Solmar Insights
Rapid growth in artificial intelligence compute and next-generation data centers is causing unprecedented electricity demand for utilities, with some AI infrastructure racks now approaching 1 megawatt of load. While traditional data centers drew 15 to 25 kilowatts per rack, current deployments for AI can reach 230 kilowatts per rack, a shift that is accelerating challenges for grid planning and infrastructure delivery.
Key figures
Traditional data center racks: 15 to 25 kW
Current AI infrastructure racks: ~230 kW
Future AI systems: up to 1 MW per rack
AI factories drive new concentration of load
Enterprise-grade data centers were once primarily built for standard applications and storage. The latest generation of “AI factories”, facilities tailored for machine learning model training and inference, are now setting entirely new benchmarks for power use. According to information shared by NVIDIA during a recent Siemens Grid Software webinar, the power density per rack for AI could surge by a factor of more than ten compared to traditional configurations.
Prospective AI deployments could see single racks consuming as much as 1 megawatt, a magnitude that puts intense localized demand on grid infrastructure. Unlike incremental data center expansion, these AI loads are more geographically concentrated and can move much faster through siting and procurement cycles than is possible for traditional energy generation and transmission buildouts. Utilities now must adapt to serve projects that require more power, sooner, and in fewer locations, significantly intensifying the load they are tasked with supporting.
Grid emerges as strategic AI factor
As the line between digital infrastructure and power networks blurs, energy availability has become a strategic parameter for AI infrastructure deployment. NVIDIA emphasized that electricity supply is integral to the stack that supports AI, meaning that the availability, reliability, and timing of grid connections are now deciding factors for hyperscale projects.
Securing adequate capacity requires coordinated planning well beyond a single utility’s operations. Developers, regulators, technology vendors, and large customers must now align project schedules, location choices, reliability standards, and potential flexibility options earlier than ever before. The experience of Northern Virginia, as highlighted by Dominion Energy in the webinar, demonstrates that integrating gigawatt-scale loads is no longer a technical challenge for utilities alone, but one that encompasses the full ecosystem of digital, real estate, and energy market participants.
Challenges in scenario planning and uncertainty
Forecasting demand for AI infrastructure is more complex than projecting conventional growth. Many possible futures compete: new facilities may be delayed, relocated, downsized, or their load profiles may change with advances in hardware and operational models. The supply side features its own uncertainties, from renewable output variability to transmission constraints and the pace of new grid upgrades.
Siemens and domain experts highlighted that deterministic demand forecasting is now insufficient for managing these risks. Utilities and developers are turning to scenario-based and probabilistic planning to prepare for a wider range of outcomes. Factoring in project probability, hour-by-hour and seasonal demand variability, flexibility from storage and distributed resources, and variable interconnection timelines helps ensure grid plans stay robust. The real aim is not infallible prediction, but identifying investment decisions that perform across various future demand and grid conditions.
Potential for flexible solutions
The rise of ultra-dense AI compute could benefit from flexibility mechanisms on both sides of the meter. With unpredictable ramp rates and evolving project specifications, non-traditional solutions, including distributed storage, behind-the-meter resources, and adaptable grid connection agreements, may become increasingly important.
To address the significant mismatch between load development speed and infrastructure delivery, new market models and regulatory approaches are likely to emerge. Utilities that embrace flexible planning and active collaboration with customers and developers may be better positioned to integrate the next generation of digital infrastructure without overbuilding or creating reliability challenges.
What this means for buyers
Data center power and interconnection capacity in major U.S. markets will see higher pressure from 230 kW to 1 MW per-rack AI infrastructure demand. These numbers mean that buyers should reassess whether available grid access and contract timelines can support AI-driven loads. This quarter, securing flexible solutions and scenario-based capacity assessments will be critical for competitive project delivery.
Reporting via the original publisher
