A new silicon modulator from Japan could help AI data centres move more data using far less energy. Researchers led by Yokohama National University (YNU) have built a slow-light silicon modulator just 320 micrometres long that reaches 200 Gb/s, according to a release published on Asia Research News on 22 October 2026. The study, titled “Reduced-DSP 200G/λ low-voltage slow-light silicon modulator enabled by built-in EO equalizer”, appeared in Nature Communications the same day. This article is based on that release and paper.
What a silicon modulator does
Inside a data centre, much of the traffic between chips and servers travels as light through optical fibre. A modulator is the component that turns an electrical signal into an optical one by switching or shaping a beam of light. As AI workloads grow, data centres need modulators that are faster, smaller and use less power.
Today, the common choice in silicon photonics is the Mach-Zehnder modulator. According to the release, conventional versions are 2 to 4 millimetres long and consume several thousand femtojoules per bit.
How the slow-light silicon modulator works
The YNU device uses a photonic crystal waveguide, a silicon structure patterned with tiny holes. These holes slow light down as it passes through, which strengthens the interaction between light and the electrical signal. That is why the device can be so short.
Slow light normally comes with a trade-off: it limits bandwidth. To get around this, the team built an electro-optic equalizer directly into the modulator. “In this paper, we created a silicon photonic crystal waveguide optical modulator that integrates a built-in electro-optic equalizer to break the bandwidth limitation while preserving the benefits of the slow-light effect,” the researchers say. The equalizer pushes the bandwidth beyond 80 GHz.
The silicon modulator’s key numbers
The modulator is 320 micrometres long. Using simple on-off signalling, it reached 140 Gb/s at an energy cost of 26 femtojoules per bit. Using PAM-4, a format that encodes two bits per symbol, it reached 180 Gb/s without any digital signal processing and 200 Gb/s with a low-complexity 11-tap digital equalizer, according to the paper.
Compared with conventional Mach-Zehnder modulators, the team says this makes the device about one order of magnitude smaller and two orders of magnitude more energy efficient. That is the “reduced DSP” in the title: needing less power-hungry signal processing to clean up the signal.
Why CMOS compatibility matters
Thin-film lithium niobate is another promising material for fast modulators, but it is not part of standard chipmaking. The YNU device, by contrast, can be made with CMOS-compatible processes. “Importantly, slow light can be fabricated using CMOS-compatible processes, simultaneously satisfying high speed, low power consumption and small footprint requirements, meeting the next-generation optical transmitter needs of AI data centers,” the team says.
The researchers see it as a building block for co-packaged optics, where optical components sit right next to processors. We have covered that trend before, including Lightmatter and Flex’s co-packaged optics trays.
Who worked on it
The work involved Yokohama National University, the Institute of Science Tokyo, the National Institute of Advanced Industrial Science and Technology (AIST) and the Photonics Electronics Technology Research Association (PETRA). Professor Toshihiko Baba of YNU is the senior author. Funding came from JSPS KAKENHI, NEDO and Lumentum Japan.
The case for and against
The case for the device is clear in its numbers: a much smaller footprint and much lower energy per bit than the standard silicon approach, made with processes chip factories already use.
The case for caution is that the researchers report a larger optical insertion loss, meaning more light is lost passing through the device. They believe design optimisation can reduce it, but that has not yet been shown. This is also a lab demonstration, not a product.
My view on the slow-light silicon modulator
As a writer, I feel this is the kind of quiet hardware progress that matters more than many headline AI announcements. Energy per bit is a real bottleneck for data centres, and a hundredfold improvement on a key component is significant if it holds up in production. I will be watching whether the insertion loss problem gets solved.
FAQ: slow-light silicon modulator
How fast is the YNU silicon modulator?
It reached 140 Gb/s with on-off signalling at 26 fJ/bit, and 200 Gb/s using PAM-4 with a small digital equalizer.
How small is it?
It is 320 micrometres long, compared with 2 to 4 millimetres for conventional Mach-Zehnder modulators.
What is its main limitation?
The researchers report a larger optical insertion loss, which they believe can be reduced through design optimisation.