Optical transceiver

An optical transceiver is a small plug-in module that converts electrical data into laser light to send over fiber, and converts incoming light back into electrical signals. It's the workhorse interconnect that links servers, switches and GPUs in data centers, and demand has exploded with AI, driving the shift to 800G and 1.6T modules.

What an optical transceiver is

An **optical transceiver** is a small, hot-swappable module that plugs into a switch, router, or server network port and does two jobs at once: it *transmits* data by turning electrical signals into pulses of laser light, and it *receives* data by turning incoming light back into electrical signals. "Transceiver" is just *transmitter* + *receiver* in one package. It connects to the outside world through optical fiber. Because light travels through glass with far less loss than electricity travels through copper, transceivers let data move faster and much farther — from a few meters between racks to tens of kilometers between buildings — than copper cabling ever could. Common form factors today are pluggable modules like QSFP-DD and OSFP that slot into a switch faceplate.

How it works

On the transmit side, an electrical signal drives a tiny **laser** (or modulator) that flickers on and off — or shifts amplitude/phase — billions of times per second, encoding bits as light. The most common laser type for high-speed data-center links is the **EML (electro-absorption modulated laser)**, prized for a clean, sharp signal over distance. The light is launched into a fiber. On the receive side, a **photodetector** turns the incoming light back into a weak electrical current, which a **trans-impedance amplifier (TIA)** boosts back to usable levels. Many high-speed modules also include a **DSP (digital signal processor)** chip that cleans up and reshapes the signal, plus retiming and error-correction. Modern 800G modules typically run eight lanes of 100G each, multiplexed onto the fiber.

Why it matters for AI and data centers

AI training and inference clusters tie together thousands of GPUs that must constantly exchange enormous amounts of data. The network *between* chips has become a primary bottleneck — Nvidia and hyperscalers have called optical interconnect the "final puzzle" of AI infrastructure. Every GPU-to-switch and switch-to-switch hop typically rides on an optical transceiver. The result is a demand explosion. The datacom optical market grew over 60% to top **$16 billion in 2025**, and **800G** modules (QSFP-DD800 / OSFP) are now the default for new AI buildouts, with over 24 million 800G-plus units shipped in 2025. **1.6T** modules are now entering production. Supply is tight: analysts expect 800G demand to outrun supply by 40-60% through 2027, largely due to limited EML laser manufacturing capacity.

Where it sits in the photonics supply chain

A transceiver is an assembly of specialized parts, so value is spread across several tiers. At the bottom are **components**: lasers (EMLs, VCSELs, DFB lasers), modulators, photodetectors, and the silicon **DSP** chip. These feed into **optical engines / sub-assemblies**, which are then built into the finished **pluggable module**. A growing alternative is **silicon photonics**, which fabricates optical functions directly on a chip rather than assembling discrete parts — improving cost and density at scale. Beyond components, **contract manufacturing and advanced packaging** (precision assembly and test of these delicate optical parts) is its own critical tier, dominated by a few specialists.

Who the key players are

Module market share is led by China's **Innolight (Zhongji Innolight)**, the largest 400G/800G supplier to US hyperscalers, alongside US-listed **Coherent** (`COHR`), a leader in lasers and modules. **Lumentum** (`LITE`) is a major supplier of lasers and, increasingly, transceivers; both Coherent and Lumentum received roughly $2 billion equity investments from Nvidia, anchoring them as core AI-cycle partners. **Applied Optoelectronics** (`AAOI`) makes lasers and transceivers and has won diversification business (including an Amazon relationship), swinging back to strong growth on 800G AI demand. **Fabrinet** (`FN`) is the key contract manufacturer that assembles and tests optical products for many of these names, including Nvidia-linked supply. Cisco and Huawei are also large players.

What's changing now

The frontier is reducing **power and latency** as speeds climb. **LPO (linear pluggable optics)** strips out the power-hungry DSP, cutting module power and latency for short links. The bigger structural shift is **CPO (co-packaged optics)**, which moves the optics out of the faceplate and places optical engines right next to the switch or GPU ASIC, dramatically improving power efficiency. Nvidia announced 1.6T co-packaged optics switches (Quantum-X / Spectrum-X Photonics) using silicon-photonics micro-ring modulators in 2025, and Broadcom is pushing CPO too. TrendForce projects the CPO/NPO market could exceed **$39 billion by 2030**, with silicon photonics' share of the optical market roughly doubling — though pluggable transceivers will remain the volume backbone for years.

Frequently asked

What is the difference between an optical transceiver and a fiber cable?

The fiber cable is just the passive glass strand that carries light. The transceiver is the active electronic module at each end that converts electrical data into light to send into the fiber, and converts received light back into electrical signals.

What does 800G or 1.6T mean on a transceiver?

It's the total data throughput: 800G means 800 gigabits per second, 1.6T means 1.6 terabits per second. Higher numbers are usually built from multiple parallel lanes (e.g., 800G as eight 100G lanes) and are now standard for AI data-center networking.

Why are optical transceivers so important for AI?

AI clusters connect thousands of GPUs that must constantly share data. The network linking them has become a bottleneck, and almost every high-speed link runs on an optical transceiver — so AI buildouts drive massive transceiver demand.

What is co-packaged optics (CPO) and will it replace transceivers?

CPO moves the optical engine next to the switch or GPU chip instead of in a pluggable faceplate module, saving power. It's emerging for the highest-end systems, but standard pluggable transceivers will remain the high-volume mainstream for years.

Which companies make optical transceivers?

Major suppliers include Innolight, Coherent (COHR), Lumentum (LITE), Cisco, and Huawei, with Applied Optoelectronics (AAOI) growing fast. Fabrinet (FN) is a leading contract manufacturer that assembles many of these products.

Related companies