The sentence that contains the problem
AEMO's 2026 Integrated System Plan describes Australian data centres in a single passage. They "account for around 2% of today's grid-supplied electricity use", and that use "is projected to grow in the ISP by around 25% annually to reach almost 10% of the NEM's underlying demand by 2050: five times the share it has today, and the equivalent of 20% of today's total demand."
Read quickly, that is one trend line: 2 per cent now, 10 per cent later. Read carefully, it is two measurements. The 2 per cent is a share of grid-supplied electricity. The 10 per cent is a share of underlying demand. Those are different quantities, and in a grid with as much rooftop solar as Australia's, they are not close to each other.
What the two bases mean
AEMO's operational, or grid-supplied, demand is the electricity that must come from centrally dispatched generation. Underlying demand is that same consumption plus the output of distributed rooftop solar, which never reaches the wholesale market because it is consumed where it is generated. Underlying demand is therefore always the larger of the two, and the gap between them is, in effect, the rooftop fleet.
The size of that gap in Australia is not marginal. The 2026 ISP records that in aggregate, rooftop solar has met just over 60 per cent of underlying electricity demand across the NEM on a sunny weekend when industrial demand was low. That is a single-period peak rather than an annual average, but it establishes the order of the quantity being divided by.
The same forecast, published two ways
The absolute number is stable across AEMO's documents. The Oxford Economics Australia report that AEMO commissioned and publishes puts Australian data centre consumption at 3.9 TWh in FY25, with the NEM accounting for 98 per cent of it and representing approximately 2 per cent of NEM grid-supplied consumption. Under Step Change it forecasts growth averaging 25.1 per cent a year to 12.0 TWh by FY30, or 6 per cent of the NEM's grid-supplied electricity, and 34.5 TWh by FY50.
That report then states the FY50 share on its own basis: "Of the 34.5 TWh, 33.8 TWh of this is forecast to be on the NEM, representing 12% of NEM grid-supplied consumption by FY50." Its figure note is explicit about the denominator, defining the share of NEM grid-supplied electricity as data centre consumption over NEM operational sent-out consumption from the 2025 Electricity Statement of Opportunities.
The ISP takes the same trajectory and states its headline share against underlying demand: almost 10 per cent by 2050. Both statements are correct. The same numerator over the larger denominator returns the smaller share.
| Published statement | Numerator | Denominator | Share | Source |
|---|---|---|---|---|
| Today | ~3.9 TWh (FY25) | Grid-supplied | ~2% | 2026 ISP; Oxford Economics for AEMO |
| FY30, Step Change | 12.0 TWh | Grid-supplied | 6% | Oxford Economics for AEMO |
| FY50, Step Change | 33.8 TWh on the NEM | Grid-supplied | 12% | Oxford Economics for AEMO |
| 2050, Step Change | Same trajectory | Underlying | almost 10% | 2026 ISP |
Rows are as each document publishes them. The FY50 grid-supplied share and the 2050 underlying share describe the same Step Change forecast on two different bases; they are not competing estimates and neither supersedes the other.
The residual is the rooftop fleet
Back-solving the two published shares gives a sense of the distance between the bases. A numerator near 34 TWh at almost 10 per cent implies underlying demand around 340 TWh; the same numerator at 12 per cent implies grid-supplied demand around 283 TWh. The difference, of the order of 55 to 60 TWh, is consumption that never clears the wholesale market. These implied totals are derived here from the published percentages as an estimate, and are not figures AEMO publishes.
The ISP does publish the components that make the residual plausible. Businesses are forecast to invest in 15 TWh more rooftop solar by 2050 and 27 TWh more large on-site generation. Household grid-supplied energy needs are forecast to fall 44 per cent to 20 TWh despite households consuming more, because more of what they consume comes off their own roofs. Business and industry, after drawing on their own rooftop solar, are forecast to need 280 TWh from the grid in 2050.
Why this matters beyond bookkeeping
On 23 July 2026 the Clean Energy Council published Powering the digital economy, which proposes a Flexible Contracting Framework requiring new data centres to match their consumption with additional firmed renewables, backed by large-scale generation certificates until 2030 and renewable energy guarantee of origin certificates thereafter, on a glide path aligned to each facility's ramp-up. The paper puts current data centre consumption at 4 TWh a year, about 2 per cent of NEM demand, across more than 250 operating facilities, and projects that share reaching almost 10 per cent by 2050.
An obligation to match consumption with new generation is written against electricity a facility draws from the grid. A demand share quoted against underlying demand is measured against a base that includes electricity nobody draws from the grid. A framework and a forecast expressed on different bases can be individually accurate and still not answer each other. Where a submission, a business case or a network application cites "almost 10 per cent by 2050", the useful follow-up question is which of the two bases the surrounding argument assumes.
Two counts of the same fleet
The facility counts also differ. The 2026 ISP states that there are currently over 160 operational data centres in Australia, with almost half in Sydney and most others in Melbourne, Brisbane and Perth. The Clean Energy Council paper describes more than 250 facilities operating. Neither document defines "data centre" in terms that make the two counts directly comparable, and a count that includes small edge and telco sites will not match one built from grid-connected load. The gap is a definitional difference to resolve before either number is carried into a model, not a contradiction to adjudicate here.
What the record will settle
The near-term parts of this forecast are checkable. At the end of the March 2026 quarter, the ISP records 11 large data centres representing over 5 GW of maximum demand in the connections pipeline, with connections historically taking about two years from application to energisation and ramping to full demand over five to ten years. Those are dated, falsifiable statements, and the FY30 waypoint, 12.0 TWh at 6 per cent of grid-supplied electricity, arrives well inside the life of this record.
One internal inconsistency is worth noting for anyone quoting the ISP directly. The document describes the 2050 share as "four times the share it has today" in one passage and "five times the share it has today" in three others. Both multipliers compare a share measured on the grid-supplied basis with one measured on the underlying basis, which is the same arithmetic the rest of this piece describes.
None of this makes either publication wrong. It makes the denominator a required field. A forecast expressed as a percentage is two claims wearing one number, and only one of them is usually stated out loud.