The short version
Australia has spent two years attracting the compute that AI runs on and about six months working out what it costs everyone else. That second part came to a head in the first week of August 2026.
On 5 August the Australian Energy Market Commission published advice to energy ministers setting out four things large data centres should be required to do: bring new clean generation with them, prove that demand is firmed, register as participants in the electricity market, and operate flexibly. The same day, the NSW government introduced a bill giving its energy minister the power to control which large loads get grid access at all. Also the same day, federal Energy Minister Chris Bowen told the National Press Club that states would be free to add more rigorous requirements but not to water them down.
Two weeks earlier, on 22 July, Bowen had lodged two rule change requests with the AEMC — Cost recovery for network augmentations Package 1 and Package 2 — aimed squarely at who pays when a very large new connection forces an upgrade to the shared network.
None of this is about your phone, your laptop or the model you talk to. It is about the fact that the thing running the model is a building, it wants hundreds of megawatts, it wants them in western Sydney or Melbourne’s west, and it wants them soon.
How much power are we actually talking about?
The reference point is the Australian Energy Market Operator’s 2026 Integrated System Plan, published 25 June 2026. AEMO’s own framing is that data centres emerged as significant consumers faster than its previous plans expected.
| Today | Data centres used about 3.9 TWh in FY25, some 98 per cent of it from the National Electricity Market — roughly 2 per cent of NEM grid-supplied consumption. Total NEM consumption is around 205 TWh. |
|---|---|
| By FY30 | Under AEMO’s Step Change scenario, data centre consumption grows at about 25 per cent a year to 12 TWh — around 6 per cent of NEM grid-supplied electricity. AEMO has described that as roughly the electricity used by every home in Victoria. |
| By 2050 | 34 TWh a year, close to 10 per cent of underlying demand, within a NEM that has itself grown to about 390 TWh. Business and industrial grid demand doubles from about 140 TWh to 280 TWh over the same period. |
| In the queue | At the end of the March 2026 quarter, AEMO counted 11 large projects above 5 MW representing 5.4 GW of maximum demand working through transmission connection, about 60 per cent in NSW and 40 per cent in Victoria. By August, trade press was reporting 17 projects and 9 GW in the connection process against a pipeline valued around A$150 billion — different counts, different cut-offs, same direction. |
| The global backdrop | The International Energy Agency projects global data centre electricity consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030. |
Two things are worth holding onto from that table. First, in absolute terms Australian data centres are still a small slice of the grid — two per cent is a rounding error next to households and heavy industry. Second, the growth rate is the story, and it is concentrated: two states, a handful of postcodes, and connection requests measured in gigawatts rather than megawatts.
The money behind that is not speculative. Amazon Web Services committed A$20 billion across 2025 to 2029 to expand data centres in Sydney and Melbourne, alongside three new solar farms in Victoria and Queensland totalling more than 170 MW. In April 2026 Microsoft announced A$25 billion (about US$18 billion) of capital and operating expenditure in Australia, which it said would grow its local cloud footprint by more than 140 per cent across 29 existing sites and three Azure regions.
The 26 per cent argument
On 3 June 2026 the Climate Council published Clouded future: managing risks of the data centre boom. Its headline finding is the number that has driven the debate since: if the extra demand is met largely with gas rather than new renewables, wholesale electricity prices could be up to 26 per cent higher in NSW and 23 per cent higher in Victoria by 2035, with about 14 per cent additional pollution from the main grid. The report counted 162 operational data centres and more than 90 in the pipeline, and noted Australia ranked second globally for data centre investment in 2024.
Climate Councillor and energy specialist Joel Gilmore put the mechanism this way: data centres are “like a giant snowball” that, unmanaged, end up “dumping massive costs onto households”.
Read the conditional clause carefully, because it is doing the work. The modelled increase is not a consequence of data centres existing. It is a consequence of new demand arriving faster than new generation, in which case the marginal generator setting the wholesale price is more often gas. That is precisely the gap the AEMC framework is designed to close.
The counter-argument, which is also real
The industry answer is that a big, flat, round-the-clock load is good for a network built to survive a few hot afternoons. NEXTDC has argued data centres soak up excess daytime solar and improve utilisation, pushing down the network component of bills — which is 30 to 40 per cent of what you pay. The NSW distributors (Ausgrid, Endeavour Energy and Essential Energy) have made the same submission: more demand across the same poles and wires lowers the share of fixed network costs borne by everyone else.
