UC San Diego tests solid-state transformer for AI power

UC San Diego will demonstrate a bidirectional solid-state transformer at an operational supercomputer centre, testing whether medium-voltage power electronics and flexible computing controls can simplify data-centre power delivery while improving interaction with the electricity grid.

The $8.48 million California Energy Commission-funded project will be hosted by the San Diego Supercomputer Center and is designed around approximately 2MW of AI computing load. The installation will combine power-conversion equipment from Alderbuck Energy with workload-control software from Emerald AI and the university’s existing campus microgrid.

Alderbuck’s solid-state transformer will convert medium-voltage alternating current directly to 800V DC through a single device. The project team expects that removing conventional conversion stages could cut the footprint of associated power equipment by more than 50% and produce energy savings of approximately 25%, although both figures are project targets that still require field validation.

The technology will undergo simulation and hardware-in-the-loop testing at UC San Diego’s DERConnect facility before it is introduced into the live computing environment. That staged approach is intended to examine electrical and workload behaviour under controlled conditions before the equipment interacts with an operational research data centre.

Power conversion moves closer to computing

Large data centres normally convert electricity several times between the incoming network and the processors consuming it. Medium-voltage AC is transformed and distributed before rectifiers and other power electronics ultimately produce the low-voltage DC required inside computing equipment.

Each stage introduces electrical losses, heat, physical equipment, and protection requirements. Those effects become more material as rack density rises because a small percentage loss applied across megawatts of continuous load also increases the cooling duty needed to remove the resulting heat.

A direct medium-voltage-to-800V DC architecture removes part of that conversion chain. The potential benefit is a more compact power-delivery system with fewer intermediate stages, but concentrating functions in solid-state equipment changes the engineering risks rather than eliminating them.

The transformer relies on power semiconductors, control electronics, and thermal management rather than the comparatively passive magnetic architecture of a conventional transformer. Reliability, protection, fault response, efficiency across different loading conditions, electromagnetic compatibility, and maintainability therefore become central to the trial.

A data-centre installation makes those questions particularly demanding because interruption tolerance is extremely low. Any reduction in equipment count has to be assessed alongside redundancy, bypass arrangements, component lifetime, and the ability to isolate or maintain equipment without taking critical computing capacity offline.

Computing demand becomes part of grid control

The project extends beyond conversion efficiency by connecting the electrical system to computing workload management. Emerald AI’s control software can slow, shift, or briefly pause workloads that tolerate delay while protecting processes that have tighter performance requirements.

That distinction allows some of the site’s electricity consumption to vary in response to grid conditions rather than presenting the network with a permanently fixed peak load. San Diego Gas & Electric will act as technical adviser and grid-integration partner, providing utility input on communications, load characteristics, and network operation.

UC San Diego also plans to use project data in a flexible load-capacity tool for utilities and policymakers assessing large new connections, including data centres and high-power EV charging. The aim is to replace part of the assumption that every megawatt of requested peak demand will be required continuously under every system condition.

There is already evidence that some AI workloads can be controlled in this way. A previous London demonstration involving National Grid and Emerald AI reduced the demand of an AI computing cluster by up to 40% during grid-response tests while keeping critical workloads operating.

The San Diego project adds physical power infrastructure to that software-led approach. Rather than managing only how much computing equipment consumes, the trial will test the conversion system supplying it and the interaction between load flexibility, distributed energy resources, and the campus microgrid.

That combination could be useful for connection planning if its behaviour proves repeatable. Utilities need confidence that flexible demand will actually respond at the required time and magnitude before they can rely on it when deciding how much additional network reinforcement a new data centre requires.

Flexibility does not create unlimited hosting capacity. Transformers, conductors, substations, fault levels, voltage limits, and generation adequacy still set physical boundaries, while computing operators will retain workloads that cannot be delayed simply because the power system is under stress.

The attraction lies in avoiding an unnecessarily rigid assumption that every large digital facility behaves as an unvarying block of maximum demand. If part of the load can be controlled and the electrical conversion system can support that control without compromising reliability, planners gain another variable when assessing new connections.

The field test should provide considerably stronger evidence than another laboratory demonstration, but the project team still has to prove the targeted efficiency, equipment-footprint, and flexibility performance under real operating conditions. For solid-state transformers, that transition from promising power electronics to dependable critical infrastructure is the part of the experiment that matters most.