PJM Interconnection, the regional transmission organization serving parts of the U.S. Midwest and Mid-Atlantic, hired Charles River Associates (CRA) to produce an independent forecast of data center electricity demand for 2027–2047. The CRA study was presented to PJM’s load analysis subcommittee on August 26, 2026.
CRA combined top-down and bottom-up methods, linking nationwide AI demand projections to individually reported data center investments. The work aims to correct possible biases in demand submissions from local distribution companies.
Two diverging trajectories: training vs. inference
The analysis separates two energy drivers tied to AI. Model training — typically the most energy-intensive activity — is expected to slow over time as publicly available training data become exhausted and chip efficiency improves.
Inference — the day-to-day operation of AI models in consumer and enterprise applications — is projected to follow a more persistent growth path. CRA draws a historical analogy to the spread of the internet, arguing that broad adoption and continuous use will sustain inference demand rather than episodic, large training projects.
Overall, the study concludes the current rapid expansion will likely moderate, and the composition of demand will shift away from training-dominated spikes toward steady inference-driven consumption.
Physical constraints and geographic reallocation
CRA emphasizes that realized demand growth will be shaped by a complex set of physical limits, not just prices. These constraints include global supply chains (for example, semiconductor and critical-raw-material markets) as well as PJM-specific factors.
PJM-relevant limits identified in the presentation include available generation capacity, long lead times to build new plants, lengths of transmission interconnection queues, coal plant retirements, natural gas availability, and permitting and market rules. The core message is that higher willingness to pay does not instantly create more generating capacity, because manufacturing and delivery timelines impose physical bounds on expansion.
The CRA model explicitly accounts for geographic reallocation: PJM’s share of North American data center demand is not fixed. If PJM’s grid becomes a tighter bottleneck than other U.S. regions, some demand and investment may shift toward areas with more spare capacity.
Next steps and PJM’s parallel forecast
The CRA analysis is part of the load analysis subcommittee’s broader program. The subcommittee reviewed prior load forecasts in July 2026 and scheduled further steps through the end of October. PJM is concurrently preparing its own forecast and will compare its results with CRA’s independent study.
Significance
The study highlights that AI adoption does not automatically translate into unconstrained, uniform load growth across a single grid. Policymakers and grid operators must consider technological trends (chip efficiency, data availability), supply-chain limits, and local infrastructure constraints to ensure reliable and flexible power supply for growing data center demand.
(Illustrative image: Getty Images)
