NextFin News - Australia's data centres are projected to consume roughly seven times more electricity by 2036 than they do today, a demand shock that collides with a national grid already racing to retire coal-fired capacity and build renewable generation faster than at any point in the country's energy history. The Australian Energy Market Operator's latest planning outlook puts data centres on a trajectory from about 4 terawatt-hours of grid-supplied power in 2024-25 - roughly 2% of the National Electricity Market - to a share approaching 10% of underlying demand by 2050, four times today's level and the equivalent of one-fifth of the entire grid's current consumption.
The projection reframes artificial intelligence from a technology story into an energy-infrastructure story, and it raises the question the market has not fully priced: when a single new class of customer can double the growth rate of national electricity demand, who pays for the firming capacity that keeps the lights on when the sun is not shining and the wind is not blowing?
The Scale of the Demand Shock
The numbers are large enough to move a wholesale market. Under the grid operator's central "Step Change" scenario, data-centre consumption grows at an average annual rate of 25.1%, reaching around 12 TWh by 2029-30 - about 6% of grid-supplied electricity - before climbing to roughly 34.5 TWh by 2050, or 12% of the NEM. The 2026 Integrated System Plan, published in July, identifies data centres supporting AI and cloud services as one of the two largest new sources of demand this decade, alongside hydrogen production at 35 TWh. Total underlying NEM consumption is forecast to nearly double from 205 TWh today to around 390 TWh by 2050, with business and industrial grid demand doubling from 140 TWh to 280 TWh while household net grid consumption falls as rooftop solar, batteries and electric vehicles spread.
"Data centres have emerged as potentially significant consumers of electricity, with their development exceeding what had been forecast in previous ISPs," the plan states.
The pipeline is already moving. At the end of the March 2026 quarter, 11 large-scale projects above 5 megawatts, representing 5.4 GW of maximum demand, were progressing through the transmission connection process - about 60% in New South Wales and 40% in Victoria, most in early stages. Australia now has more than 160 operational data centres, almost half clustered in Sydney with the remainder concentrated in Melbourne, Brisbane and Perth. In New South Wales alone, 44 data centres totalling 11.4 GW sit in the development pipeline as at 31 March 2026 - equivalent to nearly four times the capacity of Eraring, the country's largest coal-fired station at 2,880 MW.
That pipeline is where the first tension appears. Network service providers have received connection requests totalling 44 GW from data-centre proponents. If all of those requests ran continuously at requested capacity, they would approximate current global data-centre consumption. The grid operator's own commissioned analysis warns that applications are being lodged for projects that may never materialise - so-called phantom demand - meaning the headline queue materially overstates likely draw. The practical implication is that planning has to be done against a range of outcomes, not a single line.
Cyclical Surge or Structural Regime Shift
The central analytical question is whether this demand growth is cyclical - a build-out wave that will flatten once capacity is installed - or structural, a permanent step-change in the load profile of the Australian grid. The evidence points decisively to structural, for three reasons.
First, the driver is a technology transition, not an inventory cycle. AI-optimised servers consume far more power per unit of compute than conventional servers, and their adoption is compounding rather than peaking. Gartner projects global data-centre electricity consumption will rise 26% in 2026 to 565 TWh from 447 TWh in 2025, with AI-optimised server consumption jumping from 95 TWh to 175 TWh in a single year and overtaking conventional servers globally by 2027. In Australia, AI-optimised servers are expected to account for 35.7% of data-centre power in 2026 - 2.2 TWh - higher than the 31% global share. A cyclical wave reverses when inventories clear; a compute-intensity transition reverses only if AI adoption itself reverses.
Second, the load profile is fundamentally different from historical commercial demand. Data centres run 24 hours a day, 365 days a year, at high capacity factors, and they require near-perfect reliability - the industry's "five nines" standard. That makes them poor candidates for demand response and poor matches for variable renewable output without firming. A traditional office tower absorbs solar at midday and falls away at night; a hyperscale facility does not. The grid must therefore build not just energy but firm, dispatchable capacity - batteries, pumped hydro, gas peakers - sized to a load that never sleeps.
