Australia’s data centre boom is a power problem before a property one

Photo: Antti Leppänen via wikimedia (CC BY 4.0)
A hundred megawatts. That’s the number I keep hearing in planning meetings now, and five years ago it would have been a typo. When I was doing fit-outs along the M7 the big campus builds were sized at 20 or 30 megawatts across several stages, and the electrical engineers who worried about the grid connection were a small, slightly anxious group at the end of the table. They’re at the head of the table now. The land is the easy part.
The short version is this. Australia is building data centre capacity faster than at any point in the industry’s history, most of it concentrated in two corridors, western Sydney and Melbourne’s west, and the thing that decides whether a site gets built is no longer whether the developer can buy the block. It’s whether the distribution network can deliver the power, and when.
Why the number got so big #
For twenty years a data centre rack drew somewhere between four and eight kilowatts, and the building was designed around that. The racks that hold the graphics processors used for training and running large AI models draw a great deal more. Depending on whose hardware you’re installing, a single rack can now sit at 40 kilowatts, 80, or well past 100 with liquid cooling. The floor area hasn’t changed. The power density per square metre has gone up by an order of magnitude, and the cooling plant has had to follow it.
That change arrived at the same time as a general surge in cloud demand, so the operators that were already planning a next stage found themselves planning a stage three times the size, and the hyperscalers that lease from them wanted it faster. The Australian Energy Market Operator’s Integrated System Plan, which is the closest thing the country has to a long-range electricity forecast, now treats data centre load as a category worth calling out on its own. It didn’t in the editions I read as a graduate.

The corridor problem #
Sydney’s data centre industry clusters in a band running from Eastern Creek through Erskine Park and Huntingwood, for a reason that’s obvious when you drive it: it’s where the industrial land, the fibre routes and the big transmission assets already meet. Melbourne’s equivalent is the west, around Truganina and Derrimut, with the older sites closer to the city. Both corridors were chosen a decade ago because the power was there. The trouble is that “the power was there” meant 30 megawatts of headroom at a zone substation, and the requests landing now are for three times that per campus.
A zone substation is not a thing you add capacity to over a weekend. The distribution businesses (Endeavour Energy across most of the Sydney corridor, Ausgrid closer in, Powercor and Jemena in Melbourne’s west) plan their augmentations years ahead and recover the cost through regulated pricing that the Australian Energy Regulator approves in five-year blocks. A developer that turns up in year two of a five-year determination asking for a new 132 kV connection is asking for something that wasn’t in anyone’s plan. It can be done. It is not quick, and a queue forms.
I’ve seen that queue from the inside, and I’ll admit the sensible thing to say is that it’s a solvable engineering problem with a long lead time. The less sensible, more honest thing to say is that the operators who secured their connections in 2021 and 2022 are now sitting on an asset worth more than the land under it, and some of them know it.
What it costs, roughly #
People ask me what a megawatt costs and the true answer is “connected to what, and when”. A new high-voltage connection with a dedicated substation on site can run to tens of millions of dollars before a single rack is installed, and the developer usually funds it up front as a capital contribution, sometimes with a partial refund if other customers later connect to the same asset. On top of that sits the ongoing network charge, which for a large customer is negotiated rather than published, and then the energy itself, which is the part everyone talks about and is, in my experience, the smaller worry.
The energy price matters less than most people assume because the large operators contract most of their supply years out, often through renewable power purchase agreements that they also need for the carbon commitments in their tenants’ leases. What they can’t contract their way out of is the physical connection. You can sign a PPA for a wind farm in the Riverina tomorrow. You cannot sign your way to a transformer that hasn’t been built.
How a connection actually gets approved #
Since I get asked this by property people who’ve never done one, here is the sequence, stripped of the acronyms. The developer lodges a connection enquiry with the distributor stating the load, the timing and the site. The distributor runs a study on whether the local network can take it, which for anything above a few tens of megawatts almost always comes back with “not without augmentation”. The two sides then negotiate a connection offer that sets out what gets built, who pays, and when. For a load big enough to need its own transmission-level connection, the transmission business (Transgrid in New South Wales, AusNet in Victoria) joins the conversation and the timeline stretches again, because a new 132 or 330 kV connection point is subject to the Australian Energy Market Operator’s registration and technical requirements as well as the distributor’s.
The rules that govern all of this sit in the National Electricity Rules, which the Australian Energy Regulator enforces, and the parts of those rules that deal with large customer connections were written for aluminium smelters and mines, not for a building that wants 100 megawatts by Christmas and 300 by the one after. The regulator and the market bodies know this. There’s been a steady stream of consultation papers on connection reform, and some of it has landed. Most of what a developer experiences today, though, is still the old process with more patience applied.
The practical advice I give anyone entering this market is simple and unglamorous: lodge the enquiry before you buy the land, not after, and assume the answer will be measured in years. The operators that are announcing campuses today with power already secured did that work in 2021. The ones announcing campuses with “power to be confirmed” are announcing land.
The water question is coming next #
Liquid cooling at the rack has, oddly, made the water conversation both better and worse. Better, because a closed loop to a chip doesn’t consume water the way an evaporative cooling tower does. Worse, because the heat still has to be rejected somewhere, and in a Sydney summer the cheapest way to reject a lot of heat is still evaporation. Sydney Water and its Melbourne counterparts are starting to ask the same questions the electricity distributors asked five years ago, and the operators who can show a low water usage effectiveness figure will find their development applications easier. The NABERS Energy rating for data centres, which is a real and audited thing rather than a marketing claim, is going to matter more than it has, and I’d expect a water equivalent to follow.
Where this goes #
Three things, I think, and I’ll hold myself to them.
The first is that the next corridor won’t be a corridor. It’ll be a substation. Developers are already looking at sites that would have been considered too far out in 2019, in the Hunter, in the Latrobe Valley, around the old power station precincts where transmission capacity exists precisely because a coal plant used to use it. Latency to the city is a real constraint for some workloads and irrelevant for the training runs that are driving the biggest builds, and the industry is starting to separate the two.
The second is that on-site generation, which every operator has as a diesel backup and nobody wants to run, is going to be argued over. Gas turbines behind the meter, batteries sized for grid services rather than just ride-through, and eventually the small modular reactor conversation, which I’ll believe when I see a connection agreement. The distributors will resist anything that looks like a customer bypassing the network. The customers will point at the queue.
The third is that the planning system hasn’t caught up. A data centre is assessed, in most states, as a large industrial building with an unusual electrical load. It’s really a piece of energy infrastructure with some computers in it, and the states that work that out first, and plan the transmission alongside the zoning, will get the next decade of investment. New South Wales has made a start. Victoria is talking about it. Queensland, which has the transmission and the land, has been quieter than it should be.
I rode out past the Eastern Creek sites on Sunday, which is the kind of thing I do for fun and my partner does not understand. From the road you see the sheds, and the sheds look like every other shed on the M7. You don’t see the switchyard behind them, and the switchyard is the whole story.
— Declan Ferris, Infrastructure Correspondent, Technology Spectator