There is one concrete data point on that side. In May 2026 the regulator’s decision on Jemena’s Victorian distribution prices had annual network costs falling by about $189 for a typical residential customer and $516 for a small business by 2030-31 compared with 2025-26, partly because higher consumption spreads fixed costs further.
Our reading, and we are reasoning past the sources here: both claims can be true at once because they are about different lines on the same bill. Network charges are fixed costs divided by throughput, so more throughput helps. Wholesale energy charges are set by the most expensive generator dispatched in each interval, so more demand without more supply hurts. Whether a household ends up ahead or behind depends entirely on whether new generation arrives with the new load — which, again, is what the rules published on 5 August are trying to force. The industry benefit is also, for now, largely prospective; the price risk is modelled out to 2035. Neither has actually been observed on a bill yet.
What the AEMC actually recommended
The AEMC delivered its advice to the Energy and Climate Change Ministerial Council on 28 July and published it on 5 August 2026. Four changes, aimed at making very large loads carry their own weight rather than blocking them.
| Bring new clean energy | Surrender Renewable Electricity Guarantee of Origin certificates sourced from new, additional generators to offset the electricity consumed. The word doing the work is “additional” — buying certificates from existing wind and solar simply moves clean electrons from one buyer to another and adds nothing to the system. |
|---|---|
| Prove it is firmed | Demonstrate the new demand is backed by new firm capacity, so that it does not lean on the rest of the market during the hours renewables are not running. This is the clause aimed directly at the wholesale price effect above. |
| Register as a market participant | Large facilities become registered participants in the National Electricity Market, subject to market rules — which also gives AEMO real-time visibility of loads it currently sees only indirectly. |
| Operate flexibly | Connection agreements should reward shifting demand and co-locating generation and storage on site, so a data centre can act as a grid resource rather than only a drain on one. |
| Timeline | Roughly 12 months for the renewable certificate obligation, 24 to 36 months for market registration and connection reform. The rule change requests are expected before ministers in September 2026, with federal legislation flagged for early 2027. |
AEMC chair Anna Collyer framed it as a design problem rather than a fight: “Data centre growth does not have to come at a cost to other consumers, but that depends on getting the settings right from the start.” And, on the overseas experience: “The lesson from other jurisdictions isn’t that data centres are the problem, it’s that very large loads need to be planned, connected and operated as part of an integrated system.”
NSW went further, on the same day
The Electricity Infrastructure Investment Amendment Bill 2026, introduced to the NSW Legislative Assembly on 5 August 2026, takes the access-scheme machinery built for renewable energy zones and points it at large loads. The minister would be able to declare a large load infrastructure access scheme over all or part of the state, authorising or prohibiting access to specified network infrastructure for facilities capable of drawing 5 MW or more. The stated principle is that the data centre carries the cost of the network infrastructure its load requires, and that NSW electricity customers do not.
That is a meaningful shift. Grid connection in Australia has generally been a queue with technical conditions attached. A declared access scheme makes it a gate.
The technical rule nobody is arguing about
Running underneath the economics is a rule change most coverage skips. On 12 March 2026 the AEMC published a draft rule setting technical access standards for large inverter-based loads, requiring them to ride through voltage and frequency disturbances and recover their consumption within defined windows instead of dropping off the network. The draft raises the compliance threshold from 5 MW to 30 MW for facilities connecting at distribution level. Submissions closed 7 May and a final determination is due late October 2026.
The reason is a real event overseas: in July 2024, roughly 60 data centres in Virginia disconnected about 1,500 MW of load simultaneously during a single fault. As Collyer put it, “Data centres aren’t passive loads anymore; they’re active grid participants. When they fail to ride through faults, it has the potential to trigger cascading failures.” A grid can plan for a large load. It cannot easily plan for a large load that vanishes and returns in the same second.
The water nobody is measuring
Electricity gets the headlines; water is the messier number, because Australia does not really have it. Michael Vardon of the ANU made that argument in The Conversation on 3 June 2026: the Bureau of Statistics bundles data centres into broader industry categories, so the granular consumption data a policymaker would want does not exist.
