Environmental Cost of Compute

The concern runs through three physical channels. The first is electricity, which carries whatever emissions the local grid carries. The second is water, used directly for evaporative cooling and indirectly by thermal power stations. The third is hardware: the semiconductors, servers, racks and buildings, whose emissions are embodied before a single query is served. The International Energy Agency put global data centre electricity consumption at about 415 TWh in 2024, roughly 1.5% of world electricity, growing at around 12% a year since 2017, more than four times the growth rate of total electricity consumption.^[1]

The scale claim is a projection rather than a measurement. The IEA's Base Case has data centre consumption more than doubling to around 945 TWh by 2030, slightly more than Japan's total electricity consumption today, with accelerated servers, which are mainly used for AI, growing at about 30% a year against 9% for conventional servers.^[1][2] Its Lift-Off Case reaches about 1,700 TWh by 2035 and its Headwinds Case about 700 TWh, a spread wide enough that the choice of scenario largely determines the conclusion.^[2]

The distinctive part of the argument is concentration. Data centres cluster in a small number of places, which the IEA notes makes their integration into the grid harder than that of a diffuse load such as electric vehicles.^[2] China and the United States account for nearly 80% of the projected global growth to 2030, and in advanced economies data centres are expected to drive more than 20% of all electricity demand growth over that period.^[1][2] The claim is therefore less that AI will overwhelm the world's energy system and more that it will overwhelm particular grids, water basins and electricity bills.

United States figures are the best documented, because Lawrence Berkeley National Laboratory produces a recurring congressionally requested estimate. Its 2024 report, released in December 2024, put US data centre electricity use at 58 TWh in 2014 and 176 TWh in 2023, about 4.4% of total US electricity, and projected a 2028 range of 325-580 TWh, or roughly 6.7-12%.^[3] The width of that band is the honest state of knowledge, and it comes from a national laboratory rather than an advocacy group on either side.

Per-query figures are where public discussion goes wrong most often, because published numbers differ by more than an order of magnitude for reasons of methodology rather than physics. Google reported on 22 August 2025 that a median Gemini Apps text prompt used 0.24 watt hours of electricity, 0.03 grams of carbon dioxide equivalent and 0.26 millilitres of water, using a boundary that includes idle machines, CPU and RAM, and data centre overhead; the narrower accelerator-only boundary would have given 0.10 watt hours.^[7] Mistral, publishing a life cycle assessment on 22 July 2025 conducted with Carbone 4 and the French agency ADEME under the AFNOR Frugal AI methodology and ISO 14040 and 14044, reported 1.14 grams of carbon dioxide equivalent and 45 millilitres of water for a 400-token response from Le Chat.^[8] Epoch AI's Josh You estimated in February 2025 that a typical GPT-4o query with 500 output tokens used about 0.3 watt hours, roughly a tenth of the 3 watt hour figure then circulating, which derived from 2023 assumptions about output length and hardware.^[9]

The balance between training and serving has also shifted, which changes where the burden sits. James O'Donnell and Casey Crownhart reported for MIT Technology Review on 20 May 2025 that an estimated 80-90% of computing power for AI now goes to inference rather than training, and that on the Lawrence Berkeley projections more than half of the electricity going to US data centres will be used for AI by 2028.^[15] A one-off training run is a discrete, reportable event. Continuous serving is not, which is why the disclosure problem is harder than it looks.

Corporate reporting is the clearest evidence that the aggregate is moving in the wrong direction for the companies building the systems. Microsoft's 2024 environmental sustainability report recorded total emissions 29.1% above its 2020 baseline, with direct and purchased-energy emissions down 6.3% and indirect emissions up 30.9%, and attributed the increase primarily to building more data centres and to the embodied carbon of construction materials and hardware including semiconductors, servers and racks.^[5] Efficiency has improved at the same time: Google reported a fleet-wide trailing twelve-month power usage effectiveness of 1.09 for 2025 against a global average of 1.54 in the Uptime Institute's 2025 survey, and a 33-fold fall in energy per Gemini text prompt in the twelve months to May 2025.^[7][10]

The most consequential mechanism is what supplies the marginal megawatt hour. The IEA projects that renewables meet nearly half of the additional data centre demand to 2030, with natural gas and coal supplying most of the rest: gas adds over 130 TWh of annual generation for United States data centres by 2030, and coal-fired generation in China rises by nearly 90 TWh between 2024 and 2030.^[4] In its higher case, long grid connection queues mean most of the increase beyond 2030 is met by fossil generation, because a gas turbine can be sited and connected faster than a wind farm plus transmission.^[4] A company can buy renewable certificates and still have caused a gas plant to run.

The second mechanism is local. Because load concentrates, the costs of new generation, transmission and water withdrawal are borne by specific counties, utilities and river basins, while the benefit of the model accrues to users anywhere. In advanced economies data centres account for more than 20% of electricity demand growth to 2030, and in the United States nearly half.^[1] That is the pattern that produces contested planning applications, rising retail tariffs and organised local opposition, none of which shows up in a global percentage.

