AI's Invisible Power Problem

The increasing electricity demands of artificial intelligence (AI) and how current tracking systems are insufficient to capture the full picture. While large AI training campuses are visible, a growing demand from smaller, distributed "inference" facilities is becoming harder to monitor and manage, potentially leading to unseen power shortages.

New Registry: PJM, a grid operator, has proposed a Large Load Registry to track big electricity users, starting at 50 megawatts. This is a good step but doesn't capture all emerging AI power needs.

Training vs. Inference: AI training requires massive, centralized power for model building. AI inference, however, involves continuous, smaller computations for every user request, leading to distributed demand.

Inference Growth: Inference already accounts for over half of AI compute and is projected to surpass training demand by 2030.

Emerging Facilities: Alongside large campuses, smaller metro, sovereign, and enterprise inference facilities are appearing. Some will have power loads below the current 50-megawatt threshold.

Decentralized Infrastructure: Companies like Akamai and telecom operators are distributing AI workloads across thousands of edge locations, creating a fragmented but significant demand.

Location Constraints: Unlike training sites, inference facilities are often tied to specific locations (e.g., cities, universities), limiting their ability to relocate for power availability.

Invisible Shortages: Smaller projects that fail to get built due to power constraints go unnoticed, unlike the cancellation of large campuses. This hidden demand impacts other essential developments like housing and hospitals.

Solutions:

Improved Measurement: Regulators need to track demand below current thresholds, including withdrawn applications and inquiries.

Streamlined On-Site Generation: Making it easier to build dedicated power generation for these smaller facilities is crucial. Current review processes are too lengthy for the scale of these needs.

The electricity demands of AI are becoming more distributed and harder to see. The current focus on large AI campuses overlooks the growing impact of smaller, localized inference facilities. Acknowledging and measuring this distributed demand, alongside simplifying the process for on-site power generation, is essential to prevent future electricity shortages and ensure continued development in cities. 

https://www.realclearenergy.org/articles/2026/09/23/ais_invisible_power_problem_1207629.html