How fast do Australian data centre projects reach operation?

50% of the phases entering the record reached operation within four years; once construction started, 94% did, in a median of 17 months

Of the phases entering inashanu's delivery record, 50% reached operation within four years; once construction started, 94% did, in a median of 17 months, and first openings ran a median of about 50 days behind their stated targets (inashanu delivery record, as at 21 October 2026). A forecast made at mid-2021 scores no better than a flat guess and erred low, and NEXTDC's and CDC's contracted books need about 1.0 GW of new build by mid-2030, as at 21 October 2026.

Completion by starting point, competing outcomes counted

Completion incidence (Aalen-Johansen) counts withdrawal, cancellation and refusal ("withdrawn") as competing outcomes; the net median is the Kaplan-Meier duration with those outcomes censored. "Open" phases are still in progress at 21 Oct 2026. Entry is the earlier of announcement and lodgement.

What the record is, and why the model looks beyond it

The delivery record

Dated milestones for each phase of the Australian campuses that operators' disclosures and planning records describe: announcement, lodgement, approval, construction start and operation, with every withdrawal, cancellation and refusal. Phases still in progress are censored at the evidence date.

Why the model calibrates beyond it

It is a selected sample: phases of campuses that exist, mostly the large operators'. No approved phase has been withdrawn in it, so the approval cohort completes 100%. The model takes durations from the record and calibrates completion odds to population evidence (Section 01).

Completion builds slowly from entry and quickly from construction: 22% of entering phases are complete at two years, against 70% of construction starts

Share of each cohort in operation, by months since the cohort began

Incidence of reaching operation (Aalen-Johansen), in per cent. Pointwise 95% bands from a 1,000-replicate phase bootstrap are in the workbook, licensed on request (the three Followup sheets); they describe this sample, not the market.

At risk and complete, by month

Reading the curves

• Entry: 50% complete at four years and 70% by the end of the record; net median 41 months.
• Approval: 65% at four years; no withdrawal on record, which says more about the sample than about approvals.
• Construction started: 94% complete, in a median of 17 months.

NEXTDC's and CDC's contracted books need about 1.0 GW of new build by mid-2030; at the median the model delivers 556 MW from NEXTDC and 810 MW from CDC by end-2030, above both

Disclosed delivery, new IT by period (MW)

The books against what the central case delivers by end-2030 (MW)

NEXTDC: binding forward orders of 555 to 565 MW (30 Jun 2026; the range is its Malaysian site), less 103 to 113 MW of built vacancy, leave 452 MW to build, so its 197 MW FY27 conversion is 84 MW of new build. CDC: its Australian 555 MW contract, up to 750 MW (the group's 1.1 GW contracted less 350 MW deployed, New Zealand included). AirTrunk publishes no book. Delivered counts all new IT, contracted or not.

Days late against the stated target, 15 first openings

One opening per campus with a stated target period; a bar where the opening is known only to a period, an open circle where it came on or before the target. The median is about 50 days; 12 of 15 opened late. The model samples this lateness for every disclosed schedule.

M3

NEXTDC now expects M3's next increase in grid supply in FY28 (21 Oct 2026). M3 puts no new IT into service before a date drawn evenly within FY28, and capacity due earlier waits for it; the completion penalty and the sampled lateness still apply.

Scored against what happened by end-2025, a forecast made at mid-2021 reaches a Brier score of 0.221, level with a flat 57% guess: no demonstrated skill yet

Forecast against outcome, by status at mid-2021

Mean forecast is the model's probability of operation by end-2025 from the calibration known at the cut-off; completed counts the projects found operating, in whole or in part, by 31 Dec 2025.

How the test is run

• Strict cut-off. Only sources dated on or before 30 Jun 2021 enter: 16 rows published later are excluded, and 2 more that are not separate facilities. The other 49 rows describe 41 projects, and nothing learned after the cut-off informs the forecast.
• Outcomes from dated sources. Each project is read at 31 Dec 2025 from a dated operator, planning or press record; 15 outcomes are still unknown and left unscored.
• A small sample. The difference from the flat reference (0.0004) is too small to establish skill; the test exists to be repeated as outcomes arrive.

What the score says

A Brier score is the mean squared gap between a forecast probability and the outcome: 0 is perfect and a flat 50% scores 0.25. At 0.2211 against 0.2215 the mid-2021 method ranks projects no better than one flat probability, and it errs low: 19 of the 26 known projects reached operation (73%) against a mean forecast of 57%.

On capacity the backtest can match only four projects, and a national error cannot be computed: the mid-2021 operating stock is not on the public record

Matched projects: forecast against delivered IT, MW by end-2025

Forecast is the mean simulated MW for the project at end-2025 from the mid-2021 calibration; delivered is the dated operating figure. Mean error −11.8 MW; mean absolute error 11.8 MW.

What the backtest cannot yet validate

• The national stock. No public source gives Australia's operating IT at mid-2021, nor the additions outside the register as it then stood, so the national error is left blank.
• The 2026 calibration. The population evidence and the build-rate panel the model now uses were not available in 2021, so the backtest checks the method as it stood then.
• Large campuses. The four matched projects are small or mid-sized; none is a gigawatt campus.

What would make the next backtest decisive

• A dated starting stock. DC Byte puts national operating IT at 1,400 MW at end-2025, a stock the next test can start from.
• More outcomes. The 15 projects still unresolved will add scored outcomes as they settle.
• Large campuses. Gigawatt campuses entering the record would test the build-rate assumptions directly.

Reading the errors

All four matched projects delivered more than forecast, by 1.5 MW to 23.7 MW: the 2021 method ramped these campuses too slowly, the same direction as its completion forecasts on page 17.

Sources

1. inashanu delivery record, as at 21 Oct 2026
2. inashanu market model, as at 21 Oct 2026
3. NEXTDC FY26 results, as at 27 Aug 2026; release, 21 Oct 2026
4. Infratil and CDC Investor Day, as at 16 Sep 2026
5. Bouygues Construction, as at 30 Jul 2026

Related

• How much data centre capacity will Australia have in 2030?
• How much of the register's data centre pipeline will be operating by 2030?
• What is in Australia's data centre development pipeline?
• How is the Australian data centre capacity forecast built?

The full primer is published as a 103-page PDF (download). The market model and the register behind it are licensed to firms on request; the study page lists what is published and what is licensed.