Five ways to tackle the energy bill
UC Santa Barbara’s Institute for Energy Efficiency has funded five early-stage projects, according to the university’s announcement. The awards cover AI efficiency, adaptive data-center cooling and alternative computing technologies. They provide seed support for research, rather than evidence that these systems have already cut production energy use.
Two projects address the software side. Xin Eric Wang’s team is developing an open-source system to shorten successful reasoning paths and reduce the information that long-running agents repeatedly process. Mahnoosh Alizadeh and Ramtin Pedarsani are exploring energy-proportional AI, which would allocate computation according to the difficulty of a request while meeting a quality requirement.
Cooling the processors doing the work
Yangying Zhu and Igor Mezić are developing cooling controls that redistribute coolant as CPU and GPU loads change. The idea is to avoid spending the same cooling effort on components with very different workloads. The researchers estimate a successful demonstration could reduce cooling energy use by 20%. That is a proposed outcome, not a measured saving from a deployed fleet.
The other projects investigate quantum light sources based on defects in gallium oxide and purpose-built circuits for probabilistic computing. Their inclusion matters: the institute is funding several possible routes to efficiency, rather than treating model compression as the only lever.
What would make the results useful
For operators, the most useful future comparisons would connect energy savings to an unchanged task-quality target. A smaller model or shorter response is not automatically a better service if it stops completing the required work.
The university describes this funding as a way to test ideas and prepare for larger investment. Readers should therefore watch for demonstrations, measurement conditions and released software. The announcement establishes the projects’ aims and teams; it does not establish general savings that another data center could assume today.