by Alexander Moshchev

Recap: In the first part, we explored what powers a single AI output – from the computational intensity of large language models to the vast data centre infrastructure that keeps them running. We looked beyond the processors themselves to the physical and energy systems that sustain them, examined how demand scales exponentially as the models grow and situated these trends within the IEA’s energy scenarios for a digital and decarbonising world. We are now going to forward on energy forecasts – and the challenges around making the right assumptions.

Understanding forecast variations

Energy projections for AI vary considerably because each analysis makes different assumptions about three fundamental variables:

Energy efficiency per query

As detailed in Section 2, multiple technological advances are reducing energy consumption per AI query: sparse model architectures like Mixture-of-Experts, improved hardware generations, and optimised serving strategies. These efficiency gains create a critical tension in energy forecasting.

The efficiency-adoption tension

While per-query energy consumption continues falling, projections still show rising total electricity demand. This reflects forecasting’s core challenge: efficiency improvements compete against adoption growth but follow different timelines.

Hardware generations advance every 18-24 months, yet widespread deployment takes years. Meanwhile, AI feature rollouts happen within months across billions of users, explaining why forecasters see near-term demand growth despite long-term efficiency trends.

Real-world efficiency varies dramatically by use case, serving conditions, and model sizes, forcing forecasting models to create uncertainty ranges rather than precise predictions.

Adoption rates and usage patterns

Even small per-query energy consumption becomes substantial when multiplied by billions of daily interactions. This scaling factor explains why the IEA’s forecast range widens most dramatically based on adoption assumptions – whether AI becomes embedded in search, productivity software, customer support, and development tools.

Several signals indicate accelerating adoption: major technology companies rolling out AI features across their product portfolios, substantial contract backlogs at cloud and AI infrastructure providers, government initiatives for sovereign AI capabilities, and usage metrics showing rapid uptake in enterprise productivity suites.

Geographic and temporal factors

While energy consumption is determined by physics, environmental impact depends entirely on location and timing. Regional grid mix determines carbon intensity per kilowatt-hour, while cooling methods affect water consumption independently of energy use. The IEA’s modelling incorporates these geographic variations, finding that some regions experience significant strain well before global totals appear concerning.

Leading providers pursue decarbonisation strategies, including 24/7 carbon-free energy targets, renewable power purchase agreements, and strategic siting near clean generation sources. These approaches reduce emissions intensity without necessarily reducing energy consumption.

How the variables interact in practice

The IEA’s scenarios demonstrate how different assumptions about these three variables create dramatically different outcomes. The Base Case incorporates steady efficiency improvements alongside moderate adoption growth, projecting approximately 945 TWh by 2030 – just under 3% of global electricity consumption.

More aggressive assumptions yield strikingly different results. The Lift-Off scenario, which assumes accelerated adoption and streamlined infrastructure development, projects consumption exceeding 1,700 TWh by 2035. Conversely, the High-Efficiency pathway, emphasising rapid technological improvements and operational optimisation, limits consumption to approximately 970 TWh by the same timeframe.

Evaluating competing forecasts

When encountering divergent energy projections, the key question is not which analyst is correct, but rather which assumptions they made about efficiency trends, adoption rates, and infrastructure development. Did they assume widespread deployment of advanced model architectures and next-generation hardware? How quickly do they expect AI integration across major applications? What infrastructure constraints and energy procurement strategies do they model?

Understanding these underlying assumptions provides far more insight than comparing headline numbers alone. The wide range between credible scenarios reflects genuine uncertainty about technological and adoption trajectories rather than analytical disagreement.

Industry responses: Infrastructure and procurement strategies

Major technology companies are implementing strategies that primarily address emissions and grid integration rather than reducing computational energy consumption per query. These approaches focus on cleaner energy procurement, more efficient facilities, and operational flexibility to reduce environmental impact and grid strain.

Microsoft: Securing firm clean energy capacity

Microsoft signed a 20-year agreement with Constellation to restart a Three Mile Island nuclear unit, adding approximately 835 MW of carbon-free generation to the PJM grid by 2028.[1] The company also contracted with Helion for 50 MW from their first commercial fusion plant.[2]

These initiatives exemplify “venue” strategies – approaches that reduce the carbon intensity of electricity consumption without necessarily reducing energy consumption per AI query. They help ensure adequate clean generation capacity exists where data centres operate while lowering emissions per kilowatt-hour.

Google: Efficiency leadership and grid coordination

Google reports industry-leading operational efficiency with a fleet-average PUE of approximately 1.09 in 2024. The company has invested heavily in renewable energy procurement, signing multiple large-scale power purchase agreements for wind and solar projects.[3]

The company has committed to operational flexibility, agreeing to pause non-essential AI workloads during peak demand periods to support grid stability through agreements with utility providers. Google continues expanding its data centre footprint globally while implementing advanced cooling technologies and pursuing energy-efficient server designs.

