Islanded Power: Running an AI Campus Without a Utility Connection

October 28, 2026

A grid-connected campus waits on the utility. A bridged campus runs on on-site generation until the utility connection is built. The third option is islanding, a campus designed to run on its own generation with no operating grid connection, and none planned, so nothing outside the site sets its voltage or frequency.¹

Delays push some developers there. As the rules tighten, getting a large campus connected can take years. Dominion is already seeing this in Northern Virginia, where they've cited wait times of up to seven years for large connections.² NERC’s computational load standards, which passed final ballot in October 2026, would cover any site with 50 MW or more of connected load supplied at 100 kV or above.³ US grid operators are moving to curtail large loads in a grid emergency. Texas law now let’s its grid operator disconnect large loads in an emergency, ⁴ and PJM has proposed curtailing large loads that have not brought their own generation ahead of other customers. ⁵ A connected campus in that position cuts compute or moves to backup generators. An islanded campus draws nothing from the grid, so a curtailment order has nothing to act on.

Why Independence is Hard to Engineer
NERC’s guideline, writing about restoration islands, warns that uncontrolled load pickup creates large voltage and frequency swings that can destabilize and potentially collapse an island, because island stability is sensitive to changes in both.⁶ A campus island is just as sensitive and has no grid to restore from. An island also has to produce its own fault current, and the source mix decides how much it gets. Synchronous machines deliver several times rated current into a fault, while converter-based sources typically deliver 1.1 to 2.0 times ⁷, which makes protection a matter of coordination study.

Large, fast-varying loads can swing demand by tens or hundreds of MW within seconds or minutes, and NERC notes that cyclical demand can excite torsional modes in nearby generators, leading to mechanical stress and potential damage.⁶ AI training produces those swings because tens of thousands of GPUs alternate between compute and communication phases in step, at sub-synchronous frequencies where turbine-generator shafts can resonate.⁸

Engineering Design Without the Grid
A campus designed to stay islanded has no utility to fall back on at any stage. That changes what the architecture has to do.

The voltage and frequency reference has to come from a source that does not depend on the generators. The MV-UPS, the stability block between generation and the compute halls, operates grid-forming on the load side. The compute halls get a stable reference, and the generators carry a regulated draw.⁹


Islanded configuration. Gas generation feeds the data center through the MV-UPS, with an Energy Management System coordinating the site.⁹

The generators also need protection from the load’s pulsations for the life of the site, because no utility transfer ends the exposure. Pulsations that align with the shaft’s resonant frequency cause sub-synchronous torsional interaction (SSTI) and forced power oscillations, and the MV-UPS is designed to keep them off the turbine shaft.⁹ Integrated battery storage absorbs sub-second load steps, which smooths the profile the turbines carry.⁹ 

The site has to restart itself. In islanded mode the MV-UPS provides gas turbine soft starting and black start across 0 to 100% speed.

The design also has to show what happens when a generator trips. An island has no grid reserve to cover the gap, so the design record has to state the redundancy basis, whether at the converter or the generation, and the storage duration. The controller acceptance test has to show the campus riding through the loss of its largest unit.

Above the converter, the campus needs a microgrid controller, or Energy Management System (EMS), to manage dispatch and energy balance. It is a separate system from the MV-UPS. IEEE 2030.7 specifies its functions above the component control level, and IEEE 2030.8 defines the tests that verify them.¹⁰

What Staying Islanded Asks of the Operator
A permanent island needs daily operating coverage and a maintenance and spares program for the life of the site. The operator takes on the role of a power producer, with the risks of self-supply that come with it.¹ The MV-UPS addresses one of them: the electrical physics of running without a grid.⁹

Uncertain whether a campus’s load profile and supply plan call for an MV-UPS? Answer this four-question check.

References

  1. Regulatory Assistance Project, Going Off Grid: The False Panacea of Islanding High-Impact Loads, May 13, 2026.
  2. Bloomberg, Virginia Data Centers Face Seven-Year Wait for Power Hookups, Dominion Says, August 2024.
  3. Power Magazine, NERC Reliability Standards for Data Centers and Crypto Loads Pass Final Ballot, Advance to Board Review, October 15, 2026.
  4. Utility Dive, Texas Law Gives Grid Operator Power to Disconnect Data Centers During Crisis, June 25, 2025.
  5. PJM, Interim Resource Adequacy Service Proposal, fact sheet, August 2026 (proposal filed with FERC August 13, 2026).
  6. NERC Large Loads Working Group, Reliability Guideline: Risk Mitigation for Emerging Large Loads, May 2026 (Final), Chapters 4, 5 and 8.
  7. IEEE Power & Energy Society, Power System Relaying and Control Committee Working Group C32, Protection Challenges and Practices for Interconnecting Inverter Based Resources to Utility Transmission Systems, PES-TR81, July 2020.
  8. E. Choukse et al., Power Stabilization for AI Training Datacenters, arXiv:2508.14318, August 2025.
  9. GEA35589, MV-UPS: Medium Voltage Uninterruptible Power Supply, GE Vernova, Rev. 1, August 10, 2026.
  10. IEEE Std 2030.7-2017, IEEE Standard for the Specification of Microgrid Controllers, and IEEE Std 2030.8-2018, IEEE Standard for the Testing of Microgrid Controllers.

About the Author

Alpesh Saraiya

Alpesh is a global executive data center solution leader at GE Vernova. He leads data center strategy, business development, M&A, product, and marketing for power conversion and storage solutions. Previously at Honeywell and other technology companies, he spent 10 years growing the data center business of power, controls, cooling, security, AI SaaS, facility management, and IoT solutions. He started his career working at semiconductor and systems companies, including IBM, Broadcom, STMicroelectronics, and LG Electronics. He has a M.S. Computer Engineering from Syracuse University and is based in the San Francisco Bay Area.

Alpesh Saraiya

Kevin Delsol

Kevin Delsol – Senior Business Development Manager Power Stability | With nearly two decades of experience spanning business development, sales, engineering, project execution, and customer engagement, Kevin currently serves as the AI Data Center Power Stability Leader at GE Vernova. Specializing in electrical infrastructure and control systems, his focus is on developing innovative solutions that address the emerging power challenges of hyperscale AI deployments. His work includes advancing medium-voltage power conditioning architectures that enhance grid stability, power resilience, and the integration of large-scale AI data centers into future energy ecosystems.

Kevin Delsol