Summary

  • A data centre's megawatts enter AEMO's national demand forecast through a status the operator assigns — committed, anticipated or proposed — and for data centres AEMO applies a stricter stage test than it applies to other industrial loads.
  • The same operator defines anticipated a second way, on five commitment criteria, and does not reconcile the two definitions in the documents behind this piece.
  • AEMO also publishes its own attrition: more than 40% of data-centre projects since 2025 dropping out or regressing in connection status, about 27% average grid connection utilisation in FY26, and roughly 5 TWh consumed in FY2026 — about 0.57 GW of average draw, which is my arithmetic, not AEMO's.

Around 165 data centres consumed approximately 5 terawatt hours of electricity in financial year 2026, equivalent to around 3% of total NEM operational electricity consumption, and average monthly data-centre demand across the market rose from nearly 200 MW to more than 700 MW between January 2018 and June 2026 (AEMO, 2026 ESOO data-centre forecasting overview). Set beside that operating record: Projects in development: 225 data centres with 67 GW of combined capacity. These two numbers answer different questions. The 67 GW is capacity attached to projects; the 5 TWh is what was actually supplied. AEMO's own parameters explain why the gap is a schedule rather than a scandal — Grid connection utilisation: averaging ~27% in FY26, Load maturity: 5 to 10 years, and mature data centres in aggregate are projected to utilise around 50% of their network connection capacity.

AEMO describes its classification of a prospective load in its 2026 Electricity Demand Forecasting Methodology: committed, anticipated or proposed, decided on the likelihood that the project will be developed, from criteria considered in combination — the project's stage in the connection process, environmental and planning approvals, financial and contract information, and other material from its large-industrial-load survey and from network service providers. That is the methodology's description, not its operative wording; the operative rules sit in bullets and footnotes, and one footnote does most of the work for anyone counting data-centre megawatts.

Footnote 11 reads: For rapidly growing sectors such as data centres a more stringent criteria is applied: at least application stage in the connection process for the period up to and including the reliability obligation timing threshold. The methodology then explains its own term — This threshold is here termed the reliability obligation timing threshold — and fixes it for the NEM to the timeframes of the reliability instruments that may be created under the Retailer Reliability Obligation, including up to the T-3 cut-off day, where The T-3 cut-off day refers to the day that is three years away from the current reliability assessment.

So the rung a data centre stands on inside that window is a document a developer filed with a network business, on a clock set by a legal instrument. The inclusion rule is explicit: Up to and including the reliability obligation threshold, AEMO considers anticipated data centre projects with a grid connection application in the Central and High scenarios (see Figure 9). Beyond the reliability obligation threshold all anticipated projects are included in the Central and High scenarios, proposed projects with a connection application are only included in the Central scenario and proposed projects with a connection enquiry are included in the High scenario. AEMO states its reasoning in the same breath: This refined inclusion criteria reflects both the fast pace and high uncertainty of data centre developments relative to more traditional industrial loads. And the figure note supplies the comparison that makes the rule legible: Anticipated projects that are not data centre developments are not considered in this period.

Read against the generic criteria, that produces an odd plateau. Committed projects include those at application or a later stage in the connection process; anticipated projects that are not data centres need only be at least at enquiry or a later stage in the connection process. For data centres, the anticipated rung is lifted to the stage the text otherwise associates with commitment.

Then there is the second vocabulary for the same two words. AEMO's 2026 ESOO overview defines them on a different basis: Committed project: meets all five of AEMO's commitment criteria but has not yet commissioned at least 30% of their capacity. Anticipated project: made progress towards at least three of AEMO's commitment criteria. The 2023 ESOO summary names the five — land, contracts, planning, finance and construction — and adds a recency condition the 2026 overview text does not carry: Anticipated projects have made progress towards at least three of AEMO's commitment criteria and have provided AEMO confirmation or update of project status in the last six months.

Nothing in the documents behind this piece reconciles the two tests. The methodology defines status by connection stage and criteria in combination; the overview defines it by a five-part scorecard; and AEMO does not say in the retrieved text whether that scorecard reaches load projects at all or only generation and storage. The 2025 ESOO, which might settle it, could not be obtained on 19 November 2026 — the PDF returned a response too large to fetch and the newsroom page an access denial — so the six-month condition is verified from the 2023 summary.

