What Is Silicon Photonics? How Chips Send Data With Light

15 min read · Updated: October 23, 2026

Silicon photonics builds parts that handle light on the same kind of silicon wafer used for computer chips. These include paths for light, parts that make light blink, and parts that turn light back into electricity. Data centers use it to carry data as light over distances where copper wires would lose too much signal and power. Only the laser, which makes the light, comes from another material, because silicon is bad at giving off light.

What you'll learn

What you gain by carrying data with light instead of electricity
How the light path (waveguide), the modulator and the photodetector work
Why the laser is supplied separately
How it is used in data center optical transceivers and in CPO, and what problems remain

Prerequisites: it is enough to know that optical fiber carries communications. To learn how lasers work, see "What Is a Laser Diode?"

Where you see it

When you watch a video or ask an AI chatbot a question, your data passes through a data center. A data center is a building filled with rows of servers, the computers that do the work. Servers, and the equipment that sends data on to the right place, are linked mainly by optical fiber.

A small part, small enough to pinch between your fingers, plugs into each end of the fiber. It turns electrical signals into light to send, and turns arriving light back into electricity. This part is called an optical transceiver. "How Is Data Sent Over the Internet?" also touches on how light travels through fiber.

Inside these optical transceivers are silicon photonics chips. For example, Intel began shipping a 100 gigabit per second product using this technology in 2016. According to its product page, Intel has shipped more than 8 million of these chips since 2016 (as of 2025).

Why carry data with light instead of electricity?

An electrical wire sends 0s and 1s by raising and lowering the voltage on a copper line. To send faster, the voltage must switch up and down more quickly. But the faster the change, the more the current crowds near the surface of the wire. More energy is lost as heat. The board material that holds the wire also soaks up more signal energy at faster changes.

So an electrical signal gets weaker and more distorted the faster and farther it goes. To keep it readable at the far end, circuits must boost and reshape it, and those circuits use power. Why the thin wires inside a chip cause delay is explained in "RC Time Constant Explained."

Optical fiber is different. The infrared light used in communication loses only about 4% of its strength over 1 km. Blinking the light faster barely changes how much it weakens in the fiber. To keep things simple, this article leaves out the distortion caused by the spread of colors in the light.

Sending data as light, however, needs a conversion between electricity and light at both ends. The conversion parts cost power and money, so copper is still used over very short distances. NVIDIA's GB200 NVL72, announced in 2024, holds 72 GPUs (chips that do AI calculations) in one rack. About 5,000 copper cables link them together.

An analogy

Choosing between light and electricity is like choosing between a freight train and a truck. To move a lot of cargo far, a freight train wins because it uses less fuel per item. But loading cargo at a station takes effort, so a truck is better for a local delivery. Light is the freight train, the electrical wire is the truck, and loading at the station is the conversion between electricity and light.

Where the analogy breaks down: At a station, cargo waits. The conversion between electricity and light happens while the signal keeps flowing, without stopping it. Also, switching to light does not greatly shorten the time data takes to arrive (see the common misconceptions). This analogy shows only one point: the choice depends on the trade-off between conversion cost and loss over distance.

What silicon photonics is

Silicon photonics builds optical circuits on a silicon wafer, the round disc that chips are made on. It uses steps very similar to ordinary chipmaking. For example, the circuit pattern is printed with light, and unwanted parts are etched away ("How Are Chips Made?"). So tiny parts can be mass-produced with the same exact shape.

Silicon works for optical circuits because it lets the infrared light used in communication pass well. A material absorbs light when each particle of light carries more energy than a certain step. That step is the energy needed to lift an electron up, and it is called the band gap. In silicon, light with a wavelength longer than about 1,100 nm does not have enough energy for this step. It passes through almost without being absorbed. To the 1310 nm and 1550 nm light used in communication, silicon is like a clear sheet of glass.

A chip holds three main kinds of parts: paths for light, parts that make light blink, and parts that turn light back into electricity. The figure below shows the flow from the sending side to the receiving side.

The thick line is the light path. On the sending side, the modulator makes the laser light blink to carry 0s and 1s. On the receiving side, the Ge (germanium) photodetector turns the light back into current. Left out: real chips usually have both sending and receiving parts on one chip, with many sets side by side. The parts that connect the chip to the fiber, and the electrical circuits that clean up the signal, are also left out.

Part 1: the light path, or "waveguide"

In optical fiber, light travels through a thin glass core. On a chip, a waveguide does the same job. In silicon photonics, the core is a thin line of silicon. Around it is silicon oxide, the same substance glass is made of.

Light travels more slowly inside a material than in a vacuum. The number that tells how much slower is the refractive index. The larger it is, the slower light goes. For the light used in communication, silicon is about 3.5 and silicon oxide is about 1.45.

