Can Australia’s Power Grid Carry the AI Data-Centre Boom?

A project must secure a viable site, grid connection, dependable electricity, network upgrades, finance, customers and enough utilisation to cover its fixed costs. Each gate can fail on a different timetable.

That distinction is becoming more important. The Australian Energy Market Operator’s 2026 Electricity Statement of Opportunities forecasts electricity used by data centres in the National Electricity Market rising from about 5 terawatt hours in 2025–26 to 34 terawatt hours in 2035–36. Their share of grid-supplied electricity would rise from roughly 3 per cent to about 13 per cent. The opportunity is therefore large enough to matter to generation, storage, transmission and consumers. It is also too large to analyse as a simple property-development story.

Data cut-off: 28 August 2026.

The grid question is different from the AI debt question

The first question for an AI data centre is whether customers will pay enough for computing capacity. The second is how the project is financed. Those issues are covered in the companion analysis of AI data-centre debt and off-balance-sheet risk .

Australia adds a separate, local question: can the electricity system connect and serve the load at the required place and time, and who pays for the extra capacity?

A data centre can have a creditworthy sponsor and still wait for transmission, a substation, system studies or new generation. It can receive planning approval without an executed connection agreement. It can reserve a large connection capacity but ramp its actual use slowly. Conversely, a project that secures dependable power and a committed customer may create durable demand for networks and generators even before the wider AI industry becomes consistently profitable.

The investment case therefore needs two linked but distinct tests. The digital test asks whether computing revenue covers the building, chips and financing. The power test asks whether the project can obtain reliable electricity without imposing an uneconomic bill on the operator, network or other users.

Planning approval is only the first visible gate

Planning decisions are easy to report because they are public milestones. They establish that a proposed land use can proceed subject to conditions. They do not prove that a data centre is connected, energised, financed or occupied.

The practical sequence is longer:

• Planning and land: the project has an approved site, water and environmental pathway.

• Connection: the network studies the load, available capacity, technical requirements and upgrade work.

• Dependable power: generation, storage, transmission and contracts can serve the load when it is needed, not only when renewable output is abundant.

• Construction and finance: committed funding covers the build and any delay or cost overrun.

• Customer commitment: contracted demand has credible duration, pricing and termination terms.

• Utilisation: occupied capacity ramps fast enough for revenue to absorb fixed operating and financing costs.

The data-centre proposal at Long Reach in Tasmania is useful because it exposes this sequence. A local approval is evidence about the first gate. Local approval alone does not establish a grid connection, a power-supply contract, financing or committed customers, so it is not proof of operating cash flow or a conclusion about national electricity reliability.

Connection capacity is not the same as electricity consumed

AEMO’s latest forecast explicitly distinguishes connection capacity from actual use and allows for projects to drop out during the connection process. It also models a gradual demand ramp rather than assuming every announced megawatt appears at once.

That is an important defence against two opposite errors. Adding every proposal can exaggerate future demand, because some projects will be delayed, resized or cancelled. Ignoring the pipeline can understate the infrastructure task, because even a smaller completed share can become a material new source of around-the-clock demand.

Investors should therefore keep at least three numbers separate: announced capacity, approved or contracted connection capacity, and metered consumption. The first describes ambition. The second describes a more advanced technical and commercial position. The third shows realised load. A forecast should also state when each project is expected to ramp and how much attrition it assumes.

Reliability depends on timing, not just total megawatts

AEMO says reliability can be maintained if investment arrives on time. Its 2026 assessment reported no forecast reliability gaps before 2030, supported by 9.1 gigawatts of new generation and storage connected in 2025–26 and a 40-gigawatt pipeline of committed and anticipated projects expected by the early 2030s. It also forecasts about 15 gigawatts of coal and gas generation retiring over the next decade while total electricity consumption rises by more than 40 per cent.

Those figures do not support a simple “grid crisis” claim. They support a timing test. New demand, retirements, generation, storage, transmission and essential system services must arrive in a workable order. The broader analysis of Australia’s electricity-transition timing risk sets out that delivery sequence.