Every AEMO and AER passage here was read as extracted page text through retrieval, not opened in a document viewer, and the rev-parametered copy of the 2026 methodology was too large to fetch, so the quotations come from the same file at its plain URL.

AEMO then measures its own ladder, and that is the part worth sitting with. The 2026 Electricity Statement of Opportunities records that its forecasts considered the potential development of 225 known data centre projects, with more than 40% of data centre projects since 2025 either dropping out or regressing in connection status. The overview attaches a 36% cancellation rate to the projects that were in the 2025 ESOO. The media release lists what AEMO counts as the methodological work: project attrition during the connection process, the difference between connection capacity and actual electricity use, and the gradual ramp-up of demand over time. The publication itself is generally based on information available to AEMO as at 1 July 2026 unless otherwise indicated.

Some arithmetic, mine and labelled as mine. Five terawatt hours across 8,760 hours is about 0.571 GW of average draw, which makes the 67 GW of combined project capacity roughly 117 times what the operating fleet actually averaged. If every one of the 225 projects were built and ran at the 27% utilisation AEMO reports, delivered average load would be about 18.1 GW — still more than thirty times the FY2026 average.

The Step Change case of about 34 TWh by FY36 divides to about 3.9 GW of average draw, roughly one seventeenth of the pipeline, and applying AEMO's 50% mature-fleet utilisation to that implies about 7.8 GW of connection capacity rather than 67 GW. None of this makes the pipeline wrong. It makes it a different quantity from the one a reader sees first.

The regulator has already made a narrower version of that point. The AER's August 2025 Transparency Review of the 2025 Inputs, Assumptions and Scenarios Report found: We consider that AEMO has not adequately explained how these forecasts were derived, such that they could materially change without changing the description of content in the IASR. It added that stakeholders had not been given adequate opportunity to propose alternative forecasts based on the data and some specific material elements of the consultant's report, including the decision to apply an unweighted average between the trace for known projects included according to the ISP Methodology, and the projection based on economic development. AEMO's addendum to the 2025 IASR answers with its own account of the refinement: Key refinements included adjusting the scenario inclusion criteria to provide a bespoke application for data centres, reflecting the significant uncertainty and scale of proposed projects, but also their rapid development to connection. On process it is equally plain: It was AEMO's assessment at the time that the level of consultation was adequate given the significant engagement on the topic by stakeholders as part of the three FRG sessions. The criticism was read twice, in the AER report and in AEMO's reproduction of it, and the two wordings differ slightly — the report refers to the description of content in the IASR, the addendum's quotation to the description in the IASR. The addendum's own publication date is not stated in the text obtained.

Across the eight AEMO and AER documents cited here, none makes a sworn or certified attestation by a proponent about construction, financing or in-service date the condition on which a data-centre project's classification rests. That is a statement about eight documents, not about every instrument AEMO publishes. No project-level list of the 225 data centres, no connection register and no network service provider submission was read for this piece, so nothing here is a finding about a named developer or project.

The methodology's 25 August 2026 publication date comes from AEMO's guidelines page as reported by retrieval rather than from a date printed in the text obtained, and page and figure numbering, including the Figure 9 note, are unconfirmed.

What these documents show is a forecast whose most volatile component is admitted by grade rather than measured, graded by the same body that publishes the forecast and reports on reliability, and disclosed with unusual candour: AEMO prints the dropouts, the utilisation rate and the ramp. The candour is what keeps this from being a story about inflation. It is a story about the interface. A reader handed 67 GW and a class called anticipated cannot tell how much of that capacity rests on an application filed inside a three-year window and how much on a signed contract, because the class name is doing the disclosing and the class name is defined two ways. AEMO's record sits in the BTW directory entry for the operator, and the fix this piece points at is not a smaller number but a legible one: print the rung beside the megawatt, put the three-year band in the same table as the pipeline, and let the attrition that AEMO already reports land on a classification a reader can check.