The third mechanism is that efficiency stops covering volume. Between 2010 and 2018 global data centre compute rose by more than 550% while energy use rose about 6%, a period Hannah Ritchie has cited as evidence that efficiency can absorb growth.^[11] The IEA's series shows that regime ending: consumption has grown around 12% a year since 2017.^[1] Google's 33-fold improvement in energy per prompt over a year is real, and total demand still rose, because cheaper inference invites more inference.^[7]

The fourth is measurement. Almost no provider publishes a per-model footprint, and the scenario range in the two authoritative sources is very wide: 700 to 1,700 TWh globally by 2035 in the IEA's cases, and 325 to 580 TWh in the United States by 2028 in the Berkeley estimate.^[2][3] Policy set against the low end of those ranges and demand delivered at the high end is a plausible failure that requires no new technology at all.

Hannah Ritchie made the clearest sceptical case on 18 November 2024, arguing that data centres consume roughly 1-2% of world electricity including cryptocurrency, that on the IEA's World Energy Outlook 2024 projections they accounted for about 3% of global electricity demand growth to 2030, less than industry, electric vehicles or air conditioning, and that the numbers sound big mainly because they are quoted without any context about how much energy and water everything else uses.^[11] Her figures predate the IEA's April 2025 assessment, which raised the data centre contribution to just under a tenth of global electricity demand growth to 2030, so the specific percentage has moved. The structural point has not: on any published scenario, data centres are a minor term in global electricity demand and a smaller one in global emissions.

The per-query numbers that drive individual guilt have been substantially wrong. Epoch AI's analysis by Josh You on 7 February 2025 found the widely repeated 3 watt hours per ChatGPT query, a figure traceable to a 2023 paper by Alex de Vries, to be roughly ten times too high, because it assumed 2,000-token outputs, older A100 hardware and higher active parameter counts than were realistic.^[9] Google's disclosed median of 0.24 watt hours per Gemini text prompt is of the same order as Epoch's independent estimate.^[7] At those magnitudes a day of heavy chatbot use is a rounding error next to household heating or a short car trip, and consumer abstinence is not where the available reductions are.

The IEA, which produced the analysis most often cited by people worried about this, is itself measured about the conclusion. It projects data centre emissions of 300 million tonnes in its Base Case and 500 million tonnes in its Lift-Off Case by 2035, below 1.5% of total energy sector emissions, and states that the increase could potentially be offset by emissions reductions that AI adoption enables elsewhere if that adoption is widespread.^[1][12] The offset is conditional and unproven, and it is fair to note that the same body warning about demand growth does not treat it as a climate emergency.

Efficiency gains are also larger and faster than the fear usually allows. Google's fleet power usage effectiveness of 1.09 in 2025 means overhead energy is a small fraction of what a typical facility spends, against the 1.54 global average recorded by the Uptime Institute in 2025, and Google reported a 33-fold reduction in energy and a 44-fold reduction in carbon per Gemini text prompt over twelve months.^[7][10] Those are single-company figures from an interested party, and they are also the most detailed public accounting anyone has published.

Disclosure is the main regulatory lever so far, and it is thin. Annex XI of the EU AI Act requires providers of general-purpose AI models placed on the EU market to include the known or estimated energy consumption of the model in their technical documentation, and permits estimation from computational resources where actual consumption is unknown.^[13] That creates a documentation duty rather than a public figure or a limit, and it applies to model providers rather than to the operators of the data centres where most of the consumption occurs.

On the supply side, the response has been building rather than restraining. The IEA expects renewables to meet nearly half of additional data centre demand to 2030 while gas and coal supply most of the remainder, and expects nuclear to matter increasingly towards the end of the decade, with small modular reactors entering the mix after 2030.^[4] Efficiency work continues on the operator side, where power usage effectiveness has fallen far enough at the best operators that cooling and distribution overhead is no longer the dominant term; Google reported 1.09 for 2025 against an industry survey average of 1.54.^[10] The remaining gains are in the hardware and the model, not the building.

What does not exist anywhere is a binding limit on the electricity or water a training run or a data centre may consume, or a requirement to publish a per-model footprint in a comparable format. Mistral's assessment was voluntary and explicitly framed as a contribution towards a standard rather than compliance with one.^[8] Until a common boundary is agreed, published figures such as Google's 0.03 grams per prompt and Mistral's 1.14 grams per response cannot be compared directly, because they cover different models, response lengths and system boundaries.^[7][8]

Keep the global question and the local question apart. Globally, every published scenario puts data centres in the low single digits of electricity demand and below 1.5% of energy sector emissions by 2035, which makes AI a poor target for anyone trying to reduce total emissions quickly.^[1] Locally, a load that concentrates in a few grids and consumes more than a fifth of demand growth in advanced economies is a serious planning problem, and that is where the documented conflicts are.^[1][2]

When a per-query figure is quoted, ask what boundary it uses and for which model. The gap between Google's 0.24 watt hours for a median Gemini text prompt, Epoch's 0.3 watt hours for a GPT-4o query and the 3 watt hours that circulated for years is mostly assumptions about output length, hardware generation and whether idle capacity is counted.^[7][9] A figure without a stated boundary carries no information.

The indicator that matters most is what generation is built to serve new load. If additional data centre demand is met by firm low-carbon supply and the connection queues shorten, the emissions part of this fear largely resolves regardless of how much compute is used. If gas capacity is added because it can be connected faster, the fear is warranted even if total consumption grows more slowly than the high scenarios predict.^[4]