UC Santa Barbara funds five projects to cut computing’s energy demands

UC Santa Barbara’s Institute for Energy Efficiency has awarded seed funding to five projects aimed at reducing the energy demands of advanced computing. Announced by the university on October 27, 2026, the awards back work ranging from AI agents that use fewer tokens to cooling systems that send liquid where processors need it.

The selections come through the institute’s 2026–27 Research Seed, Software Impact, Gauthier Quantum and Photonics Seed grant programs. These programs support early-stage work intended to open research directions and lay the groundwork for larger investments.

Give AI less work to do

Xin Eric Wang’s team is developing an open-source system to train AI agents to use shorter reasoning paths and reduce information they repeatedly process. The goal is to maintain performance while using a fraction of the tokens—the units models use to reason and respond. Earlier work from the group lets models estimate the likely cost of different routes to a solution.

Mahnoosh Alizadeh and Ramtin Pedarsani are pursuing a different adjustment: match computing effort to each request’s difficulty. Their proposed learning system would select the lowest-energy model configuration that meets a required quality level, reserving the full model for harder requests. They plan to measure energy per correct response, rather than judging efficiency only by model size or computation counts.

Move coolant with the workload

Yangying Zhu and Igor Mezić are developing chip-level cooling with AI-enabled controls that redistribute coolant among central and graphics processors as workloads change. Rather than cooling every component equally, the system would direct resources toward busy processors and reduce unnecessary cooling of lightly loaded devices.

Tiny silicon channels would bring coolant close to heat-generating electronics, including designs where liquid absorbs heat as it becomes vapor. Predictive models would anticipate thermal changes and adjust coolant flow. The researchers estimate a successful demonstration could cut AI data-center cooling energy use by 20%, not total facility energy use.

Change the hardware’s operating demands

Sriram Krishnamoorthy and Galan Moody will investigate atomic-scale defects in gallium oxide that could emit individual photons carrying quantum information. Their goal is to demonstrate single-photon emission from the material at or near room temperature, potentially enabling compact quantum light sources with less energy-intensive infrastructure.

Dmitri Strukov is exploring specialized circuits for probabilistic computing tasks that rely on random-number generation and mathematical operations called dot products. The project asks whether purpose-built hardware can perform these tasks faster and more efficiently than conventional processors. It will generate preliminary data toward a scalable framework and a larger external funding effort.