A new $8.48 million project at UC San Diego is demonstrating how AI data centers can dramatically reduce energy waste and become active partners in grid stability, rather than passive consumers of electricity. The San Diego Supercomputer Center (SDSC) will serve as a testing ground for innovative power architecture that eliminates unnecessary energy conversion steps, potentially transforming how artificial intelligence infrastructure draws power from the grid.

Why Does AI Data Center Power Consumption Matter So Much?

The energy demands of modern AI systems have become staggering. A single rack of AI hardware can require up to 50 times more electricity than a conventional server rack from just a few years ago. This explosive growth in power consumption creates two interconnected problems: first, the sheer volume of electricity needed, and second, the inefficiency of how that power is delivered to computing equipment.

Traditional data centers rely on multiple conversion stages before electricity reaches the servers, with each step wasting energy as heat. The UC San Diego project addresses this fundamental inefficiency by developing a power delivery system that moves electricity more directly from the grid to computing equipment, eliminating unnecessary conversion stages that squander power.

How Can Data Centers Become Grid Partners Instead of Grid Burdens?

• Solid-State Transformers: San Diego-based Alderbuck Energy is developing bidirectional solid-state transformers that convert medium-voltage alternating current directly into 800-volt direct current through a single compact device, simplifying the entire power delivery chain.
• Intelligent Software Orchestration: Emerald AI's flexibility management platform, called Emerald Conductor, acts as an intelligent interface between a facility's compute system and its power infrastructure, dynamically adjusting how much electricity AI jobs draw by slowing, shifting, or briefly pausing batchable workloads.
• Real-Time Grid Response: The combined system allows data centers to maintain reliable internal power while responding to broader grid conditions, transforming passive electricity consumers into active grid resources that can help stabilize the power network.

The technology has already shown promise in real-world trials. In a peer-reviewed study published in Nature Energy, Emerald AI working with NVIDIA, Oracle, Salt River Project, and the Electric Power Research Institute successfully reduced the power draw of a live AI cluster in Phoenix by 25% for three hours without compromising performance commitments. More recent demonstrations at Nebius's AI Factory in London showed that high-performance AI infrastructure could cut electricity demand by up to 40% without any performance loss.

"Data centers are no longer just consumers of electricity, as they can be active participants in grid stability," said Brian Balderston, director of infrastructure and data centers for SDSC's Research Data Services Division.

Brian Balderston, Director of Infrastructure and Data Centers, SDSC Research Data Services Division

What Are the Projected Benefits of This New Architecture?

The UC San Diego project aims to serve two megawatts of AI computing load while achieving three major improvements. First, it will reduce the footprint of power equipment by more than 50%, meaning less physical infrastructure takes up less space in data centers. Second, it projects energy savings of approximately 25%, a significant reduction in wasted electricity. Third, by simplifying the power delivery chain, the system is expected to lower installation and operating costs for data center operators.

Beyond the immediate efficiency gains, the project includes what organizers describe as one of California's first equity-centered data center workforce development efforts, creating pathways into careers in the rapidly growing AI and energy infrastructure sectors. This addresses a critical gap as demand for skilled workers in these fields continues to accelerate.

"This research has the potential to fundamentally transform how AI data centers interact with the electric grid. By cutting costs and shrinking our environmental footprint, it paves the way for sustainably scaling this critical technology," said Chancellor Pradeep K. Khosla.

Pradeep K. Khosla, Chancellor, UC San Diego

How Will This Data Help California Plan for Future Energy Demands?

The UC San Diego team will develop a flexible load-capacity tool that utilities and policymakers can use to better understand how large electricity users, including data centers and electric vehicle fast-charging stations, can participate in grid management. Most interconnection planning currently treats a data center's peak demand as a fixed number that the grid either serves or doesn't serve, but real-world demonstrations have shown this assumption is incorrect.

The validated models and real-world workload flexibility data produced by the SDSC facility's diverse mix of scientific and AI research workloads will feed directly into this planning tool. The facility's unique position within UC San Diego's campus microgrid, which has its own distributed energy resources, provides an unusually rich testbed for understanding how AI infrastructure can adapt to grid conditions.

"Across a series of real-world demonstrations, we've shown that assumption is wrong. A meaningful share of AI computing has flexibility that can be used to actively support the grid," noted Ayse Coskun, chief scientist at Emerald AI and a professor at Boston University.

Ayse Coskun, Chief Scientist, Emerald AI and Professor, Boston University

As AI infrastructure continues to expand globally, the pressure on electrical grids intensifies. Projects like the UC San Diego initiative demonstrate that the solution isn't simply building more power plants; it's fundamentally rethinking how data centers interact with the grid. By turning AI facilities into flexible, responsive participants in grid management, utilities can better balance supply and demand while data center operators reduce costs and environmental impact. The real-world evidence emerging from these demonstrations could reshape how California and other regions plan for the energy demands of the AI era.