AI infrastructure expansion poses new risks to US climate targets without policy reforms

The rapid build‑out of AI infrastructure threatens to reshape the US power system, driving large rises in electricity demand and carbon emissions unless policy changes steer deployment toward cleaner supply and greater efficiency. According to an analysis by the Union of Concerned Scientists, US electricity consumption could rise 60–80% by 2050 under plausible growth scenarios, with data centres responsible for more than half of that increase by the end of this decade. If current policy trajectories persist, the UCS model projects a 19–29% rise in CO2 emissions from US power plants tied solely to data‑centre energy needs over the next ten years.

The UCS modelling aimed to avoid the extremes of headline figures by using mid‑range growth assumptions and by assuming only half of publicly announced data‑centre projects are actually built. Even so, the study warns that regulatory pressures on renewable deployment and rollback of climate rules could make outcomes materially worse. The analysis notes, for example, that recent federal reviews and stop‑work orders affecting onshore and offshore wind have created sizeable development bottlenecks that could constrain low‑carbon supply growth.

Independent studies and industry forecasts corroborate the scale of the challenge. BloombergNEF and other analysts warn that US data‑centre power demand could reach roughly 106 gigawatts by 2035, a step change from previous expectations and one that would stress transmission and generation in capacity‑constrained regions such as PJM and Texas. International NGOs reach similar conclusions on the emissions side: Greenpeace estimates AI‑driven electricity consumption could grow at about 200 terawatt‑hours per year between 2023 and 2030, potentially more than doubling data‑centre CO2 emissions relative to 2022.

The energy footprint of AI is driven by two linked dynamics: costly training of large models and high‑volume inference (daily user queries). Le Monde reported that training some state‑of‑the‑art language models can consume tens of gigawatt‑hours and cost in the tens of millions of dollars, while university research highlights that a single generative‑AI query may use many times the energy of a traditional web search. In aggregate, billions of such interactions magnify the problem, increasing both electricity demand and water use for cooling.

For industrial decarbonisation professionals, the policy levers highlighted by UCS and other analyses are familiar and actionable. Reinstating or strengthening tax credits for wind and solar would accelerate low‑carbon supply additions and, according to the UCS modelling, could cut CO2 emissions from the data‑centre surge by more than 30% over the next decade. The group further projects that such policies could lower wholesale power costs by roughly 4% by 2050, after a near‑term uptick in prices as new capacity is built.

Beyond renewables incentives, several practical measures would reduce the emissions intensity of AI growth. Energy efficiency improvements in server hardware and software, tighter colocation permitting and grid‑aware siting of facilities, and greater use of long‑term clean power contracts and on‑site generation all lower reliance on fossil‑fuel baseload. The UCS model and sector reports emphasise that the timing and location of data‑centre demand matter: where new load lands can determine whether incremental generation is gas, coal, or renewables‑dominated.

The political economy complicates straightforward solutions. The current federal administration’s actions to review or delay renewable projects and to scale back certain regulatory controls have already interrupted planned additions of clean capacity, according to the UCS analysis. Industry lobbying, local permitting disputes and transmission bottlenecks likewise contribute to a divergence between announced projects and achievable clean build‑out timelines; utilities have revised demand projections downward after vetting speculative data‑centre requests.

For businesses planning decarbonisation pathways, the implication is clear: demand‑side planning must be integrated with policy engagement and procurement strategy. Operators of industrial sites and large facilities can help by contracting for additionality in clean supply, investing in on‑site storage and flexible load management, and prioritising architectures and software that reduce compute‑intensity per task. According to a report from University of Technology Sydney researchers, efficiency gains at the algorithmic and hardware level can substantially reduce the per‑transaction energy burden of AI workloads.

Absent coordinated action, the combined effect of AI growth and constrained renewables could bring higher electricity bills as well as higher emissions, a politically salient outcome in an election‑year climate where energy costs are foregrounded. But the available evidence shows that targeted policy choices, chief among them restoring and expanding incentives for wind and solar, streamlining permitting for clean projects, and encouraging efficiency standards for data‑centre infrastructure, could substantially blunt both the emissions and price impacts of the AI boom.

For industrial decarbonisation practitioners, the coming decade will be defined as much by choices on grid planning, procurement and efficiency as by advances in compute. The technical pathways to keep AI growth aligned with climate targets exist; the challenge now is aligning regulatory frameworks, grid investment and corporate procurement to ensure the extra megawatts drive productivity rather than emissions.