Computing at the Speed of Light: Where Optical Computing Stands in 2026

For decades, computers have run on electrons. A growing group of researchers and companies is betting that the next leap will come from light.

Every AI model, cloud service and video call depends on chips pushing electrical signals through tiny copper wires. Those wires are now the bottleneck. They limit how fast data can move, and they turn a large share of a data centre's power into heat. As AI pushes demand for computing ever higher, the industry is turning to an old idea with new urgency: computing with light.

What optical computing is

Optical, or photonic, computing processes information using photons instead of electrons. Rather than currents flowing through transistors, beams of light interact with each other, steered by optical parts such as lenses, waveguides and beam splitters.

Light brings three natural advantages. It is fast and carries enormous bandwidth, because many wavelengths, or colours, can travel through the same channel at once, each carrying its own stream of data. It generates far less heat than electrical signals. And it happens to be very good at matrix multiplication, the core mathematical operation behind neural networks. That last point is why optical computing and AI have become so closely linked.

The field is not about replacing the processor in your laptop. Most real systems are hybrids: a conventional electronic chip runs the software and hands the heaviest maths to a photonic chip, which does it with light.

Where light is already at work

The biggest real-world use today is surprisingly unglamorous: moving data around AI data centres. Modern AI clusters link thousands of chips, and the copper connections between them consume huge amounts of power. Co-packaged optics, which places optical engines right beside the switching chip, is the leading fix. NVIDIA announced silicon-photonics switches using co-packaged optics in March 2025, and the technology is moving from pilots to volume production between 2026 and 2028.

The second application is AI acceleration itself. Photonic processors can perform the matrix maths of AI inference using a fraction of the energy a GPU needs, which matters as electricity becomes the main limit on how much AI the world can run.

Beyond AI, optical computing could cut delays in telecom networks. Today, signals travelling through fibre are converted from light to electricity and back again at each processing step; keeping the whole chain optical removes that delay. Photonics also underpins one branch of quantum computing, where single photons act as qubits and machines may run at room temperature rather than near absolute zero. Further out lie LiDAR for self-driving cars, optical sensing, medical imaging, AR and VR headsets, defence systems and even high-frequency trading.

Inside the labs

Research in 2026 is moving on several fronts at once. In China, a team demonstrated the Liuxing-I chip, which computes in parallel across 100 wavelengths of light at the same time. Earlier optical systems struggled to grow larger and faster without losing precision; this work is a direct attack on that problem.

Other groups are pushing generative AI into light. A chip called LightGen has been reported as the first all-optical processor able to run large generative models, with energy efficiency around 100 times better than leading GPUs. The figure comes from secondary reporting and deserves caution until the original paper is examined.

Some of the most striking work uses hardware that already exists. Researchers in Finland turned images into light pulses and sent them through ordinary optical fibre. The way the light naturally distorted inside the fibre acted as a hidden computing layer, and the system classified images with 91 to 93 percent accuracy, close to conventional AI hardware. Meanwhile, a Bill Gates-backed startup says it has built optical transistors 10,000 times smaller than current designs.

On the commercial side, Lightmatter is the best-known name. By early 2026 it had raised $850 million at a valuation of $4.4 billion, and its Passage L20, a 6.4 terabit-per-second, 3D-stacked optical engine, is scheduled for mass production in the second half of 2026. Ayar Labs, Luminous Computing, PsiQuantum and Xanadu are also prominent, while networking giants Broadcom, Marvell and Cisco have committed to co-packaged optics.

Hard problems remain. Many silicon optical components need constant power just to hold their settings, which wastes energy and limits how densely they can be packed. Building large optical systems is still complex and expensive. And light is fragile: lost photons and noise remain a serious obstacle, especially for quantum machines.

What comes next

The path forward is likely to unfold in stages. Over the next two years, light will take over data movement inside data centres. Lightmatter's own roadmap shows the progression: near-packaged optics in 2026, co-packaged optics in 2027 and photonic interposers in 2028, building towards a full photonic stack that combines computing and interconnects.

After that, photonic accelerators should take on a growing share of AI inference, where energy cost is now the binding constraint. Further out, photonic quantum computers and optical chips in edge and consumer devices become possible, though both depend on solving the scaling and noise problems above.

The money is following. The photonic chip market was worth about $4 billion in 2025 and is projected to reach $9.3 billion by 2032. The narrower photonic computing segment is far smaller, at $0.42 billion in 2025, but forecast to grow 34 percent a year through 2036.

The bottom line

Optical computing will not replace the chip in your laptop any time soon. The real revolution is quieter and already under way: light is replacing wires inside the world's AI data centres, and light-based AI processors are next in line. For businesses that build on AI, the effect over the next few years will show up not as new hardware on their desks but as cheaper, faster and greener AI from the cloud.