‘Sustainable AI’ requires close collaboration between data centers, grid stakeholders: Schneider Electric

• The U.S. artificial intelligence industry can help solve the looming energy challenges posed by rapid data center growth in a “Sustainable AI” scenario that balances demand for computing capacity with efficiency improvements, demand flexibility, behind-the-meter power generation and robust grid investment, Schneider Electric said on Monday.
• In the ideal scenario, data centers act as grid stabilizers “by harnessing their demand response capabilities, implementing on-site renewables or distributed energy, and optimizing their operations to provide grid stability services,” wrote “Powering Sustainable AI in the United States” report co-author Rémi Paccou, director of sustainability research for Schneider Electric.
• Paccou and co-author Fons Wijnhoven, associate professor at the University of Twente in the Netherlands, contrasted Sustainable AI with three darker scenarios where AI worsens the power sector’s environmental impact, exacerbates grid infrastructure challenges and potentially precipitates an energy crisis.

Power demand forecasts for the U.S. AI industry vary widely even in the relatively near term, Schneider said.

Schneider’s own 2030 predictions range from 16.4 GW in its infrastructure-constrained “Limits to Growth” scenario to 65.3 GW in its inefficient Abundance Without Boundaries scenario. AI power demand reaches 33.8 GW by 2030 in its balanced Sustainable AI scenario.

Other forecasts show even faster growth. RAND Corporation’s “medium confidence” 2024 forecast projected 48 GW of AI power demand in 2028 and 130 GW in 2030, while its “upper confidence” 2024 forecast — which the Schneider report called “extreme” — predicted 92 GW in 2028 and 347.2 GW in 2030.

Predicting AI power demand growth is difficult due in part to the sheer number and complexity of the variables at play, some of which are beyond the control of AI technology companies, data center operators and electricity system stakeholders, Schneider said.

Fundamental differences in demand modeling approaches affect projections as well. In general, utilities tend to overestimate future demand, the report noted. More broadly, it said, some modeling approaches overemphasize “technological determinants” for AI development and power demand while underweighting socioeconomic determinants of power demand. Predictions that AI will account for as much as 9% to 13% of all U.S. electricity demand by 2030 may thus be overstated, Schneider said.

Yet AI is virtually certain to consume more power in the future, and infrastructure decisions made in the present will determine whether its growth helps or hinders the U.S. electricity system and broader economy, Schneider said.

In its Limits to Growth scenario, Schneider sees AI growth held back by infrastructure bottlenecks and power scarcity caused by poor planning, leading to unfavorable regulatory responses, economic pressures and political blowback.

Schneider’s Abundance Without Boundaries scenario envisions unchecked growth of AI across the economy, driven by the techno-optimist view that better technology will solve future resource challenges. In the near term, something closer to the opposite occurs, driving uncoordinated and inefficient infrastructure development — including centralized data center campuses that each require up to 5 GW of dedicated power supply — that hinders economy-wide electrification efforts and exacerbates environmental harm.

The Energy Crisis vision is Schneider’s darkest: an acute scarcity scenario where AI growth strains the power system to the point of collapse, stunting the AI industry and the broader economy. Driven by “insufficient grid planning, inaccurate AI demand forecasting, uncoordinated AI governance, and the reliance on computationally intensive techniques like synthetic data and multimodal learning,” the Energy Crisis scenario ultimately causes a significant contraction in AI demand.

The Sustainable AI scenario, in contrast, benefits from a “symbiotic” relationship between the AI industry and the electrical system, Schneider said.

On the data center side, operators and technology developers work to optimize onsite power usage effectiveness and computing efficiency while pursuing load flexibility and behind-the-meter power — ideally renewables — to reduce peak demand. On the grid side, strategic site selection and supply planning are key to sustaining AI and electric infrastructure growth without undue impacts on the environment or other sectors of the economy, Schneider said.