The AI photonics supply chain

The AI photonics supply chain is the chain of companies that turn light into a data-mover for AI data centers. It runs from compound-semiconductor materials (indium phosphide, gallium arsenide) and lasers, through silicon-photonics foundries, to optical transceivers and co-packaged optics that link thousands of GPUs faster and with less power than copper.

What it is, in plain terms

Photonics means using light, instead of electrical signals over copper, to carry data. The **AI photonics supply chain** is the set of companies that build the materials, chips, lasers and modules needed to move data with light inside and between AI data centers. Why light? Training and running large AI models requires thousands of GPUs to exchange enormous amounts of data at once. Copper wires lose signal quickly at high speeds and over distance, and they burn a lot of power. Light traveling through optical fiber carries far more data, much further, with lower energy per bit. As one industry analysis put it, essentially all AI data-center interconnects are expected to be optical within about five years.

How it works: turning electricity into light and back

At each end of an optical link sits a **transceiver** (transmit + receive). On the transmit side, a tiny **laser** generates light; a **modulator** switches that light on and off (or shapes it) billions of times per second to encode the bits; the light then travels down a hair-thin glass **optical fiber**. On the receive side, a **photodetector** turns the light back into an electrical signal, and a **DSP** (digital signal processor) cleans it up. Modern modules run at **800 gigabits per second (800G)**, and **1.6 terabit (1.6T)** parts are now entering volume production. Each module typically packs eight parallel lanes at 200G each. The whole field of building these light-handling components onto silicon chips is called **silicon photonics**.

Why it matters for AI and data centers

AI clusters are network-bound: a GPU waiting for data is a GPU wasted, and the network can consume a large share of total power. Photonics attacks both problems. The newest move is **co-packaged optics (CPO)** — putting the optical engine on the *same package* as the switch or accelerator chip, instead of in a pluggable module at the faceplate. This shortens the noisy electrical path. At GTC 2025, Nvidia announced its **Spectrum-X** and **Quantum-X** silicon-photonics switches, with Spectrum-X Ethercrnet Photonics reaching up to 409.6 Tb/s of switch bandwidth and shipping in 2026. Nvidia cites roughly **3.5x better power efficiency** and **10x better resilience** versus traditional pluggable optics — the difference between a cluster that scales to millions of GPUs and one that stalls.

Where it sits in the chip and photonics stack

The supply chain is a vertical stack, roughly: 1. **Substrate materials** — wafers of **indium phosphide (InP)** and **gallium arsenide (GaAs)**, the compound semiconductors needed for high-speed lasers and detectors. InP is a known bottleneck; global laser-fab capacity is limited. 2. **Epitaxy** — growing the active crystal layers on those wafers (epiwafers) using **MOCVD** tools. 3. **Silicon-photonics foundries** — fabricating photonic integrated circuits (PICs) on silicon wafers. 4. **Lasers and active components** — the light sources and modulators. 5. **Transceivers and CPO engines** — assembling lasers, PICs, DSPs and optics into finished modules. 6. **Fiber and connectors** — the glass that carries the light between racks and buildings. A recurring theme: **vertical integration**. The strongest players own several layers, from materials through subsystems.

Who the key players are

Mapping real companies to the stack: - **Substrates:** AXT (InP/GaAs/Ge wafers), Wolfspeed (silicon carbide), Soitec (engineered substrates). - **Epiwafers and epitaxy tools:** IQE (epiwafers); Veeco and Aixtron (MOCVD growth equipment). - **Silicon-photonics foundries:** GlobalFoundries (GF Fotonix), Tower Semiconductor (now in volume production of 1.6T silicon-photonics products), TSMC (PIC/EIC/3D packaging), X-FAB. - **Lasers, modulators, DSPs and components:** Coherent and Lumentum (InP lasers, vertically integrated from materials to modules), MACOM, Marvell (PAM4 DSPs, electro-optics; acquired Celestial AI for optical interconnect), Credo (active electrical cables and DSPs). - **Transceivers / CPO assembly:** Coherent, Lumentum, Applied Optoelectronics (scaling 800G/1.6T). - **Fiber:** Corning. Note: many of the most important suppliers, including Nvidia, Broadcom and TSMC, are large diversified companies for which photonics is one segment among many.

What's changing now

Three shifts stand out in 2025-2026. **1. Hyperscaler money is reshaping the supply base.** Nvidia made direct investments into photonics suppliers — reported around \$2B each into Coherent, Lumentum and Marvell, plus a multi-year deal with Corning — to lock in laser, optics and fiber capacity for its switches. This signals that photonics is viewed as AI's next bottleneck. **2. The shift from pluggables to co-packaged optics** is moving from research to product, pulling more value toward foundries and integrators. **3. Geopolitics and materials scarcity.** China added **indium phosphide to its export-control list in February 2025**, tightening a key raw input. Suppliers like AXT (with a vertically integrated China campus) saw InP revenue surge — more than 250% sequentially in one 2025 quarter — while navigating export-license delays. The U.S. CHIPS Act is funding domestic InP capacity expansion.

Frequently asked

Why does AI need photonics instead of copper wires?

Copper loses signal quality rapidly at very high speeds and over distance, and it consumes a lot of power. AI clusters need thousands of GPUs to swap huge amounts of data simultaneously, so links must be fast, long-reach and power-efficient. Light through optical fiber delivers far higher bandwidth at lower energy per bit, which is why the industry expects nearly all AI interconnects to become optical.

What is co-packaged optics (CPO)?

CPO places the optical engine (lasers, modulators, detectors) on the same chip package as the switch or accelerator ASIC, rather than in a pluggable module at the front panel. This shortens the electrical path, cutting power and latency while improving reliability. Nvidia claims roughly 3.5x better power efficiency and 10x better resilience versus traditional pluggable optics.

What is indium phosphide and why is it a bottleneck?

Indium phosphide (InP) is a compound semiconductor used to make the high-speed lasers and detectors in optical transceivers. Unlike plain silicon, it can both efficiently generate light and operate at the very high frequencies AI links need. Global InP laser-fab capacity is limited, and China added InP to its export-control list in February 2025, making supply a strategic chokepoint.

Is silicon photonics the same as a normal computer chip?

No. A conventional chip moves electrons through transistors. A silicon photonic integrated circuit (PIC) guides light through waveguides etched in silicon and manipulates it with modulators and detectors. Because silicon does not emit light well on its own, PICs are usually paired with separate indium-phosphide lasers.

Which companies make up the AI photonics supply chain?

It spans materials makers (e.g. AXT for InP wafers, Corning for fiber), epiwafer and tool suppliers (IQE, Veeco, Aixtron), silicon-photonics foundries (GlobalFoundries, Tower Semiconductor, TSMC), and laser/transceiver integrators (Coherent, Lumentum, Marvell, MACOM, Applied Optoelectronics, Credo). Many leaders are vertically integrated across several of these layers.

Related companies