Solar and wind can provide low-cost energy, while storage, transmission, flexible demand and other dependable resources help match supply to the data centre’s operating profile. The relevant question is not whether one technology can power a facility in isolation. It is whether the contracted portfolio and local network can supply the required quality and continuity at a competitive delivered cost.

This is why a project can be valuable to the power system in one location and expensive in another. Available transmission, congestion, electricity lost in transmission, nearby generation, network upgrades and the ability to curtail or shift load all change the economics.

Who pays determines who earns the return

Large new loads can support investment in generation and networks, but demand growth does not automatically create attractive shareholder returns. The allocation of costs and risks matters.

If a data-centre customer funds a dedicated connection and signs a long contract, more of the utilisation risk may sit with the customer or project. If a regulated network builds shared infrastructure, the regulator must decide which costs enter the regulated asset base, when revenue begins and how much is recovered from the connecting party versus other users. If public support absorbs early costs, taxpayers may carry part of the risk.

For an investor, the revenue opportunity should be tested against five questions:

• Is the spending dedicated to one customer or useful to the wider network?

• Is cost recovery approved, merely proposed or dependent on future utilisation?

• Who bears delay and construction-cost overruns?

• Can the customer reduce, relocate or terminate its demand?

• Does the allowed return exceed the financing and operating cost over the asset’s life?

A regulated asset can offer visible revenue while still destroying value if construction costs rise faster than the allowed return or if regulators disallow part of the spending. A generator can sign a long power-purchase agreement but retain risk from changes in volume, hourly demand patterns, fuel costs or replacement power. A data-centre operator can secure electricity but fail to fill the building.

The AI customer still closes the cash-flow loop

Power is a necessary input, not the final source of value. Ultimately, customers must earn or raise enough cash to keep paying for computing capacity. High utilisation backed by durable contracts can convert fixed infrastructure into recurring revenue. Weak utilisation leaves the same infrastructure with high depreciation, maintenance and financing costs.

This creates a chain rather than a single AI trade:

end-customer revenue → cloud or model-provider cash flow → data-centre contract → project utilisation → electricity demand → generation and network revenue.

A break near the customer can travel backwards through the chain. The effect will not be equal for every participant. A network asset that remains useful to other customers may recover more value than a specialised building in a constrained location. A generator with multiple buyers may be less exposed than a project financed around one tenant. The AI value-chain cash-flow framework is the right companion for that upstream customer test.

What to check before calling a project investable

Project announcements should be converted into evidence that can be updated:

• planning decision and remaining conditions;

• connection agreement, capacity, expected energisation and required upgrades;

• contracted electricity volume, term, price structure and reliability provisions;

• committed construction funding and contingency;

• named or credibly described customers, contract duration and termination rights;

• occupied capacity, metered use and ramp against plan;

• network or government cost contribution and the approved recovery mechanism.

The absence of one item does not prove failure. It identifies the risk that remains with investors, consumers or government. The quality of the disclosure matters more than the size of the headline proposal.

What would change the assessment

The constructive case strengthens if projects progress from announcements to executed connections, financed construction, dependable power contracts and customer commitments; if AEMO’s expected generation, storage and transmission arrive on schedule; and if cost recovery is transparent and matched to the beneficiaries of each investment.

The case weakens if connection dates repeatedly slip, proposed capacity greatly exceeds credible customer demand, network costs are socialised without clear productivity benefits, financing depends on optimistic utilisation, or actual electricity use ramps far below reserved capacity. A wider deterioration in AI customer cash flow would raise that risk even if the Australian grid work remains technically feasible.

Australia can host a valuable data-centre industry without every announced project succeeding. The investable question is not whether AI needs electricity. It is whether each link—from customer revenue to utilisation to delivered power and cost recovery—has moved from proposal to enforceable, financed cash flow.

Sources

• Australian Energy Market Operator — 2026 Electricity Statement of Opportunities

• Australian Energy Market Operator — Digital demand surge

• George Town Council — 25 August 2026 meeting outcomes