AI and the climate are two massive, existential themes inevitably bound to shape human lives for decades. AI may prove to be one of the most powerful tools ever developed for addressing the climate crisis, but its current deployment is actively worsening the problem. The question I want to ask here is whether what is being built for tomorrow can fix the structural damage being done today.
Sam Altman of OpenAI has written that AI will “fix the climate”, a bold statement without much substance behind it. Dario Amodei of Anthropic believes AI will reach Nobel-level scientific research across multiple fields within two years, enabling us to accelerate progress towards solving the climate crisis. Demis Hassabis, Nobel laureate and CEO of Google DeepMind, is more measured: if humanity achieves AGI (artificial general intelligence), energy could become both zero-carbon and free, allowing us to “transcend the climate crisis.” But he calls for cautious optimism rather than certainty, and grounds his case in specific proven applications rather than sweeping promises. Three of the most powerful people in AI, three varying views. All three, however, are broadly in the ‘yes’ camp.
These promises are being made by the same individuals overseeing one of the largest expansions of fossil-fuel-powered infrastructure in recent history. A Wired investigation found that gas projects linked to just 11 data centre campuses in the US – serving OpenAI, Meta, Microsoft, and xAI – could emit more than 129 million tons of greenhouse gases per year. More than Morocco emitted in 2024. Goldman Sachs projects data centre power demand will surge by more than 220% by 2030. The IEA estimates that connecting new renewable energy sources to the grid can take up to 15 years. Much of the marginal electricity currently powering the AI buildout is coming from fossil fuels; renewable energy is not coming on line fast enough or at a great enough scale to provide a viable substitute.
Humanity already possesses the wherewithal required to solve the climate crisis – solar, wind, batteries, heat pumps, electrification – these are mature technologies, and their economics are increasingly compelling. What has been missing is the global political and public will to deploy them at the speed and scale required. An important question about AI and climate is not just whether it can accelerate technological solutions; it is whether AI can help change the political, commercial and social calculus that has stymied rapid progress for three decades. Can AI make solving climate change feel less costly, less disruptive, and more achievable than it feels today? Perhaps.
A rigorous attempt at understanding AI’s benefits for the climate comes from a peer-reviewed paper published last year in Nature’s Climate Action, led by researchers at LSE’s Grantham Institute and Systemiq. This study takes a bottom-up approach, analysing the projected growth of low-carbon technologies sector by sector. Its headline finding: AI applications in just three sectors (power, food, and mobility) could reduce global greenhouse gas emissions by 3.2 to 5.4 billion metric tons annually by 2035. That would more than offset all projected data centre emissions from AI across the entire global economy in the same period, and move us 36% closer to an ambitious emissions reduction trajectory versus business as usual. This would be a huge win, and powered by only three sectors.
To understand how this potential might be realised, we can look at where AI is actually creating impact. At a high level, the picture organises itself into three distinct types of contribution: scientific discovery, optimising systems, and redirecting capital.
The first, scientific discovery, is the hardest and highest-leverage category. AI is compressing decades of R&D into months in fields where human science has stalled, particularly material science and chemistry. We see this in Google DeepMind’s GNoME, which has identified more than 2 million theoretical crystal structures for next-generation battery storage, and in platforms like Orbital Materials that design new materials for industrial efficiency. Similarly, Cambridge-based Monumo applies AI to optimise electric motor systems – a critical intervention given that motors account for roughly half of global electricity consumption. Yet the commercialisation timelines remain long; materials discovered in a laboratory today take decades to deploy at global scale, leaving a wide gap between immediate promise and actual impact.
The second is optimising existing systems, where near-term wins are largest and most measurable. AI can process vast streams of real-time data to make grids, supply chains, and industrial processes continuously less wasteful. In manufacturing, platforms like Gigaton autonomously control energy-heavy processes to cut carbon, while Emerald AI manages data centre power loads dynamically during peak grid demand. Even consumer-facing efficiency scales rapidly; Google Maps’ fuel-efficient routing has prevented more than 2.4 million metric tons of CO2e emissions since 2021. However, optimisation works strictly within existing frameworks rather than transforming them – and the IEA is clear that efficiency gains alone cannot deliver a net-zero transition.
The third is redirecting capital, an administrative but powerful lever. With AI, investors and corporations can assess climate risk, verify carbon claims, and identify clean energy opportunities at unprecedented speed. Financial institutions use platforms like Clarity AI to assess transition risk across entire investment portfolios, while Station A helps corporate real estate owners evaluate commercial clean energy infrastructure. In environmental and nature markets, platforms like Stabiliti connect funding directly to verified restoration projects, and Treefera provides data intelligence across trading and insurance supply chains. Better information changes how capital is allocated but, because money ultimately follows returns, the question remains whether this capital will flow fast enough in a macroeconomic climate defined by high debt costs.
