UC San Diego Will Test $8.48M System to Make 2 MW of AI Compute Grid-Responsive

UC San Diego will lead an $8.48 million California Energy Commission-funded demonstration at the San Diego Supercomputer Center that is designed to make 2 megawatts of AI compute responsive to grid conditions. The planned system combines Alderbuck Energy’s solid-state transformers with Emerald AI software that can adjust selected computing jobs as electricity needs change.

The installation is set to use two bidirectional solid-state transformers connected to UC San Diego’s 12-kilovolt campus grid. They are intended to deliver 800-volt direct current to the computing equipment. Alderbuck’s design converts medium-voltage alternating current directly to 800V DC in one device, replacing multiple conversion stages that conventional data centers use before electricity reaches servers. Each eliminated stage is meant to reduce energy lost as heat and reduce the equipment required for power delivery.

Some compute can move; some cannot

The other half of the test is workload control. Emerald AI’s software is designed to connect the facility’s computing system and power infrastructure, slowing, shifting or briefly pausing batchable work in response to grid conditions. Jobs that cannot tolerate a delay are meant to retain their performance commitments. That distinction is central: the project is not proposing that every AI task be interruptible.

Emerald AI has previously tested this scheduling approach without the UC San Diego power architecture. The company says a Phoenix live-cluster trial with NVIDIA, Oracle, Salt River Project and the Electric Power Research Institute reduced power draw by 25% for three hours without breaching performance commitments. It also says a London test with National Grid at Nebius’s AI Factory cut demand by up to 40% without performance loss. UC San Diego’s test adds the solid-state transformers, heterogeneous scientific and AI workloads, and a campus microgrid with distributed energy resources to that workload-flexibility model.

The deliverable extends beyond the data center

UC San Diego and its partners also plan to turn the demonstration’s workload-flexibility data into a flexible-load capacity planning tool for California investor-owned utilities and policymakers. The aim is to model how large electricity users, including data centers and EV fast-charging stations, can be integrated into the grid rather than treating their peak demand as a fixed number. San Diego Gas & Electric will serve as technical adviser and grid-integration partner, contributing operational, communications-standard, load-profile and modeling input.

Two additional project commitments

• Before deployment, researchers plan to simulate AI workloads and grid interactions at UC San Diego’s DERConnect hardware-in-the-loop testbed.
• With the Good For Others Foundation, the project will develop career pathways in grid-interactive data-center operations for residents of disadvantaged communities.

Field performance is still the decision point

UC San Diego says the deployment could become the first field validation of a medium-voltage-connected solid-state transformer at an operational California data center. That is a possibility, not an accomplished validation. The consequential question is whether the new power path can meet its efficiency and footprint targets while workload controls respond to real grid conditions without disrupting delay-sensitive research work. The live demonstration, following the planned simulations, is designed to produce the evidence utilities would need to assess that tradeoff.