What estimates there are diverge sharply, and the divergence itself is informative:
| Industry-commissioned estimate | About 5.5 GL a year of direct on-site use nationally, roughly 0.04 per cent of national water use — about 0.7 per cent of Sydney’s supply and 0.2 per cent of Melbourne’s, rising to roughly 1.9 per cent and 0.9 per cent respectively by 2030, with national direct demand reaching around 17 GL. |
|---|---|
| Academic estimate | Around 300 existing Australian data centres totalling about 1.3 GW would use somewhere between 15,000 and 35,000 ML a year — still a fraction of one per cent nationally. |
| Utility evidence | Sydney Water told a NSW parliamentary inquiry that under a high-growth scenario with water-intensive cooling, cumulative demand could reach 250 ML a day by 2035. Vardon notes the 41 proposed NSW facilities could take 15 to 20 per cent of Sydney’s supply within a decade, with individual sites seeking between 5 and 40 ML a day. |
| Why they disagree | Industry models what operating facilities actually consume. Utilities model the connection applications on their desk. Both can be honest and land an order of magnitude apart — and the NSW consultation paper concedes precise water demand data is simply not available at the planning stage. |
The practical point for a reader is not which estimate wins. It is that cooling design is a choice, evaporative cooling trades water for electricity, and nothing currently requires an operator to disclose either.
What this means for you
| Nothing changes on your next bill | Every number in this article is a forecast for 2030, 2035 or 2050. No published analysis attributes a current retail price movement to data centres. Treat any marketing or campaign material that says otherwise as making a claim it cannot yet support. |
|---|---|
| Learn which line is which | Your bill has a network component (poles, wires, fixed costs) and a wholesale energy component. The industry’s cost-down argument applies to the first; the Climate Council’s cost-up modelling applies to the second. Anyone quoting one without the other is only telling you half of it. |
| The rules are still open | The two federal cost-recovery rule changes (ERC0448 and ERC0456) are lodged but not yet initiated — the AEMC publishes a consultation paper and takes public submissions when it starts. The large-load technical standard is due late October 2026. These are genuinely open processes, not announcements. |
| Watch for “additional” | When an operator says its facility is 100 per cent renewable, the question that matters is whether the certificates come from new generation built for that load or from existing plant. The AEMC has proposed making the answer mandatory. Until it is, it is a marketing claim. |
| If you live near a proposal | Water and land-use questions are decided in state planning processes and parliamentary inquiries, not in the electricity rules. That is where the disclosure gap sits, and where submissions from residents actually reach a decision-maker. |
| The concentration is the risk | Around 60 per cent of the connection queue sits in NSW. National averages will keep looking reassuring while a handful of network areas absorb almost all of the growth. Read state-level numbers, not national ones. |
The bottom line
The interesting thing about the first week of August is not that Australia decided to slow the AI build-out. Nobody in this process is proposing that; A$45 billion of announced hyperscaler investment says the opposite. What was decided is that the cost of connecting it should sit with the people building it, and that new demand of this size should arrive with new supply attached.
Whether that holds is the thing to watch over the next 18 months, and it will be settled in three unglamorous places: whether the REGO obligation survives consultation with its additionality requirement intact, whether the two cost-recovery rule changes actually reach the network upgrades a large connection triggers rather than only the connection asset itself, and whether anyone ends up obliged to publish how much water a building is using.
The forecasts will keep moving. The design question underneath them is stable, and it is the same one Australia faced with rooftop solar and again with electric vehicles: what happens when a lot of new load or new supply shows up somewhere the grid was not planned for. This time the load arrives in single 300 MW increments, with a signed lease and a deadline.
Sources
Facts, figures, dates and quotes above are drawn from these primary and reputable sources, captured 13 August 2026:
- AEMC — Data centres to bring clean, firm energy, be flexible and pay their way (5 August 2026): the four recommended changes, the REGO additionality requirement, market registration, flexibility and co-location, and both Anna Collyer quotes.
- AEMC — AEMC proposes new grid standards for data centre connections (12 March 2026): the draft technical access standards for large inverter-based loads, the 5 MW to 30 MW threshold change, the 7 May submission close, the Virginia event of July 2024 in which about 60 data centres dropped 1,500 MW, and Collyer on ride-through.
