Accidental Rainbow-On-A-Chip Discovery Could Tame AI's Soaring Energy Demands

A team of Columbia University engineers may have stumbled on a key to cooling the AI industry's energy fever. While trying to boost the power of chips for LiDAR—laser-based sensing used most notably in self-driving cars—the researchers noticed something strange. "As we sent more and more power through the chip, we noticed that it was creating what we call a frequency comb," said Andres Gil-Molina, co-lead author of a new paper titled "High-power electrically pumped microcombs." That "comb" turned out to be more than a curiosity; it might be the building block for a new generation of ultra-efficient optical links that could someday replace energy-hungry electrical wiring inside AI datacenters.

The discovery centers on "microcombs," tiny ring-shaped resonators etched into silicon nitride. When light circulates inside these rings, nonlinear optical effects generate dozens of evenly spaced laser frequencies—a sort of miniature rainbow that can carry vast amounts of data in parallel. Researchers have been chasing the dream of electrically pumped microcombs for years, but their power output was too low to be useful. Gil-Molina's team changed that by designing chips with "normal group velocity dispersion" and leveraging a nonlinear effect known as self-injection locking to stabilize the output. The result: combs delivering up to 158 milliwatts of on-chip optical power, with linewidths as narrow as 200 kilohertz—apparently, stable enough for communications and precise timing.

Those numbers might not sound dramatic, but in the world of integrated photonics, they're a serious breakthrough. Earlier versions of electrically pumped combs struggled to produce even a few milliwatts. By surpassing that barrier, the Columbia group effectively demonstrated that these light sources can be made small, cheap, and powerful enough for real-world systems. That's a milestone on the road to multiplexed optical interconnects—high-speed data links that move information with photons across multiple wavelengths of visible light.


Why does that matter for AI? Because as datacenters pack in more GPUs and neural processors, electrical wiring has actually become a major bottleneck, both in bandwidth and power consumption. Optical links based on microcombs could transmit multiple data channels simultaneously while burning a fraction of the energy. In principle, that means faster AI training, cooler chips, and much lower electricity bills for the hyperscalers building planet-sized models.

The work was led by Andres Gil-Molina and Yair Antman in the labs of Michal Lipson and Alexander L. Gaeta at Columbia University. Michal Lipson is a pioneer in silicon photonics, and her group has a reputation for turning fundamental physics into practical devices. Their latest result hints at a future where the critical interconnects that enable modern "AI factories" are literally powered by tiny rainbows—born not from design, but from a lucky detour in the lab. Sometimes, the fastest path to efficiency starts with shining a little too brightly.

_{Top image: A representation of the a frequency microcomb. Image: Columbia University.}