Third, the commitments are already being made. The Clean Energy Finance Corporation's Baringa-commissioned analysis, published in December 2025, projects data-centre capacity rising from 1.35 GW today to between 4.7 GW and 7.4 GW by 2035 - roughly a fourfold increase within a decade - with investment of A$85 billion to A$135 billion over the period. Capital of that scale, once sunk into sites, cooling systems and grid connections, does not unwind quickly. Even under a conservative demand path, the load already connected and under construction is locked in for the better part of two decades.
There is a cyclical leg layered on top - the timing of individual project commissioning, the pace of AI model deployment, the quarterly cadence of GPU shipments - and that leg will produce years that undershoot and years that overshoot the central path. But the floor has moved up permanently. The right framing is a structural regime shift with cyclical noise around the ramp, not a boom that will mean-revert.
"Data centres are emerging as a major driver of electricity demand growth that will have a key influence on our power system. Our updated methodology ensures we're capturing this transformation with the transparency and rigour it requires," said Andrew Turley, the grid operator's Group Manager of Forecasting.
The Second-Order Price Shock
The first-order effect of data-centre growth is straightforward: more load requires more generation. The second-order effect is where the market impact concentrates, and it runs through the wholesale price stack. Because data centres demand firm, always-on power, every additional megawatt-hour of data-centre load that is not matched by new renewable and storage capacity is met at the margin by gas-fired peaking generation. That pushes the marginal clearing price up for every other customer on the system.
The CEFC-Baringa modelling quantifies this. Without additional offsetting renewable generation, data-centre growth could lift wholesale electricity prices by 26% in New South Wales and 23% in Victoria by 2035 under a central scenario, driven primarily by increased reliance on gas peaking. The same analysis finds that adding 3.2 GW of renewable energy capacity and 1.9 GW of battery storage by 2035 to support the data-centre load would contain those price rises and neutralise the additional emissions. The asymmetry is stark: matched build keeps prices contained; a miss on renewable and storage delivery transfers billions in higher marginal costs to households and industrial users.
This is the already-priced conventional wisdom screen, and it does not clear. The market has largely priced the build-out of data centres themselves - the developers, the REITs, the equipment suppliers. What is not fully priced is the optionality value of firming capacity and the regional dispersion of the price impact. A state that approves renewable and transmission projects quickly captures the load and the associated grid investment; a state that delays imports the price impact without the capital. The beneficiaries are therefore not only the data-centre owners but the transmission builders, the battery developers and the renewable generators that can contract into 24/7 offtake - and the exposed are the legacy retailers and the industrial customers in states where firming lags load.
The transmission channel runs further still. If data centres absorb a growing share of firm capacity, the residual grid becomes more dependent on weather in real time, raising the value of flexibility - demand response, virtual power plants, interconnector capacity. That reprices the entire flexibility stack, not just the generation stack.
The Counter-Thesis: Phantom Demand and the Efficiency Escape Hatch
The strongest case against the sevenfold trajectory is that it confuses connection requests with realised load, and hardware intensity with lasting intensity. The 44 GW connection queue is the exhibit A: it exceeds Australia's entire current grid demand many times over if taken at face value, and the grid operator's commissioned Oxford Economics analysis explicitly models "phantom demand" - applications lodged speculatively, for sites that will not be built, or for capacities that will never be drawn. In New South Wales, the industry's own estimate is that only 1.2 GW of additional data-centre load will come online in Sydney by 2030, against a pipeline of 11.4 GW. If that pattern holds nationally, the sevenfold projection overshoots realised demand by a wide margin.
The second prong is efficiency. Server efficiency has historically improved fast enough to bend the power curve even as compute exploded, and the grid operator's own scenarios include a "Progressive Change" path under which data-centre demand reaches only 8.5 TWh by FY30 - 29.7% below the Step Change case. Water constraints, community opposition to large sites, and rising connection costs all act as brakes that no single forecast captures cleanly. A counter-thesis built on these grounds does not require AI adoption to fail; it requires only that speculative queues deflate and efficiency gains continue at historical rates.
