Introduction
For decades, building computers that process information with light instead of electricity remained largely a research curiosity. Today, a new generation of startups is revisiting those ideas, driven by an urgent problem: artificial intelligence inference is consuming enormous amounts of electricity in data centers worldwide. The optical computing sector has grown from an estimated $0.4 billion in 2022 to roughly $2.22 billion today, as investors bet that photons can handle certain AI math more efficiently than electrons.
What Happened
The concept dates to a 1978 proposal by researcher Joseph W. Goodman, who demonstrated that light could perform matrix-vector multiplication—a core operation in modern AI—by passing beams through carefully designed optical elements. Because the multiplication occurs passively as light moves through the hardware, its energy cost does not grow as quickly as it does in conventional electronic chips.
Today’s optical computing startups are taking different technical paths. Lumai, based in Oxford, England, uses a free-space approach where light travels through air and is shaped by programmable spatial light modulators. The company says its Iris Nova optical server is available for evaluation, with a follow-up system designed to handle the compute-heavy “prefill” phase of AI inference using about one-tenth the energy of a GPU.
In Austin, Texas, Neurophos is packing metamaterial-based modulators onto a chip just five millimeters across. The company claims its approach switches more than 10,000 times faster than standard liquid-crystal devices and is targeting performance beyond Nvidia’s Blackwell architecture, with rack deliveries planned for 2028.
Other firms are betting on integrated photonics built directly on silicon. Q.ANT in Stuttgart, Germany, and Lightelligence in Shanghai are manufacturing optical processors on conventional semiconductor lines. Q.ANT’s systems are already in use at German supercomputing centers, and the company recently sold a unit to cloud provider Ionos. Meanwhile, Israel-based LightSolver is using laser cavities to solve differential equations, and Leeds-based Optalysys is inserting optical math directly into data-center networking paths to handle memory-bound tasks without redesigning the entire accelerator.
Why This Matters
AI inference—the process of running trained models to answer user queries—accounts for the majority of AI’s energy footprint. As models grow larger and context windows stretch longer, data centers are under pressure to find alternatives to power-hungry GPUs. Optical computing offers a potentially different scaling relationship: efficiency can improve as matrices grow larger, which aligns well with the demands of modern generative AI workloads.
Even if optical chips do not replace GPUs outright, the broader investment is already changing data-center infrastructure. Companies such as Nvidia and TSMC are investing in co-packaged silicon photonics for circuit switching and interconnects, which could eventually open the door to more optical computation inside the network itself.
Key Takeaways
- The optical computing market has grown from roughly $0.4 billion in 2022 to about $2.22 billion today.
- Light-based matrix multiplication could theoretically scale energy use linearly rather than quadratically as problem sizes increase.
- Startups are pursuing free-space optics, integrated silicon photonics, metamaterials, and in-network processing.
- Real-world deployments are already appearing in supercomputing centers and cloud infrastructure.
- Major challenges remain, including the energy cost of encoding data into light, analog-to-digital conversion, and the fact that optics cannot accelerate every part of an AI workload.
- Industry targets range from 10- to 80-fold energy-efficiency improvements over current GPUs.
Conclusion
Optical computing is unlikely to displace electronic GPUs in the near future. Researchers caution that encoding data into optical signals and converting results back to digital can erase theoretical energy savings, and Amdahl’s law means only part of an inference workload can be accelerated by light. Still, the field is moving from theory to hardware, with commercial systems shipping and production facilities being planned.
The most likely outcome is a heterogeneous data center where CPUs, GPUs, and photonic processors each handle the tasks they are best suited for. For AI developers and data-center operators, optical computing is no longer just a laboratory curiosity—it is an emerging tool in the effort to make artificial intelligence sustainable at scale.

Discussion
Join the conversation
Thoughtful reactions, questions, and follow-up ideas help shape the next story.