The hidden climate cost of AI: how data centres are warming our cities

Every time you ask ChatGPT to draft an email, stream a show on Netflix, or ping a colleague on Teams, a bank of servers somewhere is spinning up, drawing power, and dumping heat into the air outside. That heat doesn't just disappear. According to new research out of Arizona State University, the industrial-scale exhaust from data centres is measurably warming the neighbourhoods around them — by as much as 4 degrees Fahrenheit (roughly 2.2°C) in parts of Phoenix.

For Australians watching hyperscale server farms sprout across Sydney's west, Melbourne's north and the outskirts of Canberra, the finding lands at an awkward moment. We are being told that AI will help solve climate change. But the physical infrastructure that runs it is quietly making our hottest suburbs hotter.

What the Phoenix study actually found

The research, reported by Tech Xplore and picked up by Gizmodo, modelled the way large data centres interact with Phoenix's already sweltering microclimate. The headline number — up to 4°F of extra warming in adjacent neighbourhoods — is a big deal in a city that regularly clears 45°C in summer and where heat is the leading weather-related killer.

The mechanism isn't complicated. Servers convert electricity into computation and, as a by-product, waste heat. That heat has to be removed to stop the hardware failing, so it is blown outside via massive cooling systems — chillers, evaporative towers, and increasingly, direct-to-chip liquid loops. What comes out the back of a data centre is essentially a permanent hot exhaust plume the size of a warehouse.

As Facilities Dive notes, the effect is highly localised — worst in the streets immediately downwind — but it stacks on top of the existing urban heat island effect from roads, concrete and rooftops. In cities where data centre campuses are clustered, the compounding is significant.

Why AI is making the problem worse, fast

Data centres have existed for decades. What has changed is density. Training and running large language models like GPT-4, Gemini and Claude requires racks of GPUs that consume five to ten times more power per square metre than the servers that ran your bank's website a decade ago. A single AI-optimised rack can draw more than 100 kilowatts. Older enterprise racks drew five.

More power in means more heat out. Tech Briefs frames the challenge bluntly: the pace of data centre growth is now outstripping the planning tools cities use to manage urban temperature. Zoning rules written for warehouses and light industry never anticipated buildings that behave, thermally, like small power stations.

The AI boom is only accelerating the trend. Global data centre electricity demand is projected to double this decade, and hyperscalers — Amazon Web Services, Microsoft, Google, Meta — are racing to secure sites near cheap power and fibre. That often means the outer suburbs of major cities, close enough to users to keep latency low, but on land where nearby residents didn't sign up to live next to a 24/7 industrial cooling operation.

What this means for Australia

Australia is one of the fastest-growing data centre markets in the Asia-Pacific. Sydney alone hosts more than 80 facilities, concentrated in suburbs like Macquarie Park, Erskine Park and Eastern Creek. Melbourne has its own hub across Port Melbourne and Tullamarine. Canberra, thanks to federal government contracts, punches well above its weight.

These are also suburbs that already run hot. Western Sydney regularly records temperatures 6–10°C higher than coastal parts of the city on summer afternoons. The Bureau of Meteorology has flagged Penrith and Richmond as among the hottest places in metropolitan Australia during heatwaves. If the Phoenix findings translate — and thermodynamically, there's no reason they wouldn't — every new hyperscale campus in Sydney's west is potentially adding a couple of degrees to already dangerous local conditions.

That matters because heat is not an inconvenience in Australia; it is a public health emergency. More Australians die from extreme heat than from bushfires, floods and cyclones combined. Elderly residents, people on low incomes without adequate air-conditioning, and outdoor workers bear the brunt. A data centre that nudges the local ambient temperature up by even one degree during a heatwave shifts hospitalisation and mortality risk in a measurable way.

The cooling paradox

There's a cruel irony baked into the physics here. As neighbourhoods warm, residents run their air conditioners harder. Those air conditioners also dump heat outdoors. So do the data centres' own cooling systems, which have to work harder as the outside air gets hotter. The result is a feedback loop: hotter ambient temperatures mean less efficient cooling, which means more waste heat, which means hotter ambient temperatures.

Water is the other pressure point. Many older data centres use evaporative cooling, which is water-hungry — a single hyperscale facility can drink millions of litres a day. In dry parts of Australia, that competes directly with agriculture, environmental flows and household use. The industry is shifting toward closed-loop liquid cooling and air-cooled designs, but retrofits are slow and expensive.

What can actually be done

The good news is that the Phoenix researchers, and the industry engineers cited in Tech Briefs, aren't just pointing at a problem — they're pointing at levers. Several matter for Australia:

None of this is technically exotic. What's missing is the regulatory push. Australian state governments approve data centre developments largely on the basis of power supply, land use and traffic — not thermal impact on neighbours. That's a gap the Phoenix study should help close.

The bigger question

Behind the engineering is a question we haven't really debated as a country: what is all this compute for, and who bears the cost? A resident of St Marys or Kemps Creek did not vote for their suburb to become the physical substrate of the global AI industry. Yet they may end up living with hotter summers, higher energy prices, and stressed local water supplies so that someone on the other side of the world can generate a marketing image in three seconds instead of five.

The 4-degree figure from Phoenix is not a doomsday number. It's a warning shot. It tells us that the digital economy has a physical footprint, that footprint has a temperature, and Australia — hot, dry, and increasingly server-farmed — has more skin in this game than almost anywhere else.

Related on Bleen

Sources