- AEMC — Cost recovery for network augmentations Package 1 (ERC0448) and Package 2 (ERC0456): both lodged by the Hon Chris Bowen MP on 22 July 2026, both still pending initiation and consultation.
- PV Tech — Australia’s data centre energy rules take shape as AEMC publishes framework and NSW moves to control grid access (7 August 2026): the NSW Electricity Infrastructure Investment Amendment Bill 2026 introduced 5 August, the large load infrastructure access scheme power, the 5 MW threshold, the cost-allocation principle, and the 12-month and 24-36 month implementation windows.
- Energy-Storage.news — Australia advances data centre energy policy on three fronts (11 August 2026): the 28 July delivery to the Energy and Climate Change Ministerial Council, Chris Bowen’s National Press Club remarks that states may add more rigorous requirements but not water them down, federal legislation flagged for early 2027, the A$150 billion pipeline figure, and 17 projects totalling 9 GW in the connection process.
- W.Media — AEMO: data centres could account for almost 10% of NEM electricity demand by 2050 (25 June 2026), reporting AEMO’s 2026 Integrated System Plan: 205 TWh today rising to about 390 TWh by 2050, 34 TWh of data centre demand at close to 10 per cent of underlying demand, about 160 operational facilities, business and industrial demand doubling to 280 TWh, and 11 projects totalling 5.4 GW in the connection process at the end of the March 2026 quarter.
- PV Magazine Australia — AI is driving a data centre boom but can the grid keep up? (10 June 2026): 3.9 TWh consumed in FY25 with 98 per cent from the NEM and about 2 per cent of grid-supplied consumption, 12 TWh and 6 per cent by FY30 under Step Change, and the IEA projection of global consumption rising from 485 TWh in 2025 to 950 TWh in 2030.
- Climate Council — Clouded future: managing risks of the data centre boom (3 June 2026): the 26 per cent NSW and 23 per cent Victorian wholesale price modelling to 2035 conditional on gas meeting the demand, 14 per cent additional grid pollution, 162 operational data centres and more than 90 in the pipeline, Australia second globally for investment in 2024, single-site water requests up to 40 million litres a day, and the Joel Gilmore and Amanda McKenzie quotes.
- RenewEconomy — Data centres say they will drive down the network part of bills (11 June 2026): NEXTDC’s argument on flat load and solar soak-up, the network component being 30 to 40 per cent of bills, the Ausgrid, Endeavour and Essential Energy submission on utilisation, and the May 2026 Jemena decision cutting typical residential network costs by $189 and small business by $516 a year by 2030-31.
- The Conversation — How much water and power will AI data centres use in Australia? Ironically, we don’t have the data to know (3 June 2026), Michael Vardon, ANU: about 300 existing facilities at 1.3 GW using 15,000 to 35,000 ML a year, the 41 NSW proposals potentially taking 15 to 20 per cent of Sydney’s water within a decade, individual sites seeking 5 to 40 ML a day, the CEFC estimate of 11 per cent of electricity within a decade, and the ABS disaggregation gap.
- ArchitectureAU / University of Melbourne Pursuit — What AI data centres really mean for Australia’s water supply (3 August 2026), M. Reza Hosseini: 5.5 GL a year and 0.04 per cent of national water use, 0.7 per cent of Sydney and 0.2 per cent of Melbourne supply rising to 1.9 and 0.9 per cent by 2030, national demand reaching about 17 GL, Sydney Water’s 250 ML a day by 2035 evidence to the NSW inquiry, and why utility and industry forecasts diverge.
- Austrade — AWS plans to invest A$20 billion to expand digital infrastructure in Australia by 2029 (July 2025): the A$20 billion 2025-2029 commitment, Sydney and Melbourne data centres, three new solar farms in Victoria and Queensland totalling more than 170 MW, and the Albanese and Garman quotes.
- CNBC — Microsoft expands AI footprint in Australia with $18 billion investment (23 April 2026): the A$25 billion (US$18 billion) capital and operating commitment announced at the Microsoft AI Tour in Sydney, and cloud footprint growth of more than 140 per cent across 29 existing sites and three Azure regions.