How much new data-centre load can Australia's grid carry without new supply?

Data centres are among the NEM's fastest-growing loads, and new ones are expected to bring their own supply

One base scenario, one sweep file, 25 full-year runs

A NEMSweep study starts from a base scenario, one file that describes the grid, and a sweep, a short second file that lists the values to try for one or more of its inputs. Each value becomes its own scenario, run for a full year, and the results line up side by side.

1. Describe the grid

   The baseline is the NEM for FY2025-26: five regions with the generation and storage each region declares, the interconnectors between them, hourly demand and one typical weather year per region.

   The base scenario points to each region's hourly demand and weather files, which you provide. The ones used here are in the ready-to-run NEM dataset (opens in a new tab); for another grid, set up a custom grid ↗.

   {
     …
     "regions": [
       {
         "regionId": "NSW1",
         "demandFile": "demand-nsw1.json",
         "weatherFile": "weather-nsw1.json",
         "generatingFleets": [
           { "technology": "Solar", "nameplateCapacityMw": 5356.2, … },
           …
         ],
         "storageFleets": [
           { "technology": "Battery", "initialEnergyCapacityMwh": 5516.7, … },
           …
         ]
       },
       …
     ]
   }

2. Choose what to vary

   Here that is one field, the always-on data-centre load in each region, at 25 values from 0 to 12,000 MW. The sweep file sets each region's share of the load at each point.

   {
     "pointId": "p1",
     "axisValue": 500,
     "label": "+500 MW",
     "overrides": {
       // dataCentreNameplateMw adds a fixed load, drawn every hour of the year,
       // on top of the region's hourly demand from the base scenario.
       "regions": [
         { "regionId": "NSW1", "dataCentreNameplateMw": 200 },
         { "regionId": "QLD1", "dataCentreNameplateMw": 100 },
         { "regionId": "SA1", "dataCentreNameplateMw": 100 },
         { "regionId": "VIC1", "dataCentreNameplateMw": 100 }
       ]
     }
   }

3. Run and compare

   Each scenario is dispatched hour by hour for a full year: the model decides which generators and batteries run in each hour. Where a region falls short of the reliability target, the model grows that region's battery storage until it passes; it never adds generation or transmission. Cost, reliability, storage, emissions and renewable share are recorded for every run.

   What the model does in each scenario

   Every scenario runs through the same loop until the target is met, or more storage can no longer help.

   Scroll sideways to see the whole loop.

   The storage sizing loop the model runs for every scenario. More storage stops helping when adding it no longer reduces unmet demand, when the year has less generation energy than demand, or when the search reaches the storage limits the scenario sets.

Cost

System levelised cost: what it takes to build and run the whole system per MWh it delivers, in 2026 Australian dollars at a 7% real discount rate (the return capital is assumed to need, after inflation). A system cost, not a retail price.

Reliability target

At most 0.002% of demand unmet, the level of the NEM reliability standard.

Storage

Power capacity in MW (how fast it can charge or discharge) and energy capacity in MWh (how much it can hold). Figures are batteries: about 16,000 MWh of existing pumped hydro is held fixed in every scenario and left out of them. Results are for the whole NEM unless a figure or table says otherwise.

From 500 MW the grid needs more batteries; from 5,000 MW the battery search stops closing the gap

Wind and solar produce when the weather allows, not when demand needs it, so the grid relies on firming to move energy into the hours when output is low. This sweep adds the data-centre load as a flat draw, the same in every hour, so each step leaves more hours short. With nothing added, the batteries the scenario declares are enough. From 500 MW they are not, and the model adds more.

Between 3,500 MW and 4,000 MW the battery energy capacity needed nearly doubles from one scenario to the next, from about 46,000 MWh to about 85,000 MWh. From 5,000 MW, adding batteries no longer brings the grid back to the target, and the runs beyond it report where the sizing search stopped rather than what meeting the target would take. Generation and transmission were held fixed, so this sweep cannot say which of them is the next constraint. That is a question for the next sweep.

Cost per MWh falls to 3,500 MW, but the annual bill rises at every step

Cost per MWh delivered falls from $148 to $141 up to 3,500 MW, because the fixed cost of the existing fleet is spread over more energy and some wind and solar output that was being curtailed now finds a use. From 3,500 MW to 4,000 MW the storage part of that (batteries plus existing pumped hydro) cost rises by $4 per MWh while the generation part falls by $1, and the total turns up.

The whole annual cost rises at every step, from about $28 billion to about $43 billion a year. Read on its own, cost per MWh understates what happens past 5,000 MW: it puts no cost on the demand left unmet, so read it with unmet demand and the annual cost beside it.

Past 5,000 MW, cost per MWh reaches $153 at 12,000 MW. Storage's part of each MWh (batteries plus existing pumped hydro) goes from $8 at 3,500 MW to $32 at 12,000 MW, because the model keeps adding batteries that no longer bring the grid back to the target.

At 12,000 MW the grid leaves 4.7 per cent of demand unmet, across about 5,000 of the year's 8,760 hours.

Emissions rise 56 per cent as gas and coal supply most of the extra energy

Grid-scale renewable share is wind, solar and hydro as a share of all generation delivered. Rooftop solar is netted out of demand before the model sees it, so the share is not comparable with a rooftop-inclusive target.

Six simplifications make each turning point approximate

Every model simplifies. Some of the simplifications here understate what the grid would need, and some overstate what it costs. The sweep samples 25 load levels, and the simplifications below can shift where results change, so treat each turning point as approximate. Read each number alongside the scenarios either side of it.

• No supply or backup of their own

  Here the data centres add load only. In practice they are expected to secure new supply (Department of Industry, Science and Resources, 2026), and sites keep backup generation for outages, most commonly diesel (Oxford Economics Australia for AEMO, 2025). This sweep models neither.

  Effect: overstates the strain on the grid

• Data-centre load is flat

  Each scenario adds the same load in every hour of the year. Real facilities do not draw a constant load, so this simplifies both the peaks and the quiet hours.

  Effect: could move results either way

• One typical weather year

  Each region uses one typical year. Hard years, such as a long wind drought, are missing, so the storage figures probably understate what a harder year would need.

  Effect: probably understates the batteries needed

• Batteries only, with no foresight

  Storage charges and discharges by a simple rule, not a plan. The search that sizes it adds batteries in steps, up to the limits the scenario sets, so where it stops it has not proved that no battery would do. The model never adds generation or transmission.

  Effect: could move results either way

• System cost, not a power bill

  Build and run costs only. Network, retail, scheme and tax costs are not included.

  Effect: leaves out costs a bill would include

• Transmission costed on the high side

  Each direction of an interconnector is charged for the full line length, so the transmission part of cost is overstated.

  Effect: overstates the transmission cost

The effects are indications of direction, not measured corrections. A high-fidelity study could give a different answer.

Quote the 4,500 MW limit with the batteries it takes and the basis it was run on

On these inputs, the NEM meets the reliability target to about 4,500 MW of added data-centre load, and needs about seven times its declared battery energy capacity to do it. Beyond that it misses the reliability target, and more batteries stop closing the gap.