Deploying AI on a large scale? ‘Use photonic chips’
Photonics can play a crucial role in making AI more energy efficient.
Published on September 11, 2025
Our DATA+ expert, Elcke Vels, explores AI, cyber security, and Dutch innovation. Her "What if..." column imagines bold scenarios beyond the norm.
AI helps society move forward, but it has a downside: it consumes energy. So there is work to be done. Photonics can play a crucial role in making AI more energy efficient. By transferring data using light instead of electricity, AI chips and other systems consume much less energy and can also process information faster. We interviewed Martijn Heck, professor at the Photonic Integration group at Eindhoven University of Technology (TU/e). Heck is one of the speakers at the Watt Matters in AI conference.
What is the key message you want to convey at Watt Matters in AI?
"The message is quite simple: computers and chips are becoming increasingly powerful, but we also need to ensure that they are energy efficient. Moore's Law shows that the number of transistors on a chip doubles approximately every two years. This means that computers, supercomputers, and even the chips in your phone are becoming faster and more powerful. But if every transistor used the same amount of energy, we would need enormous amounts of energy, which is obviously not feasible. That's why there is Koomey's Law, which states that the energy per calculation must be halved every two years. This is working surprisingly well; your current phone is many times more powerful than it was ten years ago, but it can last a day on a single battery. Photonics, among other things, can play an important role in making systems and chips more energy efficient."
Tell us: how can photonics make AI more energy-efficient?
“One of the biggest challenges remains the transport of data. Whether it's exchanging data between servers in an AI data center or moving data on a chip itself. Currently, about half of the energy goes to transport. This is because electrical signals encounter resistance from copper wires: they lose energy in the form of heat. This is where photonics can play an important role, because data can be moved much faster and more efficiently with light instead of electrical signals."
What is a noteworthy development in the world of photonics that you are enthusiastic about?
"Nvidia, for example, has developed prototypes of chips whose edges not only have metal pins, but also photonic transceivers – small transmitters that communicate via fiber optics. They call this co-packaged optics. In short, by integrating light directly into the chip architecture, the network itself becomes much more efficient and more power can be directed to the calculations instead of to data traffic. This is crucial for AI factories with hundreds of thousands of GPUs, among other things."
What is your main focus within photonics?
“My group is strong in integration. In the media, you constantly hear about ‘groundbreaking solutions’. For example, about someone inventing a new laser or modulator. But if such a component does not work well with the others, then a groundbreaking development is not of much use. We are the only group in the Netherlands that focuses on this kind of photonic integration and ensures that all these different components work together. An important challenge is to make technologies that are still new producible."
Do you have an example of a project you have worked on?
"In the past, for example, we have worked on waveguides on a chip. We use these waveguides to direct light, a bit like mini fiber optics. We have minimized the losses for the light so that we can build small networks on a chip. Now we are working on adding small switches through what we call heterogeneous integration. These switches must, of course, be as energy-efficient as possible. This will ultimately enable us to build chips that can switch thousands of gigabits per second in the nodes of a data center or AI supercomputer."
In short, do you think that (integrated) photonics will make AI a lot more energy-efficient in the future?
"Definitely, and it's already happening. The large-scale application of photonics within AI chips may take some time, but they are already being interconnected via photonics in a data center. When we talk about electronic chips, we are talking about the most advanced technology in the world, and we can't really run any faster than we already do. If you design a chip on the most advanced platforms, such as those from TSMC, it can easily cost around €700 million these days. So you have to be sure that the design is correct and that the production process works well. That's why new technologies are always postponed as much as possible. So, only when there is really no other option do we replace the copper lines with optical waveguides. This happened in the past within data centers, and within a few years, it will happen within the chips."