Both points are valid, but neither overturns the structural call. Phantom demand affects the timing and the peak of the ramp, not the direction - the 5.4 GW already in the transmission connection process is real, permitted capital, and the 1.35 GW operating base is already drawing power. Efficiency gains reduce the power per unit of compute, but total compute demand is growing faster than efficiency is improving, which is why global consumption is still projected to double by 2030 even with efficiency baked in. The counter-thesis argues for a lower multiple, not for no multiple.
The falsifying signal is specific: if data-centre electricity consumption in the NEM is below 8 TWh in FY30 - that is, below the Progressive Change scenario rather than merely below the Step Change case - the structural regime-shift call is wrong and the demand wave should be treated as cyclical. A print at or above 12 TWh confirms the central path. Watch the FY30 actual, reported by the grid operator in its Electricity Statement of Opportunities.
Policy Response: Make Data Centres Pay Their Way
The regulatory system is moving to ensure the new load does not socialise its costs. On 5 August 2026, the Australian Energy Market Commission advised energy ministers that data centres should bring their own clean, firmed energy, operate flexibly and connect efficiently, so that other electricity consumers are not worse off as the facilities connect to the grid. Ministers had agreed in May that data centres should offset their electricity demand with renewable generation and firming and report transparently on energy use; the ministerial council welcomed the advice in July and agreed to progress regulatory arrangements, including a guarantee-of-origin scheme.
"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. That's exactly what this advice is designed to do," said Anna Collyer, chair of the market commission.
The commission is also progressing a rule change to update technical access standards for large inverter-based loads like data centres, with a final determination due in late October, alongside measures to close gaps in how data centres pay for the network upgrades their connections trigger. The intent is explicit: the businesses building large loads should bear the costs and risks they create, not the households and small businesses on the same grid. That shifts the investment case - data centres that can pair with firm renewable capacity keep their economics; those that cannot will face the full marginal cost of a coal-retiring, gas-constrained system.
What Comes Next: Beneficiaries, the Exposed, and the Scenarios
Cashing the mechanism into impact, the beneficiaries are the owners of dispatchable and renewable capacity that can contract into firm offtake: battery and pumped-hydro developers, transmission builders, and renewable generators with land and grid access in the load corridors of New South Wales and Victoria. The exposed are legacy retailers without long-dated hedge cover, industrial customers in states where firming lags, and the emissions trajectory itself - a data-centre load met by gas peakers keeps fossil capacity online longer than current retirement schedules assume.
Split by time horizon, the picture diverges. In the short term - the next two to three years - the story is liquidity and connection queues: projects that can secure firm power get built, the rest stall, and volatility clusters around commissioning dates. In the medium term - to 2035 - the story is the price stack: whether the 3.2 GW of renewables and 1.9 GW of storage identified as sufficient actually get permitted and connected before the load arrives. In the long term - to 2050 - the story is structural: a grid whose largest marginal customer never switches off, and whose planning framework has been rewritten around always-on load.
Three scenarios frame the path. The base case tracks the Step Change trajectory: 12 TWh by 2030, prices contained where matched renewable and storage build keeps pace, and data centres settling at roughly 10% of underlying NEM demand by 2050. The upside case - for prices, not profits - is the Higher Demand sensitivity: an additional 39 TWh of demand by 2050 concentrated in South Australia, Sydney-Newcastle-Wollongong, Melbourne-Geelong and Gladstone, which would force a faster, costlier build and lift clearing prices above the contained path. The downside case is the phantom-demand unwind: queues deflate toward the 1.2 GW Sydney-realisation pattern, efficiency outpaces deployment, and realised load lands near the Progressive Change scenario at 8.5 TWh by 2030.
The signal that separates base from downside is the FY30 print against the 8 TWh threshold. The signal that separates base from upside is the pace of transmission and storage approvals in the load-heavy states - specifically whether the 5.4 GW in the connection queue converts to energised capacity on schedule.
Australia's grid was built for a century of predictable, weather-independent load. The data-centre decade ends that assumption, and the price of the transition will be set not by how much compute the country hosts, but by how fast it can build the firming capacity to host it without passing the bill to everyone else.
Explore more exclusive insights at nextfin.ai.
