AI Energy Demand Creates 'Adaptation Trap' Despite Renewable Growth

The narrative that surging AI energy demand will naturally drive clean energy investment faces a fundamental challenge: the economics of frontier AI development may actually lock in fossil fuel dependence even as renewable capacity grows.

Research from Tinglong Dai, Bernard T. Ferrari Professor of Business at Johns Hopkins University's Carey Business School, and co-author Luyi Gui reveals a troubling dynamic in how AI labs scale their models. The core issue is a race between two growth rates: how quickly a model's market value increases with capability versus how rapidly its energy costs rise. When value consistently outpaces cost—the current reality for frontier AI—developers push to maximum scale regardless of power source.

The marginal megawatt-hour problem

The International Energy Agency projects renewables will supply the majority of new data center electricity through 2035, with fossil fuels covering only 15% of new capacity. But this aggregate figure masks what happens at the margin. When a frontier lab races to train the next generation model, it draws on whatever power remains available after dedicated clean supply is exhausted—typically fossil-heavy grid power.

Adding renewable capacity in this environment doesn't displace fossil generation. Instead, it raises the ceiling on total compute available, allowing developers to scale even larger models that still run partly on coal and gas. Aggregate renewable growth and marginal carbon intensity for AI's most demanding workloads are two different metrics, and the industry consistently highlights only the first.

Why it matters

This creates what Dai and Gui call an "adaptation trap"—a feedback loop where AI's genuine climate adaptation value (improved forecasting, disaster response, grid management) justifies continued scaling on carbon-intensive infrastructure. As climate damage worsens, adaptation value rises, making carbon-intensive AI expansion appear increasingly worthwhile. The researchers' estimates suggest current U.S. and Chinese AI markets sit closer to this trap than most realize.

Policy battles over data center interconnection

These dynamics are playing out in real-time U.S. energy policy. In June 2026, the Federal Energy Regulatory Commission ordered all six regional grid operators to justify or reform how they handle data center interconnection, following an October 2025 Department of Energy directive. Critically, nothing in the federal push ties faster interconnection to clean energy procurement.

State-level responses offer a potential counterweight. Twenty-three states have approved large-load tariffs as of May 2026, with seven more pending. Ohio's version requires data centers to pay at least 85% of contracted capacity on twelve-year minimum terms; after implementation, the utility's large-load forecast reportedly fell by half. By raising the real cost of speculative scaling, such tariffs discourage oversized load requests without mandating specific fuel sources.

One proposal under discussion would reserve fast-track interconnection specifically for data centers that sign firm contracts for new, nearby renewable and storage projects capable of meeting load requirements in most hours. This would directly address the marginal power source problem rather than just adding aggregate capacity.

The path forward

Dai and Gui's research points to a different regime that could emerge when energy costs grow faster than the market will pay for additional capability. In this "resource-led" scenario, developers become genuinely cost-sensitive, and cheaper clean power directly constrains model size. But reaching that regime requires policy intervention—carbon pricing, clean-matching requirements, or grid-emissions-linked rules that make the marginal megawatt-hour powering AI clean.

The details were first reported by Forbes contributor Anjana Susarla, a professor of Responsible AI at Michigan State University's Eli Broad College of Business, based on her conversation with Dai about the research.