AI’s potential for system-wide emissions reduction

The question of whether AI can offset its own energy consumption through emissions reductions elsewhere has attracted significant research attention. Early evidence suggests substantial potential, though realising these benefits depends on rapid deployment and careful implementation.

Research findings on sectoral impacts

A 2025 peer-reviewed analysis by Stern and colleagues, published in npj Climate Action, quantifies potential emissions reductions from AI applications across three major sectors. Their modelling suggests that AI deployment in power systems, food production, and transportation could reduce global emissions by 3.2 to 5.4 gigatons of CO₂ equivalent annually by 2035. This potential reduction substantially exceeds their estimate of AI’s own emissions footprint, projected at 0.4 to 1.6 gigatons of CO₂ equivalent by the same timeframe.[4]

Mechanisms for emissions reduction

Power systems optimisation accounts for the largest potential impact, with an estimated 1.8 gigatons of CO₂ equivalent reduction annually. AI enhances renewable energy forecasting and dispatch optimisation, potentially increasing load factors for wind and solar installations by up to 20 per cent. These improvements enable more effective integration of variable renewable sources, reducing reliance on fossil fuel backup generation and optimising energy storage and distributed energy resource utilisation.

Food production and agriculture present opportunities for 0.9 to 3.0 gigatons of annual emissions reductions. AI applications could accelerate adoption of lower-carbon protein alternatives through improved cost-effectiveness and quality, while also enhancing market matching between supply and demand. Additionally, AI-driven optimisation of existing agricultural supply chains could reduce process emissions and minimise food waste throughout distribution networks.

Transportation systems offer potential reductions of 0.5 to 0.6 gigatons annually. AI applications could influence modal choices toward shared mobility options, optimise routing and traffic flow to reduce fuel consumption, and accelerate electric vehicle adoption by improving affordability assessments and charging infrastructure integration.

Critical limitations and uncertainties

Several important factors temper these optimistic projections. The analysis covers only three sectors, meaning real-world net benefits depend heavily on broader adoption patterns and supportive policy frameworks.

The modelling excludes rebound effects – where efficiency improvements lead to increased consumption – which could partially offset projected gains.

Additionally, the research assumes relatively static grid carbon intensities between 2022 and 2035 for simplicity. In reality, grid decarbonisation would reduce AI’s own emissions footprint, while continued reliance on fossil fuels would increase it. The ultimate climate impact will depend on the parallel evolution of both AI deployment and electricity generation sources.

Conclusion

The energy consumption of artificial intelligence ultimately comes down to three interconnected variables: the efficiency of individual queries, the scale of adoption, and the characteristics of where and when the computation occurs. While a single AI response consumes minimal energy – comparable to powering an LED bulb for a few minutes – the aggregate impact depends on multiplication across billions of daily interactions.

Current trends suggest a complex trajectory. Technological advances continue to reduce energy consumption per query through more efficient model architectures, improved hardware, and optimised infrastructure. Mixture-of-Experts models, advanced accelerators, and better data centre design all contribute to this efficiency improvement. However, these gains occur alongside rapid expansion in AI deployment across search, productivity tools, and enterprise applications.

Credible projections from the International Energy Agency indicate that total data centre electricity consumption will continue rising through 2035, with scenarios ranging from conservative growth to substantial increases depending on adoption rates and infrastructure development. These forecasts represent possibilities rather than certainties, as each of the three fundamental variables – efficiency, volume, and location – remains subject to technological and policy influences.

The geographic and temporal aspects of AI deployment increasingly determine environmental impact beyond absolute energy consumption. Strategic choices about clean energy procurement, facility location, and operational flexibility can significantly reduce emissions intensity and grid strain without necessarily changing computational energy requirements. Meanwhile, emerging research suggests that AI applications in power systems, transportation, and agriculture could potentially offset AI’s own emissions through system-wide efficiency improvements – though realising these benefits requires widespread deployment and supportive policy frameworks.

The path forward involves managing all three dimensions simultaneously: continuing efficiency improvements, deploying AI thoughtfully as adoption scales, and ensuring that the infrastructure supporting these systems aligns with broader climate objectives.

Footnotes

[1] Constellation Energy. “Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to The Grid.” Press Release, September 20, 2024. 20-year PPA with Microsoft to restart Three Mile Island Unit 1, ~835 MW. https://www.constellationenergy.com/newsroom/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html

[2] https://www.helionenergy.com/articles/helion-announces-worlds-first-fusion-ppa-with-microsoft/

[3] Same as 5

[4] Stern, Nicholas, Mattia Romani, Roberta Pierfederici, et al. “Green and Intelligent: The Role of AI in the Climate Transition.” npj Climate Action 4, no. 1 (2025). AI-enabled abatement ~3.2–5.4 GtCO₂e/year by 2035; AI emissions ~0.4–1.6 GtCO₂e. https://www.nature.com/articles/s44168-025-00252-3