Korean photonic chip has learned to slow down light. This could change the architecture of AI servers

In Seoul, they presented a photonic circuit that can slow down light and rearrange its behavior right during operation. This is a rare signal for the AI ​​industry: for the first time, there is a component that is not fixed in silicon, but is able to adapt to load.

The light that had to be stopped class=»notranslate»>__GTAG8__

Data centers around the world — and in Russia too — live in a mode of continuous expansion, writes xrust. Generative models are growing, requests are multiplying, and electronic processors are working to the limit. Data center operators are increasingly saying that infrastructure is limited not only by the cost of electricity, but also by the physical limitations of silicon chips.

Optical computing has long been considered a natural solution: light travels faster, consumes less, and the throughput of optical lines is several times higher than electronic ones. But light has a character — it does not know how to wait. It is difficult to delay, synchronize, and temporarily hold it inside the circuit. And without this it is impossible to build normal buffers and memory.

Korean researchers solved this problem not by complicating the resonators, but by changing the light control logic itself.

What exactly did they do

In classic CRIT systems, the delay is achieved due to the interference of several resonators. Everything works, but only in one mode — as it was made, it lives. Engineers who need to change the frequency range or increase latency have to design a new chip. It's expensive, time consuming and doesn't scale well.

A team from Seoul National University proposed a different approach: they combined light and dark modes into one controlled degree of freedom. Added to this are two loop connectors that can be reconfigured.

The result is a diagram where the behavior of light is not fixed. It can be slowed down, accelerated, changed the shape of the pulse, adjusted the frequency range — and all this without making a new device.

An interesting detail: they consider interference between modes as a single parameter, and not as two independent states.

How it looks in action

Inside such a circuit, a light pulse can literally be “supported”—delayed for the required time without destroying its shape. If necessary, you can change the bandwidth, rearrange the frequency range, or redirect the signal so that it arrives exactly at the moment when the system is ready to receive it.

In one of the calculations, the researchers showed that the pulse speed can be changed directly while the circuit is operating. Not after flashing, not after replacing a module — in real time.

There is also an unexplained detail: in one configuration the delay time was higher than expected. The team mentions this in passing, without comment — and it seems like an effect that they cannot yet explain.

Reality check

To understand whether the design would withstand industrial standards, it was run through 3D electromagnetic modeling. They checked everything that usually breaks photonic devices: material loss, differences in the quality of resonators, backscattering, phase errors, thermal crosstalk.

The scheme survived. And because it's built on silicon nitride, the platform used in real photonic chips, it means the technology won't remain a laboratory toy.

How the industry might perceive this

It is important here not to pass off wishful thinking as fact. Therefore, this is only the author’s generalization.

In recent years, companies working with optical accelerators have been striving to make photonic circuits less rigid and more adaptive. It is logical to assume that the ability to dynamically change the delay and waveform within the chip will be perceived as a step towards the very flexibility that optical systems have been lacking.

Data center operators typically look at such developments through an infrastructure lens: if a single photonic component can perform multiple functions, this reduces the number of specialized modules in the racks and reduces power consumption.

It's too early to talk about mass adoption, but the architectural idea itself — tunable light flow — looks like an element that could change the approach to optical computing in the long term.

Historical context

Attempts to slow down light have been made for a long time. In the early 2000s, several research groups in the United States experimented with ultracold atoms, where the speed of light could be reduced to meters per second. These were large laboratory installations, far from practical use.

Now we are talking about a chip that can be integrated into the server infrastructure. The scale of the changes is obvious.

What will happen next

The authors of the study say that the main achievement is not the chip itself, but the design principle. It allows you to rearrange the flow of light inside a photonic circuit as flexibly as programmers rearrange software logic.

The team plans to move on to large systems based on silicon photonics and test the technology experimentally.

At this point the text could be completed with a conclusion, but let’s leave the fact: the researchers are confident that rethinking the physics of resonators opens the way to functions that were previously considered impossible.

Sources

1. https://www.sciencedaily.com/releases/2026/07/260718010149.htm (sciencedaily.com in Bing)
2. https://onlinelibrary.wiley.com/journal/21989119 (onlinelibrary.wiley.com in Bing)
3. https://en.snu.ac.kr/research (en.snu.ac.kr in Bing)

Xrust A Korean photonic chip has learned to slow down light. This could change the architecture of AI servers