The Carbon Reduction Math for AI Doesn't Add Up Yet

At the UN Climate Conference in Antalya, Turkey, AI was placed front and center on the agenda. The host nation called on all countries to commit to ensuring AI is designed, powered and deployed in ways that support climate action (COP31 related coverage). The catch is that there is still a huge gap between these public pledges and actual verified emission cuts.

AI companies have been loudly touting their climate benefits. Back in 2024, OpenAI CEO Sam Altman stated that AI would deliver "stunning victories," including "fixing the climate." A year earlier, Google's head of sustainability and a co-author wrote in an op-ed that AI could reduce global emissions by 5% to 10% by 2030 (Fortune op-ed).

These claims are not entirely unfounded. A peer-reviewed study published last year, led by researchers from the London School of Economics' Grantham Institute and Systemiq, estimated that AI could help the world cut 3.2 to 5.4 billion tons of CO₂ annually by 2035. After subtracting AI's own energy footprint, the upper end of that range is roughly equivalent to total annual US energy-related emissions (Nature study, US EIA data). The projected benefits are concentrated in three sectors: power, meat and dairy, and transportation.

But projections have to be measured against real-world reality. An AI-assisted review overseen by Jeremy Tamanini, the study's second author, combed through public disclosures from major market players across the three sectors (review report). The findings are far from optimistic: very few companies have announced they are following the specific emission reduction paths outlined in the research, and even fewer claim to have delivered tangible climate progress through these paths. In other words, the 2035 emission reduction numbers currently look more like a wishlist than a realistic target.

The power sector has the most concrete use cases. AI-powered supply and demand forecasting can get up to 20% more power from existing wind and solar farms onto the grid, cutting around 1.8 billion tons of emissions a year. The review covered grid operators in China, the US, and Europe, and public records show only one notable deployed project: UK National Grid ESO partnered with Open Climate Fix to cut solar forecasting errors by half, avoiding hundreds of thousands of tons of CO₂ emissions every year (Open Climate Fix). Many other grid operators have no public disclosures of meaningful progress on this type of AI integration. Some leeway is needed here, as some projects may not have been publicly disclosed.

The gap between promise and reality is even wider in the meat and dairy sector. To hit 2035 emission reduction targets, AI needs to help alternative protein products close the gap with animal products on taste and cost. NotCo set up a joint venture with Kraft Heinz, using its Giuseppe platform to analyze plant molecular data to find combinations that replicate the flavor, texture and scent of animal products (Giuseppe platform). This is one of the very few examples of deep, integrated deployment. Almost no other alternative protein companies are taking similar approaches. The sector is also facing far more fundamental challenges: mainstream consumer adoption is slow, and investment flows are drying up. Investment in cultivated meat startups has fallen from a peak of $989 million in 2021 to roughly $83 million last year (AgFunder report).

The transportation sector shows more promising signs. CATL, which controls roughly 40% of the global EV battery market, launched an AI system trained on more than 100,000 battery design cases and 600TB of test data. The company says the system cuts candidate material screening time by two orders of magnitude (CATL announcement). Independent researchers also found that DiDi uses AI to optimize routing and reduce unnecessary detours, cutting emissions per trip by an average of 16% (related study).

There is another side to this climate accounting. Oil and gas companies are also keen to use these same AI tools, including to boost fossil fuel production (Nature article). Tech companies that have previously pledged to meet sustainability goals need to answer one key question: are their tools simultaneously helping increase output from high-carbon fossil fuels? On top of that, integrating AI is only the first step. The standard, established workflows for certification, permitting, construction, packaging, marketing and sales will not automatically speed up just because AI is introduced.

The real world is already sending warning signs. The rapid adoption of AI models and expansion of data centers is driving up energy consumption and emissions. Major firms including Microsoft and Google have already fallen well short of their stated climate targets. Under the IEA's baseline projection, global data center emissions could exceed 300 million tons per year by 2035 (IEA report). This figure covers both AI and non-AI computing tasks.

That is why the carbon reduction math for AI does not add up right now. Larger, more precise AI deployments could narrow the gap between 2035 projections and real-world outcomes, but that will require sustained, targeted effort. Industry leaders should prioritize deploying AI in use cases with the highest emission reduction potential. Enterprise customers can choose models with lower environmental footprints, and partner with AI developers that are delivering verifiable carbon cuts (CLEER guidance). State and regional governments will need to create incentives or introduce requirements to push companies to operate more responsibly.

Mary Nichols is Distinguished Advisor at the UCLA School of Law and former chair of the California Air Resources Board. Jeremy Tamanini is the founder of Dual Citizen, a sustainable research and consulting firm.