Koç University Researcher Proposes Two Optical Systems for AI Computing

Two new systems from Koç University move part of the computation for artificial intelligence out of electronic circuits and into light. Their author, Assistant Professor Uğur Teğin of the Department of Electrical and Electronics Engineering, described the approaches in two separate papers. The goal is clear: make AI faster and cheaper in energy terms. The announcement gives no specific numbers, and that is the main thing to keep in mind when reading it.

Why light is suddenly on the table

AI's appetite for electricity has long since become an infrastructure problem. According to the International Energy Agency, data centers consumed about 415 TWh in 2024, and by 2030 that figure could more than double, to roughly 945 TWh. AI will account for a significant share of the increase. Conventional electronics runs into two limits: heat, and the cost of moving data between memory and compute units.

Optics attacks both problems at once. Light generates almost no heat as it propagates, and many neural network operations, above all matrix multiplications and convolutions, are performed physically as a light signal passes through a suitable medium. The calculation happens as the light propagates, with no clock cycles and no switching of billions of transistors.

What we know about Teğin's work

The published announcement says little. Here is what it does say, in list form.

• Author: Uğur Teğin, assistant professor, Koç University, Department of Electrical and Electronics Engineering.
• Format: two separate scientific papers, two different systems.
• Principle: AI computations are performed by light, not by conventional electronic circuitry.
• Stated goal: faster, more energy-efficient AI systems.
• Not disclosed: accuracy on standard tasks, latency, energy per operation, model size, feasibility of mass production.

The last item matters most. Optical computing has produced impressive lab results for decades and rarely makes it into a data center. Without a comparison against a GPU on a specific task, it is premature to talk about savings.

Where optics usually stumbles

Light-based computing has typical bottlenecks, and any new work in the field gets tested against them.

First: data conversion. Input data almost always lives in electronic form, and every conversion into light and back costs energy. If the conversion costs more than the computation itself, the gain disappears.

Second: nonlinearity. Neural networks need nonlinear activation functions, and light in linear media does not provide them. You either have to go back to electronics or use optical nonlinear effects, which require substantial power.

Third: accuracy. Optical computation is essentially analog, and it is sensitive to noise and temperature drift. That is tolerable for inference, but much harder for training.

Fourth: integration. For a system to become part of a stack with PyTorch and familiar accelerators, it needs drivers, compilers and scalable manufacturing.

The phrase "part of the computational load" in the headline is what looks realistic here. The likely scenario is that an optical block takes over the heavy linear operations, while electronics remain responsible for logic and memory.

What to watch next

For engineers and anyone planning AI infrastructure, the practical plan is simple. Read both papers in full and find three numbers: energy per operation, latency, and accuracy relative to a digital baseline on the same task. Then check what model size the work was tested on: toy datasets like MNIST say nothing about transformers. Finally, check what exactly the optics does inside the system and how much electronics remains around it.

For the market, the signal is broader. Research on photonic computing is multiplying precisely because energy has become the limiting resource for AI. The distance from university groups' work to production accelerators running on light remains large, but the direction deserves attention. When Teğin has measurable metrics, we will return to this topic with numbers.