Problem, or solution? The energy and water costs of AI are measurable and rising. So are its contributions to climate modelling, wildfire response, and grid optimisation. Neither side has closed the balance sheet.
Craig Stanley Studio · The Debate · Issue 02 of 09 · craigstanley.work
The energy and water costs of training and running AI systems are real, growing, and measurable. So are AI's contributions to grid efficiency, climate science, and environmental monitoring. The honest answer is that nobody has yet produced a rigorous net calculation — and that gap is itself a problem.
The IEA projected global data centre electricity demand approaching 1,050 TWh by 2026 — roughly equal to Japan's entire national electricity consumption. At the same time, DeepMind's AI-powered cooling optimisation reduced Google's data centre energy use by 40%, and AI climate models are compressing what once took weeks of supercomputer time into hours. Both of those things are true simultaneously. The question is whether the scale of the cost and the scale of the benefit are even in the same order of magnitude — and on that question, there is no settled answer.
This is not primarily a technical debate. It is a transparency and methodology debate. Corporate sustainability pledges are commitments about the energy mix powering data centres — not calculations of whether the AI running inside those centres does more climate good than harm in aggregate. Until those calculations exist, the debate will generate more heat than light.
The sourced claims on both sides. Neither column wins by default — the honest task is to hold them simultaneously and ask what a shared methodology for measuring net impact would actually look like.
IEA: global data centre electricity demand approaching 1,050 TWh by 2026 — roughly equal to Japan's total electricity consumption. The growth is driven primarily by AI workloads, which are more compute-intensive than conventional cloud applications.
Microsoft's 2025 sustainability report recorded total emissions 23.4% higher than in 2020, attributing the rise directly to AI and cloud expansion — while simultaneously maintaining a pledge to become carbon negative by 2030. The divergence between trajectory and target has not been explained.
Nature Sustainability (2025): US AI server deployment could generate 24–44 Mt CO₂-equivalent per year between 2024 and 2030, with a water footprint of 731–1,125 million m³ annually. Water consumption is the cost least often cited in corporate disclosures.
The rebound effect: efficiency gains in AI models are continuously offset by increased use. Cheaper inference enables more queries, more applications, and larger model deployments — not fewer data centre resources. This is a modern Jevons paradox, documented in peer-reviewed literature on AI energy consumption.
DeepMind reduced Google data centre cooling energy by 40% using its own AI optimisation tools. Separately, AI is being used to optimise renewable dispatch on power grids — enabling more variable solar and wind capacity to be absorbed without curtailment or storage losses.
AI-powered wildfire prediction models are now operating in California, Australia, and parts of Southern Europe. Earlier warnings reduce both human cost and the eventual carbon load of the fires themselves, which are among the largest single emissions events in those regions.
Satellite-based AI systems monitor deforestation in near-real time. Global Forest Watch and comparable services can flag illegal clearing within days, enabling enforcement responses that would previously have taken months. Preserved forest is preserved carbon sequestration.
AI accelerates the processing of atmospheric and ocean data, compressing the time to generate high-resolution climate projections from weeks to hours. This enables faster iteration in climate science and more timely inputs to policy — a multiplier on existing scientific infrastructure.
The honest answer is that no one has closed the balance sheet yet. The energy and water costs are real, measurable, and rising. The climate benefits are real, measurable, and significant — but not yet sufficient to offset the costs in aggregate at the scale they are currently operating.
The missing thing is a shared methodology for measuring the net effect. Corporate pledges — carbon-free by 2030, carbon negative by 2030 — are not that. They are targets for the data centre's own energy mix: a commitment to source electricity from renewable generation. They are not a calculation of whether the AI running inside that data centre does more climate good than harm when its full application portfolio is taken into account. Those are different questions, and the second one has no accepted answer yet.
The scientific literature is beginning to assemble the pieces. The Nature Sustainability study provides an estimate of the cost side. The npj Climate Action review provides an estimate of the benefit side across grid optimisation, wildfire, and environmental monitoring. But these use different scopes, different baselines, and different attribution methodologies. Adding them together to produce a net figure is not yet a settled practice. That gap — the missing methodology, not the missing data — is the real frontier of this debate.
Worth naming the gap: Ireland's data centres already account for 21% of national electricity consumption (Carbon Brief, 2025). That cost falls on the local grid, the local water supply, and the local planning authority — not on the company making the sustainability announcement from a headquarters in another jurisdiction. The localised burden of AI infrastructure is systematically absent from aggregate corporate disclosures.
There is currently no jurisdiction with binding legislation requiring AI companies to disclose the net climate impact of their systems — as opposed to the energy mix of their data centres. The EU AI Act's sustainability provisions are limited to energy efficiency requirements for high-impact models. Mandatory lifecycle carbon disclosure for AI products does not yet exist anywhere. The regulatory gap between what companies are required to report and what would constitute a genuine net-impact calculation is wide.
Data centre construction is accelerating faster than the grid decarbonisation that was supposed to make it sustainable. Microsoft, Google, Amazon, and Meta have all announced data centre expansions in 2025–2026 that will significantly increase electricity demand over the next three to five years. The renewable energy procurement commitments attached to those expansions are real — but they are procurements, not guarantees that the marginal electricity consumed will come from zero-carbon generation at the moment of consumption. Grid accounting and physical reality are not the same thing.
The most useful contribution the AI industry could make to this debate is not a more ambitious pledge — it is a standardised methodology for measuring the net climate effect of an AI deployment, accounting for both the energy cost and the application benefit. That methodology does not currently exist. Until it does, every claim that AI is "net positive" for the climate — or "net negative" — is a position, not a finding. The debate cannot be resolved without the tools to resolve it.
Primary sources cited in this issue. Dates and URLs verified at time of publication.
The energy and water costs of AI infrastructure are measurable and rising. The climate benefits — in grid optimisation, wildfire response, deforestation monitoring, and climate modelling — are also measurable and significant. This issue lays both sides out and names what we don't yet know.
Every issue of The Debate presents the strongest sourced case on both sides of a live question in AI. The ledger is not a verdict. The bridge names what both sides leave unresolved. The line says where law and policy currently sit.
Craig Stanley Studio · craigstanley.work · The Debate series · Issue 02 of 09 · 2026