Penn State researchers reported software that reduces data center cooling energy use by up to 25 percent, with a 24 percent reduction for cryptocurrency mining loads and a better than 8 percent improvement in Bitcoin mining profitability. The university's account credits a team led by architectural engineering professor Wangda Zuo, with doctoral candidate Viswanathan Ganesh as first author and co-authors Hongjun Li and Jiyuan Sui. The work was accepted for the IEEE ITherm conference in May 2026 and funded by the National Science Foundation and the Department of Energy.
The method is physics-informed reinforcement learning trained against a digital twin of the facility. An agent learns to set cooling controls in real time based on climate and economic conditions, rather than holding the fixed conservative setpoints most operators run.
Cooling accounts for roughly 40 percent of a data center's total electricity. A 25 percent cut in that slice is a 10 percent reduction in whole-facility energy with no change to the chillers, the pumps, or the racks. That is the part worth sitting with. The biggest barrier to cooling efficiency has rarely been the absence of better hardware. It has been operators running wide thermal margins because a controls mistake risks a thermal event and a hardware refresh does not pay back fast enough to justify the disruption. Software that closes that margin safely captures savings that were already physically available and left on the table.
The conservative-setpoint problem exists because operators cannot safely experiment on a live floor. A digital twin removes that constraint by letting the agent learn aggressive policies in simulation before any of it touches production. The physics-informed part is what keeps the learned policy from proposing states that violate real thermal limits. This is the same logic behind running facilities at higher water temperatures, the way Nvidia's 45C hot-water cooling eliminates chillers. Both move the operating point toward the efficient edge. One does it with architecture, the other with control policy, and they compound when used together.
This is not staying in a conference proceeding. Zuo co-founded Glacian Technologies, which has a commercial agreement to integrate the software into the Alerify data center in Harrisburg, Pennsylvania, with integration planned for late 2026. A controls optimization layer with a named commercial deployment is a different object than a research result. It is the start of a product category: a software tier that sits on top of existing mechanical plant and sells a percentage of the cooling bill back to the operator. For cooling vendors, that is both an opportunity and a margin question. The hardware sale is one-time. The optimization layer is recurring, and it is increasingly where the defensible value sits.