The cloud has a postcode. Every chatbot reply, every AI image, every cached video happens inside a real building somewhere. And in Australia, most of those buildings are being built in two states where the grid still runs largely on coal.
A data centre breaking ground today will still be running in 2050. By then, the engineers who designed it will probably have retired. The electricity it draws, and the emissions that come with it, will keep showing up in Australia’s emissions numbers for another two decades.
That’s the quiet bet being placed right now across western Sydney and the outer suburbs of Melbourne, where the global AI boom is meeting an Australian grid that hasn’t quite caught up. The grids in New South Wales and Victoria are about twelve times dirtier than Tasmania’s, and around three times dirtier than South Australia’s. Yet that’s exactly where the next decade of AI infrastructure is heading.
This article uses five interactive charts to map that mismatch. It looks at how the world is locating its AI compute against where water and clean electricity actually exist. It then zooms into Australia, asking where the data centres are clustering, what grids power them, and whether decarbonisation is moving fast enough to keep up with what’s coming.
The question isn’t whether to build AI infrastructure. It’s where, and on what kind of grid. Once it’s built, we’re stuck with the answer.
Before looking at where AI compute is being built, it helps to see how much electricity the world is using to begin with. Global demand has grown steadily since 2000, and data centres are now adding to that load at an accelerating pace.
Figure 1: World electricity demand, 2000–2025 (TWh). The total has roughly doubled in twenty-five years. Hover over the line for year-on-year change.
Source: Adapted from Yearly electricity data, by Ember, 2025 (https://ember-energy.org/data/).
Global electricity demand was around 15,000 TWh in 2000. By 2025 it had roughly doubled (31,747 TWh). Most of that growth came from industrialisation and the electrification of transport. Data centres are a small but fast-growing share of the picture, and AI is what is making the curve steepen.
Sixteen countries host most of the world’s data centres. To see whether they are also the countries running on dirty grids or stressed water systems, switch between the three buttons below. The country order stays the same. What changes is how each one ranks on the metric you’ve chosen. The countries that stay near the top across all three measures are the lock-in problem.
Figure 2: Top sixteen data-centre countries scored on three measures. Country order stays fixed by data-centre count. Switch the metric using the buttons.
Sources: Adapted from Yearly electricity data, by Ember, 2025; Aqueduct 4.0 country rankings, by World Resources Institute, 2023; and Australia data centre market overview, by Cloudscene, 2026.
Try clicking through the three buttons. The country order on the left does not change, only the bars do. What you should see is that the countries running the largest data-centre fleets, like the United States, the United Kingdom, Germany, and China, also have long bars when you switch to grid carbon. Many of them stay near the top when you switch to water stress as well. Australia is mid-pack on facility count but lands well into the upper half on grid carbon, which is the heart of the lock-in problem.
That global picture is uncomfortable wherever you sit, but the question for Australia is more specific. How dirty is Australia’s grid, exactly, in the company of the countries it shares the data-centre map with? The chart below ranks those same sixteen major data-centre countries by grid carbon intensity, cleanest at the top, dirtiest at the bottom. Australia is highlighted in red.
Figure 3: Major data-centre countries ranked by grid carbon intensity, latest year available. Bubble size shows data-centre count. Australia highlighted in red.
Sources: Adapted from Yearly electricity data, by Ember, 2025; and Australia data centre market overview, by Cloudscene, 2026.
Most of the major data-centre countries on this chart run on grids that are at least twice as carbon-heavy as the cleanest performers. France, Brazil and Sweden lead, mostly thanks to nuclear and hydro. The United Kingdom and Germany have made real progress over the past decade and now sit roughly half as carbon-heavy as Australia. China and India sit just below Australia, but the gap is small. Australia is firmly in the dirty half of the table, and well above the world average shown by the dashed line.
Australia’s grid is not really one grid. The five states on the National Electricity Market each run on different fuel mixes, and they are not getting equal shares of the data-centre boom. The chart below maps each state by where it sits on grid emissions, against how many data centres it hosts. Bubble size is annual electricity generation. Colour is local water stress. The story is in the upper right.
Figure 4: Australian NEM states by grid emissions intensity and data-centre count. Bubble size shows annual generation. Colour shows industrial water stress.
Sources: Adapted from NEM regional and national electricity data, by OpenElectricity, 2026; Australia data centre market overview, by Cloudscene, 2026; and Aqueduct 4.0 country rankings, by World Resources Institute, 2023. Excludes Western Australia, which is not on the NEM.
New South Wales hosts most of Australia’s data centres, on a grid still dominated by coal. Victoria’s grid is even dirtier per unit of electricity, with the second-largest cluster. Queensland sits in similar carbon territory but with a smaller data-centre footprint. Tasmania, the lower-left point, runs on hydro and would be the natural fit for clean compute, yet its data-centre count is a fraction of Sydney’s. South Australia is the interesting outlier: a relatively cleaner grid that has built fewer data centres than its capacity could carry. The state-level picture is the same problem the global chart shows, only with sharper edges.
None of this matters if the grid is decarbonising fast enough to absorb whatever AI throws at it. The chart below shows what Australia’s progress has actually looked like since 1999. The grid has cleaned up. The question is whether it has cleaned up fast enough.
Figure 5: NEM-wide annual emissions intensity, 1999–2025 (kgCO₂/MWh).
Source: Adapted from NEM regional and national electricity data, by OpenElectricity, 2026 (https://openelectricity.org.au/). Annual averages computed from monthly data.
Australia’s NEM grid has cleaned up by about 46% over twenty-five years. That works out to less than two per cent per year on average. Global data-centre electricity demand has been growing closer to ten to fifteen per cent per year, and AI is accelerating that. The grid is heading in the right direction. The load it has to absorb is moving faster, in the wrong one. The gap between those two curves is the lock-in problem this article has been about. Where the next decade of data centres get built, on what kind of grid, decides which line wins.
Claude (claude-opus-4-7, Anthropic, 2026) was used as a scaffolding and debugging assistant for parts of this assignment. Claude helped with project setup, chart-code scaffolding, and debugging R errors. All chart designs, story decisions, data interpretation, and final integration are the author’s own. All written content was reviewed and rewritten in my own words. Per the RMIT Library GenAI referencing guide.
Anthropic. (2026). Claude (Version Opus 4.7) [Large language model]. https://claude.ai/
Cloudscene. (2026). Australia data centre market overview. Retrieved June 2026, from https://cloudscene.com/market/data-centers-in-australia/all
Ember. (2025). Yearly electricity data [Data set]. https://ember-energy.org/data/
OpenElectricity. (2026). NEM regional and national electricity data [Data set]. https://openelectricity.org.au/
World Resources Institute. (2023). Aqueduct 4.0 country rankings [Data set]. https://www.wri.org/data/aqueduct-40-country-rankings