Lighting the Way: Where Photonics Fits Into the Next Generation of AI

UCI researcher Maxim Shcherbakov explains why computing with light could make AI systems faster and more energy efficient, and what challenges remain.

Sept. 15, 2026 — As AI becomes embedded in modern life, the hardware powering it is running up against its physical limits. Meanwhile, its energy consumption is straining energy grids and increasing scrutiny over AI's environmental footprint.

Engineers are racing to find more energy-efficient approaches to AI hardware, and light-based, or photonic, technologies offer promising alternatives. Using light to perform certain computations could lead to more efficient AI, which would reduce its energy footprint and shrink the its environmental impacts.

UC Irvine professor Maxim Shcherbakov examined the current state of photonics in the AI hardware landscape and what challenges lie ahead in a new Perspective in Nature Photonics, published on September 14. Co-authored by UCI researchers Dmitry Nechipurenko and Abhishek Gautam, the piece maps out where light-based accelerators might make an impact and what the field needs to get there.

We spoke with Shcherbakov, assistant professor of electrical engineering and computer science at the Samueli School of Engineering, to learn more.

What is a photonic accelerator, and why is the AI industry looking to light-based computing as a solution to its growing energy and performance demands?

The core idea behind optical computing is simple: instead of using light only to move information around, what if we could use it to actually compute? That is essentially what my lab does — we pick components of traditional computers and replace them with their photonic counterparts.

The key infrastructure currently driving most AI computation — the servers running behind ChatGPT, Claude, and similar tools — relies on GPUs, or graphics processing units. Much of the work a GPU performs involves matrix and vectors operations, which are fundamental to neural networks. Every matrix-vector multiplication implemented in an electronic circuit produces losses in the form of heat. At the scale of modern AI systems, these losses add up, and that's why AI servers require so much cooling.

If you were to use photons instead, the heat problem would be far less significant. Photons have no charge, so moving light through a device does not create the same losses as moving electrons through the circuits. As a result, the operations using photons can, by their nature, be performed with very little dissipation.

The piece draws a key distinction between cloud-oriented and edge-oriented photonic systems. What is that distinction, and why does it matter for how this technology develops?

The picture I have described — simply replacing electrons with photons without introducing new tradeoffs — is fairly pie-in-the-sky. The field has evolved from fundamental demonstrations that light can assist with AI-related tasks to a more rigorous understanding of exactly how fast and energy-efficient photonic processors actually are.

It turns out that if you try to build a direct optical analog of a GPU, the technology is not quite there yet. Most photonic processors that exist today, including those from influential startup companies, are still playing catch-up with standard GPUs.

However, there are use cases where photonics can outperform current electronic technology, particularly at the edge. Cloud computing, running on GPU servers, is designed to be general — it can run a wide range of AI models and workloads. For edge devices, the computational loads are typically much smaller because the data volumes are more modest. But there are real constraints: you cannot put a server-sized GPU in a self-driving car. Edge devices need to be small, and their latency — reaction time — is sometimes more critical than raw computation throughput.

So in a sense, the cloud-edge distinction matters because photonic accelerators may find their first practical use in specialized edge systems rather than as a direct replacement for GPUs.

What are the biggest technical obstacles standing between where photonic accelerators are today and widespread deployment in cloud data centers?

For general-purpose photonic accelerators, the main bottleneck is the conversion between the electronic domain and the photonic domain. You need lasers modulated by electricity and detectors that convert light back to electricity, and most of the energy and bandwidth is lost at these conversion points. Until we solve how to efficiently integrate photonic chips into the existing electronic ecosystem, we will remain bottlenecked by these fundamental constraints.

At the edge, things look more promising. Many of the signals that edge processors work with are already optical. Images are the most prominent example — consider a self-driving car continuously recording visual information, which can add up to a substantial workload. Photonic co-processors could enable AI-related image classification before the information even enters the electronic domain. A small device attached to each camera could classify objects — this is a person, this is a dog, this is a street light — before the optical signal enters the electronic domain, dramatically reducing the load on the electronic system.

How realistic is the promise of "greener AI" through photonics, and how close are we to realizing that promise?

Some projections suggest that AI servers could account for a significant share of global energy consumption by the end of the decade. We really want to reverse that trajectory. AI is remarkable, and I know many people appreciate the productivity gains it provides. But we need to step up our game in making it greener. I can't speak to every approach — I'm not a specialist in electronic systems — but I'm sure there are many researchers working on making electronic AI computing more efficient.

What I can say is that photonics offers a compelling approach to greener AI, rooted in the fundamental nature of light and the potential to perform certain computations with low dissipation. Once we've cleared the bottlenecks I described, we hope to establish — as a community — a more energy-efficient platform for AI, even if it's initially for specialized tasks. Ultimately, the goal is to reduce AI's overall energy and environmental footprint while continuing to improve its performance and accuracy.

• Paul McQuiston