Microsoft and Constellation Seal 20-Year Deal to Revive Three Mile Island for AI Workloads

Constellation Energy announced a landmark 20-year power purchase agreement (PPA) with Microsoft to restart Unit 1 of Pennsylvania's Three Mile Island facility, rebranding the site as the Crane Clean Energy Center. The deal will restore roughly 835 megawatts of clean baseload electricity to the PJM Interconnection grid, with commercial operations targeted for 2028 pending Nuclear Regulatory Commission approvals. The unit originally shuttered in 2019 due to market economics, sitting adjacent to Unit 2, which suffered a partial meltdown in 1979. For DevOps leaders, enterprise architects, and cloud infrastructure engineers, this agreement clarifies the operational reality of generative AI. Modern AI clusters require relentless, high-density power delivery that cannot be supported by intermittent solar and wind without massive, unproven battery storage systems. By securing dedicated nuclear baseload, Microsoft ensures its East Coast availability zones can expand AI accelerator deployments without triggering regulatory caps or local utility moratoriums tied to grid congestion. This development marks a new era in cloud data center architecture. Hyperscale providers have spent the past decade executing virtual PPAs for wind and solar credits, but the exponential power draw of dense GPU clusters (now frequently demanding dozens of kilowatts per rack) has outpaced local utility capacity. Following Talen Energy's Susquehanna arrangement with AWS, Microsoft's move confirms that dedicated nuclear power—both through recommissioning traditional gigawatt-scale plants and investing in future small modular reactors (SMRs)—is rapidly becoming the standard playbook for Tier-1 cloud operators. In practice, engineering teams should expect cloud providers to bifurcate workload placement strategies based on regional energy resilience. Ultra-low-latency and training-heavy AI workloads will increasingly concentrate in regions backed by dedicated baseload agreements like PJM, while secondary regions may face stricter quota allocations, throttled GPU instances, and higher spot instance volatility during peak grid stress. Platform engineering teams must factor carbon accounting and regional grid stability into multi-region disaster recovery and model inference architectures.