When light in a slow material hits a faster material at a shallow angle, it all bounces back at the boundary. This is called total internal reflection. Light in the silicon core keeps bouncing off the boundary with the silicon oxide around it. So it moves forward without escaping.

The bigger the difference in refractive index, the more tightly light can be held in a small core. A common core is about 0.5 µm wide and 0.22 µm thick. A human hair is about 0.05 to 0.1 mm thick, so the core is less than a hundredth of that. Light barely leaks even at sharp bends, so parts can be packed into a small area.

The top drawing is a cross-section, cut across the direction light travels. The small gray rectangle is the silicon core, and the colored oval shows where the light is. The light spreads a little outside the core. The bottom drawing looks down on the core and draws light as a line (a ray). Left out: real light travels as a wave spread across the whole core, not in a sharp zigzag as drawn. The oxide layer under the core and the silicon base below it are also left out.

Part 2: making light blink, the "modulator"

The laser stays on, and a separate part makes the light stronger or weaker. This part is called a modulator. In silicon, a voltage adds or removes electrons and holes (empty seats for electrons) in the waveguide. This changes the refractive index slightly. The light wave then runs a little behind or a little ahead.

A small shift in the wave alone does not change how bright the light is. So a Mach-Zehnder modulator splits the light into two paths. It shifts the wave in only one path, then joins the two again. If peaks meet peaks, the light is bright. If peaks meet troughs, they cancel out and the light is dark. The change in silicon's refractive index is small, so the paths are a few mm long.

The top drawing is a view from above. Incoming light splits into two paths, and a voltage shifts the wave in the upper path only before the paths rejoin. Below are the two waves that rejoin (the upper two) and their sum (below the line). Compare them at the dotted line: on the left, peak meets peak; on the right, peak meets trough. On the right, the waves are shifted by half a wavelength and cancel out. Left out: real devices often have electrodes on both paths, and brightness changes smoothly between bright and dark. When the output is dark, the light does not vanish. It escapes at the joining point or leaves through a second output.

Another method uses a ring resonator. A tiny loop of waveguide, 10 to a few tens of µm across, sits right next to the main waveguide. Here, light of different wavelengths is called "light of different colors" (it is invisible infrared). Only a color whose wavelength fits a whole number of times into one trip around the loop enters it and circles round and round. In 2005, a team at Cornell University in the US made light blink with a 12 µm loop.

A voltage slightly changes the refractive index of the loop, which shifts the color that enters it. Light of a given color then switches between "passing straight by" and "entering the loop and getting weaker." Rings can be made far smaller than Mach-Zehnder modulators, by orders of magnitude. Their weakness is that they are sensitive to temperature (explained in the challenges section).

Part 3: turning light back into electricity, the "Ge photodetector"

On the receiving side, light is turned back into electricity. This is done by a photodiode, a part that lets current flow when light hits it. When a semiconductor absorbs a particle of light, it creates a pair of an electron and a hole. That pair becomes current.

But silicon hardly absorbs 1310 nm or 1550 nm light. The property that made it a good light path now works against it. So germanium (Ge) is grown as a crystal on top of the waveguide. Germanium is in group 14 of the periodic table, the same group as silicon. Its band gap is smaller, so it can absorb this light and turn it into current.

A light receiver made this way is called a Ge photodetector. Germanium can be grown on silicon, and its steps fit easily into a silicon factory's process. So the receiving parts can also sit on the same chip.

The laser is supplied separately

Only the light source cannot be made from silicon. Silicon is an indirect band gap semiconductor. When an electron meets a hole in silicon, most of the energy turns into heat instead of light. For details, see "What Is a Laser Diode?" and "direct vs. indirect band gap."

So lasers made of materials that are good at giving off light, such as InP, are added. InP (indium phosphide) is a semiconductor made of two elements, a type called a compound semiconductor. There are two main approaches. One bonds a small piece of InP onto the silicon chip. The other makes the light in a separate part and feeds it into the chip through fiber. The second approach keeps the laser away from hot areas and makes it easier to replace.

Use 1: optical transceivers in data centers

The biggest use is optical transceivers in data centers. Ethernet is a common way of connecting network equipment, and its speed standards are set by the IEEE, an association of engineers. The 400 gigabit per second (400G) standard was approved in December 2017, and the 800G standard in February 2024.

For example, the 400G type called "DR4" sends 100 gigabits per second of light down each of 4 fibers, reaching up to 500 m. The 800G "DR8" type does the same over 8 fibers. Many identical sets of parts must be lined up precisely on one chip. Making tiny, uniform parts in rows is what chip factories do best, so silicon photonics suits this job.

Use 2: CPO and AI data centers

Most optical transceivers today plug into ports on the front of a switch, the equipment that directs data. Between the large chip at the center and those ports, signals travel over electrical wiring on the board. As the speed on each wire rises, loss and power on this stretch have become a problem.

