Silicon photonics

Silicon photonics builds optical components - waveguides, modulators, detectors - on a silicon chip so data travels as light instead of electrical signals over copper. It uses standard chip factories to make cheap, fast, low-power optics, and has become critical for connecting GPUs inside AI data centers.

What silicon photonics is, in plain terms

Silicon photonics is a way to build optical components - the parts that send and receive **light** instead of electricity - directly on a silicon chip, using the same factories that make ordinary computer chips. The idea is to take the bulky, expensive glass-and-laser optics used in fiber networks and shrink them onto a chip that can be mass-produced cheaply. The signal that travels is light (photons) rather than electrical current (electrons over copper wire). Because light can carry enormous amounts of data over distance with very little loss and heat, packing optics onto silicon lets engineers move far more information, far more efficiently, than copper allows.

How it works: the key building blocks

A silicon photonic chip routes light through tiny channels called **waveguides** - silicon (or silicon-nitride) wires that confine and steer light, much as a copper trace steers electricity. To put data onto the light, a **modulator** rapidly switches the beam on and off (or shifts its phase) by changing the electrical charge in the silicon, encoding bits. At the far end, a **photodetector** - usually made of germanium grown on the silicon, because pure silicon can't 'see' the infrared wavelengths used in telecom - converts the light back into an electrical signal. The one thing silicon is bad at is *making* light: it doesn't emit efficiently, so a separate **laser** (typically indium phosphide) is attached or coupled in to supply the light. This 'electronics for logic, photons for transport, an external laser for the source' division of labor is the defining trait of the technology.

Why it matters for AI and data centers

Training and running large AI models requires tens of thousands of GPUs working as one machine, and they must exchange staggering amounts of data. Traditional copper links run out of reach and burn too much power at the speeds AI now needs, while pluggable optical modules add cost and energy at the edge of the switch. Silicon photonics solves this by integrating optics tightly with the switching and compute chips. The leading edge is **co-packaged optics (CPO)**, where optical engines sit right next to the switch ASIC instead of plugging into the front panel. NVIDIA says its Spectrum-X and Quantum-X photonic switches use roughly 4x fewer lasers and deliver about 3.5x better power efficiency than conventional pluggable optics - a decisive advantage when a single AI 'factory' can draw tens of megawatts.

Where it sits in the supply chain

Silicon photonics is a layered ecosystem. At the bottom, **specialty wafer** makers supply the silicon-on-insulator (SOI) substrate that gives waveguides their low-loss optical confinement - Soitec is the dominant supplier of photonics-grade SOI. Next, **foundries** fabricate the photonic integrated circuits (PICs); Tower Semiconductor and GlobalFoundries (with its GF Fotonix platform) are leading merchant foundries, and TSMC supplies the advanced packaging that stacks an electronic die on a photonic die for CPO. **Fabless designers and packagers** then turn PICs into optical engines, transceivers, and interposers - POET Technologies, for example, sells an 'Optical Interposer' that fuses lasers, PICs and electronics at wafer scale. Finally, **system vendors** like NVIDIA, Broadcom, Cisco and the big transceiver makers integrate it all into switches and modules.

Who the key players are

On the **systems** side, NVIDIA and Broadcom are driving CPO into AI networking - Broadcom's Bailly integrates 6.4 Tbps optical engines into the switch package, while NVIDIA announced the first 1.6T-per-engine co-packaged switches in 2025. On the **foundry** side, Tower Semiconductor (TSEM) has turned its 300mm silicon-photonics process into a fast-growing business, while GlobalFoundries (GFS) bolstered its position by acquiring Singapore's Advanced Micro Foundry in late 2025. Soitec (SOI on some trackers) supplies the underlying photonics-SOI wafers, and POET Technologies (POET) is a smaller fabless player commercializing hybrid-integrated 1.6T transmitter and receiver products with partners like Semtech. Note these are illustrative of the supply chain, not investment recommendations.

What's changing now

The industry is racing up the speed curve: 100G-per-lane optics are mainstream, 200G-per-lane links are arriving in 2026, and 800G and 1.6T modules are ramping. The bigger shift is the move from *pluggable* optics to *co-packaged* optics and even optical I/O built into the compute package itself - many analysts expect AI-cluster interconnects to become predominantly optical within a few years. Open questions remain around reliability, serviceability (a failed co-packaged laser is harder to swap than a pluggable module), and whether external-laser sourcing can scale, but the direction - more light, less copper - is now widely accepted as the path to scaling AI infrastructure.

Frequently asked

What is silicon photonics in simple terms?

It's the practice of building optical parts - the components that transmit and receive light - on a silicon chip using normal semiconductor factories. That lets data move as pulses of light instead of electrical signals over copper, which is faster and more energy-efficient over distance.

Why can't a silicon chip make its own laser?

Silicon is an 'indirect bandgap' material, so it doesn't emit light efficiently. Engineers can make waveguides, modulators and detectors in silicon, but the laser light source almost always comes from a separate material like indium phosphide that is bonded or coupled onto the chip.

How is silicon photonics different from co-packaged optics (CPO)?

Silicon photonics is the underlying chip technology. Co-packaged optics is a packaging approach that uses silicon photonics: it places the optical engines right beside the switch or processor chip instead of in pluggable front-panel modules, cutting power use and latency.

Why does AI need silicon photonics?

AI clusters link tens of thousands of GPUs that must exchange huge volumes of data. Copper can't deliver the needed bandwidth at acceptable power and distance, so optical interconnects built with silicon photonics are becoming essential to scale AI data centers.

Which companies are involved in silicon photonics?

It's a layered supply chain: Soitec supplies SOI wafers; Tower Semiconductor, GlobalFoundries and TSMC handle fabrication and packaging; POET Technologies and others build optical engines and interposers; and NVIDIA, Broadcom and Cisco integrate it into AI networking switches.

Related companies