Caltech researchers have printed optical circuits from the same glass used in optical fiber directly onto 8- and 12-inch silicon wafers, achieving signal loss at visible wavelengths twenty times better than the silicon nitride record. The team reports the work in Nature, in a paper titled "Towards fibre-like loss for photonic integration from violet to near-infrared" (Nature 649, 338, 2026).

The platform could feed a generation of photonic integrated circuits for optical clocks, chip-scale atomic sensors, ion-trap quantum systems, and lower-energy AI data center links. Whether any of this becomes product depends on manufacturing transfer, but the numbers are the story here.

What did Caltech actually build?

The waveguides use germano-silicate, the glass family of telecom fiber, adapted to lithography-based wafer processing. Instead of straight runs, the paths wind into spirals, packing a long optical path into a few square centimeters. It is the spool of fiber, reimagined at nanoscale.

"For years, we have been working to translate the spool-based fabrication of optical fiber onto silicon wafers, while trying to preserve the fiber's hallmark of ultralow loss," says Kerry Vahala, the Ted and Ginger Jenkins Professor of Information Science and Technology and Applied Physics at Caltech. "This shift toward fiber-like performance, especially in the visible bands, will enable new technologies that benefit from negligibly low circuit energy loss."

The compatibility claim matters commercially. The technique targets the same wafer sizes as standard CMOS lines, which is the difference between a lab curiosity and a foundry option.

How large is the performance gap?

At near-infrared wavelengths, the new devices match the best previous silicon nitride components. The visible band is where the gap opens.

"At visible wavelengths, our recent platform exceeds silicon nitride's record by a factor of 20, and we have more room to improve," says lead author Hao-Jing Chen, a Caltech postdoctoral scholar. The trick is thermal: germano-silicate's comparatively low melting temperature lets the team reflow the waveguide surfaces in a furnace until their smoothness reaches the level of individual atoms, suppressing the scattering loss that has limited visible PICs.

Loss compounds quickly in resonant devices. Every factor of ten reduction in loss yields a factor of one hundred improvement in laser coherence. Lasers built on the platform already show more than a 100-fold coherence improvement over previous designs.

Why does kilometer-scale loss matter on a 2-centimeter chip?

Co-author Kellan Colburn concedes the target sounds, in his word, "a little ridiculous" for chips only 2 centimeters across. The reasoning holds up anyway.

Ring resonators, basic devices in research and data transmission, circulate light repeatedly. A ring of a few millimeters behaves, optically, like a much longer path when per-lap loss is low. "The longer light can circulate, the higher the performance of resulting devices can be," Colburn says. Meter and kilometer loss figures are shorthand for how many laps survive.

The demonstration devices in the paper include ring resonators, several laser types, and nonlinear resonators that generate a range of frequencies.

Who is behind it, and what comes next?

Chen shares lead authorship with graduate student Kellan Colburn. Co-author Henry Blauvelt, a visiting associate at Caltech and CTO of photonics specialist Emcore, frames the data center angle: "Germano-silicate waveguides are also readily adaptable to efficiently transfer light between optical fibers and semiconductor lasers, which is of paramount importance in reducing the overall energy cost of server infrastructure."

Funding came from DARPA, the Air Force Research Laboratory, the UK's Engineering and Physical Sciences Research Council, and the Kavli Nanoscience Institute at Caltech. Vahala calls the results a waypoint: "We haven't gone as far as we want to go, but we've made significant progress over the last five years."

Silicon nitride will not vacate the field quickly, and the paper claims records, not production yields. Still, a fiber-material platform that runs on standard 8- and 12-inch wafers with a further-improvement path sets up a genuine contest for the visible and near-infrared photonic layers that AI interconnects increasingly demand.