- by Graham Holton
- The Guardian
- Issue #2166
AI data centres in the United States are expanding at an unprecedented rate, powering everything from ChatGPT to autonomous vehicles. Most of the world’s 8,000 data centres are located in the US, and the number continues to rise. The rapid growth of AI data centres increases demands for fossil fuels and water supplies with increasing carbon emissions and increased environmental impacts.
US data centres consume 5 per cent of total US electricity (130 TWh/year), equivalent to Australia’s annual use. AI workloads will likely double by 2030, because AI training large language models (LLMs) require up to fifty times more electricity than conventional cloud workloads. Training GPT-4 consumed 1.1 GWh of electricity, equivalent to powering 120 homes for a year. A single massive ‘hyperscale’ AI facility can consume 500 MW of electricity, equivalent to that used by a small city. The environmental consequence of generating this much power is that tens of millions of tonnes of CO2 are released annually.
AI hardware, such as GPUs, generates far more heat than standard servers, so water cooling is critical. A large AI data centre can use 19 million litres of water per day. Many US data centres are in arid states, such as Arizona and Nevada, where water scarcity is a major problem. Google’s data centre in The Dalles, Oregon, used over 1.2 billion litres of water in 2021. Microsoft consumed 1.7 billion litres of water in 2022. With increased consumption AI data centres compete with homes and farms for electricity and water, particularly during droughts and heatwaves. Waste heat from concentrated facilities can alter local microclimates.
AI-optimised facilities emit 50 per cent more CO2 per unit of computing than traditional data centres. The training of a single large AI model can emit 600 tonnes of CO2, equivalent to the lifetime emissions of more than fifty cars.
This rapid AI expansion can overwhelm the existing electricity infrastructure, with states like Virginia and Texas struggling to meet current demand. To meet power surges, utilities have fallen back on gas and coal plants, reversing previous decarbonisation gains. Dominion Energy in Virginia has delayed retiring fossil plants to keep up with data centre growth.
Then there’s space. Massive facilities such as Hyperscale AI data centres often cover more than 40 hectares, contributing to habitat fragmentation and local ecosystem loss. Suburban farmland and forested areas are increasingly converted into data campuses, particularly in Northern Virginia, which has the world’s largest data-centre hub.
Microsoft released 30 per cent more greenhouse gases in 2024 than in 2020. By 2030, data centres will consume 16 per cent of the total power consumption in the US. In comparison power consumption was only 2.5 per cent before ChatGPT was introduced in 2022.
Reducing the environmental burden of AI infrastructure depends on future policy, grid planning and transparency. The tech giants are doing a bit of that, working on renewable energy, advanced cooling, and circular economy practices. They’re also trying to improve computer efficiency to reduce their power consumption, with Google, Microsoft, Oracle, and Amazon using a low-power process in their data centres to minimise consumption by 15 per cent. Nvidia’s new AI chip models consume 25 per cent less power. As the demands of AI continue to grow, tech companies are trying to divert an energy and resource crisis which will have major environmental and social impacts.
Tech companies are capitalist companies, so they’re also resorting to secrecy in the face of community cencerns, as well as choosing communities that have been historically targeted for exploitation and have the least amount of civic power wielded by local residents when deciding where to put oversized and destructive data centres.
