The San Diego Supercomputer Center will test a new power architecture for artificial-intelligence data centers through an $8.48 million project funded by the California Energy Commission, UC San Diego announced Monday.
Plans call for two solid-state transformers and a controller to connect with UC San Diego's 12-kilovolt grid and deliver 800-volt direct current to two megawatts of AI computing load. The project team aims to cut the footprint of power equipment by more than half and projects energy savings of about 25% by reducing conversion steps between the grid and computer hardware.
For UC San Diego and AI Data Centers, the strongest claims are the ones tied to transparent methods, appropriate comparison groups, disclosed limitations and results other researchers can test. Funding and institutional support make the work possible, but they do not settle its conclusions.
Emerald AI software will adjust flexible computing workloads in response to grid conditions, while San Diego Gas & Electric will advise on utility operations and integration. UC San Diego will also develop a planning tool for large electrical loads and join a workforce program designed to create data-center career pathways for disadvantaged communities.
For UC San Diego and AI Data Centers, a large number can describe reach without describing effectiveness. The more revealing questions concern who was included, what was measured, how outcomes changed and whether another team could reproduce the finding.
The local stakes in UC San Diego and AI Data Centers become meaningful where a public announcement meets daily life. The useful questions are who can use the program or place, who carries the cost, when the change becomes visible and what evidence will show whether it worked.
The next questions about UC San Diego and AI Data Centers concern methods, limitations, peer review, replication and access to underlying data or code where appropriate. If the work points toward a practical application, later testing will matter more than the promise of eventual use.