A server at the Leibniz Supercomputing Center in Germany is using only light to perform AI inference somewhere inside a beige rack. There are no flickering transistors. There are no electrons racing through copper. The speed at which photons, shaped and guided through a chip, perform math is far faster than that of conventional silicon. The device, manufactured by the German startup Q.ANT, was one of the first commercial photonic processors installed in a functional data center last year. A brief moment, but the kind that usually counts in retrospect.
For many years, photonic computing was in the disappointing category of “promising but always five years away.” The laws of physics held true. The lab demonstrations were impressive. However, the engineering challenges—manufacturing at scale, performing nonlinear operations, and confining light in small spaces—kept pushing the technology back into the realm of academia. Both the AI boom and the accompanying electricity bill have contributed to the changes. Large model training and operation now require so much power that even small efficiency improvements begin to seem revolutionary, and photonics no longer seems exotic.
It was difficult to ignore that change after the work from MIT in December of last year. A group at the Research Laboratory of Electronics developed a fully integrated photonic chip that could perform all of the essential functions of a deep neural network, including the challenging nonlinear ones, in the optical domain. With 92% accuracy, the chip finished a classification task in less than 0.5 nanoseconds. That is competitive with traditional hardware, which wasn’t really an option for an optical system a few years ago.
Microsoft, on the other hand, revealed an analog optical computer earlier this year that was constructed using micro-LEDs, lenses, and components taken from phone camera sensors. Although the prototype is small, the design may be able to scale to billion-parameter models with over 99 percent accuracy, according to the team’s digital twin. Future iterations, according to their researchers, will be 100 times more energy-efficient than current GPUs. That figure is hypothetical. The more intriguing development is that anyone is saying it while maintaining a straight face.
This focus is primarily due to thermodynamics. Photons produce very little heat and move quickly. In contrast, electrons leak energy at every intersection, and at current chip densities, cooling accounts for an astounding portion of an AI data center’s electricity bill instead of computing. IBM takes a more cautious approach, measuring advancement in picojoules per bit and using light to transfer data between chips instead of doing the computation itself. They are between five and ten today. The objective is less than one. It sounds tiny. It isn’t at data-center scale.
Photonic systems continue to struggle with size, as some skeptics—and there are good ones—point out. In a September paper, Cornell researchers made the case that an optical computer with metasurfaces and other cutting-edge techniques that aren’t yet used in mass production could theoretically be only a centimeter thick and perform the linear operations that underpin something like ChatGPT. Additionally, they point out that larger optical setups don’t always perform proportionately better due to diminishing returns.
The ultimate winner might be a hybrid, with silicon handling the messy logic and light handling the heavy linear math, rather than being pure photonic or pure electronic. With its modest but significant investments in the market, Nvidia appears to be making a wager along those lines. The gap between lab and rack is closing more quickly than most people realize, whether photonics completely replaces electricity inside AI inference or just coexists with it. Already, the light is on.
