What Is Silicon Photonics — and Why Does AI Depend on It?

Silicon photonics is a chip-making technique that moves data using pulses of light instead of electrical current, by etching microscopic optical components — the equivalent of lasers, mirrors, and fiber-optic links — directly onto a silicon wafer with the same manufacturing process used for ordinary chips. Its appeal to artificial intelligence is straightforward: as AI clusters grow to hundreds of thousands of GPUs, the connections between chips are becoming as important as the chips themselves, and light can move far more data with far less energy than copper wire.

How it works

A silicon photonics chip is built on the same kind of silicon-on-insulator wafers used for standard semiconductors, but instead of transistors switching electrical current, it carves nanometer-scale channels called waveguides that guide beams of infrared light much like a hallway guides a person. A modulator encodes data onto that light by rapidly varying its intensity or phase, and at the receiving end a germanium photodetector converts the light pulses back into an electrical signal a processor can use. Because many different wavelengths of light can travel down the same waveguide without interfering with each other, a single optical channel can carry dozens of parallel data streams at once — something copper wiring cannot do.

Why AI data centers need it

Standard data-center wiring uses copper: cheap, reliable, and easy to manufacture. But copper has a hard physical limit — at the data rates modern AI training and inference now require, electrical signals lose strength quickly over distance, so engineers must either shorten the cable, use thicker (heavier, costlier) copper, or accept escalating power draw just to push a signal reliably from one chip to the next. As GPU clusters scale into the hundreds of thousands, those short reaches and power penalties add up, and networking — not compute — becomes the bottleneck.

Light doesn’t have this problem: it travels much farther with far less signal loss and no electromagnetic interference, so an optical link can move data between racks, or even between buildings, almost as easily as between two neighboring chips. The industry’s current focus is co-packaged optics (CPO), where the light-generating and light-detecting chiplets sit directly next to — or inside the same package as — the GPU or network switch itself, shortening the remaining electrical hop to millimeters and cutting the energy it wastes.

Who’s building it

Nvidia has said its next-generation networking gear, including its Quantum-X switches, will lean on co-packaged optics to link GPUs at higher speed and lower power than today’s pluggable optical transceivers. In the US, GlobalFoundries signed a letter of intent for $300 million in CHIPS Act funding to scale up its silicon photonics manufacturing, targeting a roughly fivefold gain in energy efficiency for optical interconnects; the deal also gives the US government a small equity stake in the company. Chip foundries like GlobalFoundries and TSMC are racing to industrialize the process for AI chips, since packaging a photonic chiplet next to a GPU die is itself a form of advanced chip packaging — one more reason interconnects and chip manufacturing are becoming inseparable problems.

None of this is entirely new: silicon photonics has quietly carried internet and telecom traffic for well over a decade. What’s changed is the volume and density of data AI systems need to move between chips sitting centimeters apart, which is pulling a technology built for long-haul networks onto the chip package itself.

In the news

GlobalFoundries recently won a $300 million US grant to scale up its AI chip photonics work, one sign of how central optical interconnects have become to the next generation of AI hardware.