So a newer approach puts the optical parts in the same package as the switch chip. A package is the holder that contains a chip and connects it to the board. This approach is called CPO (co-packaged optics) and is starting to reach real use. The electrical wiring becomes very short, which saves power. Packages are explained in "What Is Semiconductor Packaging?"

In March 2024, Broadcom announced it had delivered CPO switches to customers. They pair a 51.2 terabit per second switch chip with 8 silicon photonics optical parts. The company says the optical part uses 70% less power than pluggable modules. NVIDIA also announced CPO switches in March 2025, saying power efficiency improves 3.5 times. At the time, it planned to release them in stages through 2025 and 2026. Both numbers are each company's own figures, using its own way of comparing.

AI training links a very large number of GPUs so they work like one big computer. NVIDIA says it aims to use CPO to connect millions of GPUs. The more links there are, the more a tiny power difference per link adds up. This trend of bringing optical and electronic circuits closer is also called photonics-electronics convergence, a term widely used in Japan.

Remaining challenges

Connecting to optical fiber

In a standard optical fiber, the area the light travels through is about 9 µm across. A chip's waveguide is only about 0.5 µm wide, so the sizes differ by about 20 times. If you simply butt them together, much of the light leaks away.

One method gradually narrows the tip of the waveguide so the light spreads out. The tip then meets the fiber at the edge of the chip. Another method cuts fine grooves into the waveguide to bend the light upward and out of the chip. Either way, the fiber's position must be aligned with micrometer-level precision. That adds assembly work and cost.

Sensitive to temperature

Silicon's refractive index changes with temperature. So the color a ring resonator pulls in shifts by about 0.08 nm for every 1 °C rise. At 10 °C, that is about 0.8 nm. WDM (wavelength division multiplexing) sends many colors of light through one fiber, and some systems space the colors only about 0.8 nm apart. So a 10 °C change can move a ring onto its neighbor's color. Rings therefore get tiny heaters, plus circuits that watch the temperature and adjust. Those also use power.

Assembly and replacement

A pluggable optical transceiver can be pulled out and replaced on its own if it fails. With CPO, the optical parts sit inside the package, so this is not possible. Testing before assembly becomes important. Some designs move the heat-sensitive laser outside the package so it can be replaced. In its 2025 announcement, NVIDIA said it had cut the number of lasers to a quarter of earlier designs.

Common misconceptions

Myth: All the wiring inside a chip will be replaced by light

In fact, copper wires still connect the transistors inside a chip. Optical parts are far bigger than transistors: even a waveguide is about 0.5 µm wide, and a modulator is a few mm long. Over short distances, the conversion can also use more power than it saves. Light handles the longer distances out of the chip or the package. The phrase "chips wired with light" makes people think light goes all the way inside.

Myth: Using light makes the computing itself faster

In fact, silicon photonics is a technology for moving data, and transistor circuits still do the computing. An optical signal travels through a silicon waveguide at about a quarter of the speed of light in a vacuum. Being light does not make it arrive dramatically faster. The gain is sending more data, farther, for the same power. Some research uses light itself to compute, but that is a different technology.

Myth: Data center optical links have already switched to CPO

In fact, CPO products only began to appear around 2024. Pluggable optical transceivers are still widely used. People are watching how far CPO spreads as it tackles problems such as hard replacement and assembly cost. New products get big news coverage, so CPO can look like it is already mainstream.

By the numbers

Item	Number	Note
Waveguide core	About 0.5 µm wide, 0.22 µm thick	Under a hundredth of a hair
Refractive index	Silicon about 3.5, silicon oxide about 1.45	A bigger gap holds light more tightly
Optical fiber core	About 9 µm across	About 20 times the waveguide
Ring resonator	10 to a few tens of µm across	The 2005 example was 12 µm
Ring color shift	About 0.08 nm per 1 °C	Tuned with a heater
400G and 800G standards	2017 and 2024	Approved by the IEEE
CPO switch	51.2 terabits per second	Announced in March 2024

51.2 terabits per second equals 64 ports of 800G. Each optical part handles an eighth of that, 6.4 terabits per second.

Learn more

The light source: "What Is a Laser Diode?" (including why 1310 nm and 1550 nm are used)
Compare in a table: single-mode vs. multimode fiber, NRZ vs. PAM4 (how many brightness levels to use)
Terms for the next step: plasma dispersion effect (changing the refractive index with the number of electrons), grating coupler and edge coupler (ways to connect to fiber), propagation loss (how much light weakens in a waveguide)
See it in figures: more detailed figures of the Mach-Zehnder modulator, the ring resonator and the silicon waveguide cross-section are in the list of figures
Technologies that combine chips: "What Is a Chiplet?" and "What Is HBM?"
The big picture of where chips go next: "Beyond Scaling"