Powering the AI Factory: Utility Interconnection, Generation, and Transmission for Hyperscale Artificial Intelligence Compute Campuses
A White Paper on Utility Interconnection, Generation, Transmission, and Reference Architecture for Hyperscale Artificial Intelligence Compute Campuses
Executive Summary
The transition from conventional cloud and colocation data centers to artificial intelligence (AI) factories has redefined how utility power is procured, designed, and delivered. AI factories are no longer evaluated on rack count, square footage, or megawatts of nameplate load. They are evaluated on speed-to-power, voltage class, generation portfolio, transmission corridor access, and the maturity of the governance system that connects developers, hyperscalers, utilities, regional transmission organizations, regulators, and equipment suppliers. This white paper documents that change and provides a comprehensive operating framework for executives, engineers, and capital allocators responsible for delivering gigawatt-scale AI compute.
A modern AI factory is best understood as a programmable industrial load with a power profile that is dense, growing, volatile in millisecond timescales, and procurement-led on a multi-year horizon. A 1 gigawatt (GW) AI training campus draws as much electricity as a mid-sized U.S. city, and the pipeline of announced projects in 2026 already exceeds the available transmission and generation infrastructure across most regions. The North American Electric Reliability Corporation (NERC) projects summer peak demand on the bulk system to grow by 224 GW over the next decade, with roughly 90 GW of that growth attributable to data centers. Regional transmission organizations such as PJM Interconnection, the Electric Reliability Council of Texas (ERCOT), the Midcontinent Independent System Operator (MISO), the California Independent System Operator (CAISO), and the New York Independent System Operator (NYISO) are all actively rewriting their tariffs, queue procedures, and large load interconnection standards in response. The Federal Energy Regulatory Commission (FERC) has issued Order Number 2023 to reform generator interconnection and, in December 2025, ordered PJM to file new co-location and large load rules under section 206 of the Federal Power Act. Texas Senate Bill 6 introduced statutory study fees and a 75 megawatt (MW) screening threshold for new large loads. These changes are arriving in parallel with multi-year supply chain shortages for large power transformers, generator step-up units (GSUs), gas-insulated switchgear, and high-voltage circuit breakers, and they are arriving while the underlying load profile of AI training campuses is itself shifting toward 800 volt direct current (VDC) high-density rack architectures pioneered by NVIDIA and its supply ecosystem.
The result is a planning environment in which speed-to-power has become the primary site selection driver and conventional grid build alone cannot deliver the throughput hyperscalers and frontier AI laboratories require. End-to-end timelines from interconnection request through commercial energization commonly run five to seven years for greenfield campuses depending on region, voltage class, and the depth of network upgrades required. Long-lead equipment alone now consumes a significant fraction of the schedule: power transformer lead times have stretched to roughly 128 weeks and GSU lead times to roughly 144 weeks, both more than double their pre-pandemic baseline. Switchgear, breakers, and reactor lead times have risen in parallel, and the equipment shortfall is the single largest root cause of the seven-gigawatt 2026 capacity gap that has defined the current cycle.
Inside that envelope, this white paper makes four core arguments. First, the AI factory power problem is an architecture, capital, and governance problem before it is an engineering problem. Treating it as an isolated engineering decision discards critical decision rights and buries cost-shifting risks that surface during interconnection studies, in tariff filings, or in stranded asset disputes after construction begins. Second, the only reliable path to deploy gigawatts at the speed AI workloads demand is a hybrid power strategy that combines accelerated grid interconnection, behind-the-meter or co-located generation, battery energy storage, and demand-flexible operating modes. Pure grid dependence is too slow; pure off-grid generation is too capital-intensive and exposes the developer to fuel and emissions risk. Third, the customer and the utility each carry distinct, non-transferable responsibilities, and a clear allocation of decision rights between them is the single most useful instrument for compressing schedule. Fourth, the technical reference architecture itself is shifting from the legacy 480 volt alternating current (AC) topology to 800 VDC distribution paired with solid-state transformers (SSTs), grid-forming battery storage, and direct liquid cooling, and that shift requires utilities and customers to align on grounding, harmonics, dynamic load behavior, and disturbance ride-through earlier in the design process than has historically been the practice.
This white paper documents the full process for building an AI factory with the focus on power. It addresses the studies that must be performed by the utility, the customer, and the regional transmission organization, including the feasibility study, the system impact study, the facilities study, and the affected systems study. It addresses what the utility must do on its end, including transmission and substation engineering, equipment procurement, network upgrades, ratemaking, and resource adequacy planning. It addresses what the customer must do, including site control, demand profile disclosure, financial commitments, network upgrade funding, backup generation mandates, cybersecurity compliance, and demand response coordination. It documents how long the process takes under conventional assumptions and under accelerated hybrid strategies, with realistic schedule examples for each. It compares interconnection processes and tariffs across the major U.S. ISO and RTO regions and across the principal vertically integrated utilities now hosting hyperscale data center load. It documents the supporting reference architecture for the power chain itself, including 800 VDC distribution, SSTs, grid-forming BESS, modular substations, and the integration points with liquid cooling. It closes with a process map, a responsibility matrix, a risk register, and a set of practical recommendations.
The intended audience is the executive who must approve gigawatt-scale capital deployments, the infrastructure leader who must execute them, the utility planner who must engineer them, and the regulator or board member who must govern the resulting cost and reliability outcomes. The document is written to be read by all four stakeholders without requiring separate translations.
Full white paper below
