AI’s electricity problem may have a solution the size of a watch crystal

Every time you use a chatbot, stream a video, or ask your phone a question, a data centre somewhere is burning electricity. A lot of it. And a surprising amount of that electricity is wasted before it ever reaches the chip doing the actual work.

Engineers at the University of California San Diego have built a prototype chip that attacks this waste at its source. Their design, published in Nature Communications, could make this conversion significantly more efficient – using a technology borrowed from guitar pickups and ultrasound machines.

The hidden cost of powering a GPU

Most people think of power consumption in terms of what a chip does: running calculations, loading models, generating text. But before any of that happens, the electricity arriving at a data centre needs to be converted into a form processors can actually use.

Data centres receive power at 48 volts. The GPUs inside them run at between 1 and 5 volts. Something has to bridge that gap cleanly and efficiently, trillions of times per second. That job falls to a component called a DC-DC step-down converter – found in virtually every piece of electronics ever made.

The problem is that today’s step-down converters rely on magnetic coils called inductors, and those inductors are running out of room to improve. Decades of engineering have pushed them close to their physical limits.

“We’ve gotten so good at designing inductive converters that there’s not really much room left to improve them to meet future needs,” said Patrick Mercier, the study’s senior author and a professor of electrical and computer engineering at UC San Diego.

Every fraction of a percent lost in that conversion step, multiplied across millions of servers running around the clock, adds up to a significant and growing electricity bill.

Crystals that vibrate to carry power

The UC San Diego team’s answer is to replace the magnetic coil with a piezoelectric resonator. Piezoelectric materials generate electricity when squeezed or bent, and they vibrate mechanically when electricity is applied to them. The same principle powers the crystal inside a quartz watch and the probe that creates ultrasound images.

In a power converter, a piezoelectric resonator stores and moves energy through mechanical vibration rather than through a magnetic field. The physics allows for a much smaller component that can, in principle, handle large voltage differences more efficiently than an inductor.

Previous attempts to build piezoelectric converters stumbled when the voltage gap was large – exactly the situation in AI data centres. The UC San Diego team solved this by pairing the resonator with a network of small capacitors in a carefully designed circuit. The result is multiple pathways for power to flow simultaneously, which spreads the load, reduces waste, and keeps the resonator in its most efficient range.

The prototype converted 48 volts down to 4.8 volts with a peak efficiency of 96.2 percent, and delivered roughly four times more output current than any previous piezoelectric-based design.

Not ready yet – but pointing in the right direction

Mercier is careful not to oversell where the technology stands. The converter works in a laboratory, but getting it into a real data centre will require solving a packaging problem: piezoelectric resonators physically vibrate as they operate, so they cannot simply be soldered onto circuit boards the way most electronic components are.

“Piezoelectric-based converters aren’t quite ready to replace existing power converter technologies yet,” Mercier said in a statement released by UC San Diego. “But they offer a trajectory for improvement. We need to continue to improve on multiple areas – materials, circuits and packaging – to make this technology ready for data centre applications.”

That honesty matters. AI data centres are expected to consume increasing shares of national electricity grids over the coming decade, and the gap between where today’s converters can take us and where AI’s power demands are heading is only going to widen.


Sources

Primary paper Jae-Young Ko, Wen-Chin B. Liu, Patrick P. Mercier. “A Hybrid Piezoelectric Resonator-based DC-DC Converter.” Nature Communications, 17 March 2026. https://doi.org/10.1038/s41467-026-70494-0

Institutional press release University of California San Diego. “New chip design could boost efficiency of power management in data centers.” 8 April 2026. https://today.ucsd.edu/story/new-chip-design-could-boost-efficiency-of-power-management-in-data-centers