How computers are learning to use light

How computers are learning to use light

Something amazing happens inside a photonic chip. Light travels through microscopic circuits. Optical modulators encode information onto those signals, and photodetectors read the results. In one recently demonstrated system, four external lasers fed a silicon photonic chip containing more than 16,000 photonic components, with optical signals representing the numbers used in matrix calculations.

It sounds almost too elegant for a computer. 

We’re used to thinking about computing in terms of electricity: transistors switching, electrical signals moving, processors carrying out billions of operations. But photonics adds another medium to that toolkit – it uses light to transmit information and has the potential to perform particular kinds of computation. 

The result is that computers are becoming bilingual. 

Light is joining the calculation 

In April 2025, two papers published in Nature gave us a glimpse of how far the idea has already progressed.

The first introduced a photonic AI processor capable of executing advanced AI workloads including ResNet and BERT, alongside an Atari deep reinforcement-learning algorithm originally demonstrated by DeepMind. The researchers reported near-electronic precision across many workloads.

That’s significant because the photonic system was taking part directly in computation.

Modern AI relies heavily on tensor and matrix operations: repeated calculations involving large arrays of numbers. In photonic systems, optical signals can perform multiplication and accumulation as they travel through guided circuits. The behaviour of light becomes part of the way those numerical operations are carried out. 

There’s a pleasing physicality to it. Numbers are encoded into optic signals, and those signals pass through a carefully designed circuit. Their interactions help produce the result – which means light is doing useful computational work.

16,000 photonic components on a single chip

The second Nature paper, published on the same day, pushed the idea towards a larger integrated system.

Researchers built a photonic accelerator containing more than 16,000 photonic components on a single chip. It was designed for matrix multiply-accumulate operations – a type of calculation used extensively in machine learning. 

The accelerator operated at frequencies of up to 1 GHz, with reported cycle latency as low as three nanoseconds. The researchers also used the system for computationally difficult optimisation problems. In those experiments, the minimum effective loop latency reported for the Ising solver was five nanoseconds. 

And the partnership happening inside the hardware is particularly intriguing. The photonic chip performs the optical matrix operations. A co-integrated electronic chip supplies logic, memory and control. The two are connected using an advanced 2.5D hybrid packaging approach.

This offers a useful way to think about where computing may be heading. Electronics can continue doing the jobs they’re really good at, while photonics can take on tasks where optical signals bring useful advantages. 

The two technologies can work together inside the same system. And when you look at it like that, the bilingual computer feels more like an emerging engineering philosophy.

Light is moving closer to the network switch

Photonic computing is only one part of this field; the other is communication. 

AI systems rely on huge numbers of processors working together. Those processors need to exchange vast quantities of information, often across complex networks inside data centres.

Silicon photonics provides a way to move some of that data using light.

In May 2026, NVIDIA announced that Spectrum-X Ethernet Photonics was in production as part of its Vera Rubin platform. The technology combines co-packaged optics with Spectrum-X switching, and NVIDIA says the switches use 200Gb/s SerDes – serializer/deserializer technology for converting and transmitting high-speed data streams. 

Co-packaged optics brings optical technology into much tighter integration with the switching system. NVIDIA is positioning it as infrastructure for connecting extremely large AI computing clusters.

For anyone following the evolution of computing hardware, this is a production milestone worth taking note of. Photonics is appearing in commercial infrastructure, not just in laboratory breakthroughs.

Then comes the factory

Of course, once a technology begins moving towards wider use, we face a whole new set of engineering questions. 

  • How do you manufacture it at scale? 
  • Which optical materials work best? 
  • How do you package photonic and electronic components together? 
  • Which processes can reliably produce large numbers of devices?

And those questions are attracting serious attention.

On 29 July 2026, GlobalFoundries signed a letter of intent with the US Department of Commerce under which the department’s CHIPS Research and Development Office is expected to award the company $300 million for next-generation silicon photonics. The proposed programme covers optical materials, wafer technologies and advanced packaging, including work related to near-packaged and co-packaged optics. 

Reuters reported the planned support as part of an effort to develop faster, more energy-efficient optical links for AI computing systems. 

In a press release, GlobalFoundries said its SCALE silicon-photonics platform is targeting 400Gb/s performance and a fivefold increase in energy efficiency over current-generation implementations. 

The wider field now includes lasers, modulators, photodetectors, foundry processes, packaging technologies, networking equipment and software designed to help optical and electronic systems work together.

And that’s part of what makes photonics so interesting: it touches several layers of computing at once.

So what happens when light gets more jobs?

One useful way to follow the field is to watch where light is moving closer to computation. 

  • Optical links connecting processors. 
  • Photonic circuits carrying out specialised matrix calculations.
  • New manufacturing techniques that bring optical and electronic components together in the same package.

The recent evidence shows that progress is underway. Each of the developments we’ve touched on here tackles a different part of the computing stack – but they all suggest real potential for machines in which electrons and photons work side by side.

And for us at LEAP, that creates some exciting questions: 

  • How much computation can move into optical circuits?
  • How close can lasers and photonic components get to processors?
  • Which applications benefit most when data can travel optically?
  • And what entirely new kinds of hardware might appear once engineers become fluent in both?

Those questions are starting to move from research papers into real products and manufacturing programmes. Computers are fluent in electronics – but now they’re learning a second language.