University of Pennsylvania, engineers have created a chip that allows algorithms to function not with circuits and silicon, but with pulses of light.

It’s the world’s first photonic processor that can train artificial intelligence systems in real-time using beams of light. This technology could upend how we train the AIs that increasingly shape our lives.

Training AI is an energy hog. This chip flips the script

In just a couple of years, AI has gone from promising “someday” technology to something we use pretty much every day. Behind the scenes, neural networks make it happen — massive webs of artificial “neurons” tuned through training. These neural networks are very robust but they need endless calculations and electricity-hungry hardware.

Today, most AI runs on specialized chips called GPUs. These chips are fast but energy intensive. The cost of training cutting-edge AI models like GPT-4 can run into millions of dollars — and even more in carbon emissions.

“Nonlinear functions are critical for training deep neural networks,” says Liang Feng, the paper’s senior author, referring to the training phase. “Our aim was to make this happen in photonics for the first time.”

That’s where the new chip from Penn Engineering steps in. It trains neural networks entirely with light, slashing energy use and dramatically increasing speed.

It’s not just efficient. It’s something more radical: a new computing paradigm.

How does this work?

To understand the leap, let’s talk about how AI “thinks.”

At its core, most AI systems today use neural networks. These networks are composed of nodes (akin to neurons) connected by weighted links. When data flows through the network, some connections amplify the signal, others dampen it, and only certain pathways activate.

But there’s a twist: these systems don’t just rely on adding and multiplying numbers. The real magic happens in the nonlinear functions — mathematical operations where small inputs can lead to big changes. Nonlinearity is what gives AI its power to detect patterns, recognize faces, or drive cars.

For decades, engineers dreamed of using photons instead of electrons to compute. Light is fast and doesn’t heat up like electricity. It can also travel in parallel beams, handling multiple signals at once.

But light has a problem. It travels in straight lines, and in most materials, it behaves in a linear way. That means it’s great for adding things up — but terrible for the nonlinear twists AI needs. While many teams developed light-power chips capable of handling linear mathematics, none managed to use light for non-linear functions. Until now.

The secret lies in a special semiconductor

The team’s innovation starts with a special semiconductor that reacts to light. Think of it as a thin film that can become more or less transparent, depending on how you shine light into it.

The team then uses two beams: one carries the data (“signal” light), and the other acts as a kind of invisible hand (“pump” light), sculpting how the material responds. By precisely tuning the shape, intensity, and spatial pattern of the pump light, they can control how the signal light is absorbed, amplified, or altered. This interaction simulates nonlinear mathematical functions, the kind used in neural networks to make decisions.

And it’s all done without changing the chip’s physical structure.

“We’re not changing the chip’s structure,” says Feng. “We’re using light itself to create patterns inside the material, which then reshapes how the light moves through it.”

A new paradigm for AI

This type of design is excellently suited for machine learning. Because the chip can simulate different nonlinearities on demand, it can adapt its behavior during training. This is what makes it programmable — not just in setup, but in operation.

The team tested it on classic machine learning challenges, like distinguishing species of iris flowers or recognizing spoken words. The chip trained itself, adjusting its internal light patterns to improve accuracy over time.

For the iris dataset, it achieved 96.7% accuracy — outperforming comparable linear photonic systems. When tasked with identifying spoken words like “Bird” and “Tree,” it reached over 91% accuracy using far fewer connections than traditional digital networks. The team also proved that they can achieve major simplification in hardware and massive savings in energy.

The technology is still in its early stages. The chip uses sophisticated optical setups and precise holographic light patterns to control behavior. Scaling this up will take engineering and manufacturing breakthroughs.

But the core concept — programming light to compute nonlinearly — is now proven.

Next steps include integrating the chip with existing silicon photonic platforms, increasing the number of inputs and outputs, and exploring real-world applications in vision, speech, and robotics.

You might not notice it today, but this could be the moment we remember as the dawn of light-trained AI. It won’t just be faster, it will be fundamentally different.

The study was published in Nature Photonics.