooftop solar reduced daytime consumption from central generators. That made retiring ageing coal stations somewhat easier to accommodate. AI changes the arithmetic, and so does electrification. Electric vehicles require electricity, electric heating requires electricity, electrified industrial processes require electricity and green hydrogen, if produced on anything like the scale once envisaged, requires staggering amounts of electricity. Add hyperscale AI facilities and the supposedly declining-energy economy becomes an electricity-hungry economy.
There is nothing wrong with that if supply expands first or simultaneously. The danger arises if political ambition expands faster than physical infrastructure. Imagine a severe summer heatwave in the early 2030s. Household air conditioners are running hard across Sydney, solar production falls towards evening just as residential demand remains high, several ageing generators have already retired, wind generation happens to be weak, batteries discharge, interconnectors operate heavily and the wholesale electricity price rises sharply. Sitting beside this system are gigawatts of data centres whose computers do not particularly care that Sydney is hot.
The preferable solution would be contractual demand flexibility. Some computational work can potentially be delayed, shifted geographically or reduced temporarily, and data centres could even become useful grid participants by adjusting workloads when electricity becomes scarce. But not every computing task is flexible. Cloud services, financial systems, communications, government systems and real-time AI applications require very high availability. Data-centre operators sell reliability precisely because customers expect their services to work continuously, which is why many data centres possess backup generators in the first place.
There is a wonderfully Australian irony emerging in the Northern Territory. Developers are proposing enormous data-centre projects associated with Beetaloo Basin gas. Rather than waiting years for transmission connections and sufficient renewable capacity, they can potentially locate generation beside the data centre and burn gas directly. One proposed campus near Darwin has been envisaged at an extraordinary 2 GW. Thus, the AI revolution may produce an outcome environmental policymakers did not anticipate: after years spent attempting to move electricity away from fossil fuels, governments may discover that the world's fastest-growing electricity consumers value reliability so highly that gas once again becomes attractive.
That is why the 20 GW of connection enquiries around Sydney should cause policymakers to look very carefully at their timetables. The issue is not whether all those projects will be built, because they won't. The issue is what happens if even a substantial fraction of them are built while coal retirements, transmission delays and rapidly rising general electricity demand occur simultaneously. Australia would then have manufactured an energy squeeze through the interaction of two policies that individually sounded visionary: become a renewable-energy superpower and become an artificial-intelligence superpower. Now supply the electricity for both.
There is a final question that receives surprisingly little attention: what does the ordinary Australian obtain in return? If data centres create high-value Australian employment, domestic technological capability, tax revenue and productivity improvements, additional electricity infrastructure may be an excellent national investment. But data centres are unusually capital-intensive facilities and employ relatively modest numbers of people once construction is complete. If Australians instead pay higher network charges, surrender scarce grid capacity and finance additional infrastructure so multinational technology corporations can run machines serving customers around the world, the national cost-benefit calculation becomes considerably less obvious.
That is why the emerging principle that data-centre developers requiring network expansion should pay for it makes sense. The same principle should apply throughout the system. Bring the AI investment, build the data centres and develop the technology, but bring the generation, storage and transmission necessary to support them as well. Australia cannot run a twenty-first-century artificial-intelligence economy on promises of electricity that will arrive sometime after the computers do. The grid has no interest in political slogans; it has only supply and demand, and when demand exceeds supply, even artificial intelligence cannot negotiate with the laws of physics.
https://www.macrobusiness.com.au/2026/09/data-centres-wreck-the-energy-transition/