The pipeline of solutions is real, well-funded, and growing. The non-profit Climate Change AI catalogues dozens more credible applications across energy, transport, industry, and climate science. But the Grantham Institute / Systemiq paper is also explicit that this is potential, not destiny. The positive scenario assumes AI will be deployed toward climate solutions at scale and speed, and that requires active policy choices that are not yet in place. Can these new AI-driven hopes galvanise the world’s political leaders into action? It remains to be seen. No doubt we will need a change of guard at the White House for this to be even considered a plausible outcome.
And what happens if those policy choices are not made? And, even if they are, is the positive case really as strong as it looks? A rigorous attempt to answer these questions comes from a forthcoming MIT study, which asks a simple question: what is the net climate impact of AI, when you properly account for both its emissions and its potential to reduce emissions elsewhere in the economy? Their framework surfaces four problems that the optimist case consistently underestimates.
The first is timing. AI’s emissions are happening now. Its climate benefits, if they materialise, come later – years or decades later. The authors call the difference an ‘AI carbon debt’. And paying off that debt does not undo the damage done in the interim. A tonne of CO2 emitted today warms the planet every day until it leaves the atmosphere – a process that takes more than a century. The harms from that additional warming, including accelerated ice sheet melt, sea level rise, and more frequent extreme weather, are not reversed even if AI eventually achieves carbon neutrality. Carbon neutrality, they argue, is not the same as climate neutrality.
The second is scale. Data centres currently account for around 1.5% of global electricity demand. The Electric Power Research Institute projects that figure could reach between 9% and 17% of US electricity generation alone by 2030. But AI’s direct energy footprint is arguably the smaller concern. The MIT study argues that AI’s indirect emissions – through boosting productivity, accelerating economic growth, and increasing consumption across the entire economy – are likely to be larger than its direct emissions. The headline data centre numbers are huge, but potentially the tip of the iceberg.
The third is the rebound effect. As AI becomes more efficient and cheaper, we use dramatically more of it. Greater use means greater energy demand, which erodes or reverses efficiency gains. The IEA notes that modern AI chips use 99% less energy to perform the same calculations as chips from 2008, which is a remarkable improvement. But the total electricity consumed by AI is rising rapidly regardless, because the scale of deployment is growing faster than the efficiency gains. This is Jevons’ paradox applied to the most powerful general purpose technology in history: efficiency improvements tend to increase rather than decrease overall consumption, because lower costs drive greater use. The IEA’s own base case projects global data centre electricity consumption growing by around 15% per year between now and 2030, more than four times faster than the rest of the electricity sector combined.
The fourth is the fossil fuel problem. AI is not only being deployed to accelerate clean energy solutions. The IEA notes that oil and gas companies have been among the earliest and most aggressive adopters of AI, using it to reduce the cost of discovering and extracting hydrocarbons. The MIT study makes a critical observation here: the delays in AI’s benefits for fossil fuel extraction are shorter than the delays for AI climate solutions. In other words, AI is helping the problem faster than it is helping the solution. This connects to what systems thinker Nate Hagens calls the Carbon Pulse – the brief window in geological time during which humanity is burning through millions of years of stored solar energy. We are near the peak of that pulse, and AI is being used in part to extend it.
Taken together, these four problems present a fundamental challenge to the industry’s narrative. The MIT paper’s conclusion is stark: AI is likely to worsen climate change even if it stimulates large emissions reductions in the future. The conditions under which a net-positive outcome might be achieved – such as global carbon pricing, mandatory disclosure of full system-wide environmental costs, and aggressive state intervention to direct compute toward public goods – are not currently in place.
Altman and the optimists promise that AI will fix the climate while building infrastructure that currently makes it worse. This argument conveniently implies that today’s energy consumption is irrelevant because future AI will generate abundant clean power to clean up the damage. That’s not a sound climate strategy; it’s a hopeful promise.
The core issue is not capability; rather it is a question of whether the scale and speed of what is being built now can deliver on what the biosphere actually needs, on a timeline that matters. Right now, the scales are tilting the wrong way. The buildout adds carbon to the atmosphere today against a promise of solutions that may arrive tomorrow – or may not arrive at the pace required. What stands between those two outcomes is not technology, but speculative hope.
We need to stop treating AI as a distant saviour for the climate and start treating it as a resource that requires immediate, targeted governance. While governments are increasingly waking up to the vast capabilities of this computational shift, the strategic link between deploying AI and actively solving climate change is not yet explicit or structurally well understood. The priority needs to shift from accepting vague promises of future breakthroughs to actively establishing that link, and direct this awesome new analytical power toward our most pressing climate bottlenecks – forcing transparency onto supply chains, accelerating grid integration, and proving the economic case for rapid decarbonisation.
We do not need AI to dream up a distant, perfect future; we need public and private strategy to align, using it to manage the messy, material reality of the present before the technology’s carbon debt becomes a problem impossible to solve in the future.
Phil Verey
Phil Verey has spent twenty years scaling AI and sustainability businesses, most recently as CEO of Provenance. He is completing a diploma in AI for Business at Oxford University and is writing a series of essays on AI and the physical world.
