China’s Optical Supply Chain Is Becoming AI Infrastructure’s Critical Link

China’s importance to AI infrastructure is most visible in a component that users rarely see: the optical transceiver. These modules convert electrical signals into light and back again, allowing GPUs, switches, racks and data centers to exchange enormous volumes of data without relying solely on copper connections.

The strategic advantage is not that China controls every part of the optical stack. It is that Chinese manufacturers have built exceptional scale in the final, difficult-to-qualify stage: designing, integrating, packaging, testing and shipping high-speed modules in large volumes. That manufacturing base is now difficult to reproduce quickly elsewhere.

The ranking signal: China moved into the top tier

LightCounting’s 2024 supplier ranking illustrates how quickly the market has changed. Innolight’s transceiver revenue exceeded $3.3 billion, up 114% year over year, while Eoptolink grew 175% to $1.2 billion and rose from seventh place in 2023 to third in 2024. 52

Other reports describe seven Chinese companies among the global top 10, but the precise order varies by reporting period and methodology. The reliable conclusion is broader than whether Eoptolink is second or third: Chinese vendors have become a dominant force in the high-speed Ethernet-transceiver market, while many Japanese and U.S. suppliers have reduced or exited parts of the business. 5051

That shift matters because Ethernet optics are the fastest-growing portion of the transceiver market. LightCounting said Ethernet represented 75% of the broader transceiver market in 2025 and that Innolight and Eoptolink benefited from focusing on large cloud customers. 51

Why AI clusters need so many optical modules

AI training and inference require dense connections among GPUs and switches. As clusters expand, optical links increasingly connect racks, buildings, campuses and separate data centers. Each new generation of networking raises the bandwidth requirement: 400G has given way to 800G, while 1.6T products are moving into volume production and qualification.

A 2026 market estimate projects global shipments of 49 million 800G modules and 22 million 1.6T modules, while still forecasting a multimillion-unit supply shortfall, particularly for 1.6T. 1 Innolight has also reported mass production across successive generations of high-speed modules, including 400G, 800G and 1.6T products. 6

At this scale, manufacturing execution becomes a competitive technology. A supplier must coordinate optical devices, electrical chips, substrates, packaging, alignment, automated testing and customer qualification. A module that works in a laboratory is not enough; hyperscale operators need consistent quality, high yield and predictable delivery across millions of units.

China’s role is concentrated in the manufacturing layer

The supply chain is better understood as a cross-border division of labor than as a purely Chinese system.

• Chinese vendors are particularly strong in module design and integration, packaging, high-volume assembly, testing and delivery.
• DSP and analog-chip suppliers provide the high-speed electronics that condition, recover and process optical signals.
• Specialist photonics companies supply EML lasers and other advanced optical components.
• Asian manufacturing networks contribute components, tooling, contract manufacturing and final-module capacity across mainland China, Taiwan, Thailand and other Southeast Asian locations.

This structure gives Chinese companies a powerful position even when some of the most valuable inputs originate elsewhere. Their advantage comes from assembling the whole production system around fast product transitions and large cloud orders.

Why replacing China is harder than building a new factory

The United States and other countries can add factories, but new floor space alone does not recreate an established optical ecosystem. A qualified 800G or 1.6T transceiver depends on synchronized access to chips, lasers, InP substrates, packaging equipment, optical-alignment tools, testing capacity and engineering expertise.

It also requires customer qualification. Hyperscalers must validate performance, reliability and interoperability before deploying a module throughout an AI network. Years of process learning can therefore matter as much as nominal factory capacity.

Planned expansion by companies such as Coherent, Lumentum, Nokia, Applied Optoelectronics and Corning could improve geographic resilience. It is unlikely, however, to replace Asia-Pacific production capacity immediately. The more realistic outcome is diversification: more manufacturing outside mainland China alongside continued dependence on Chinese companies, engineers and suppliers.

Chinese vendors are already pursuing that strategy. Innolight and Eoptolink have expanded in Thailand, and other suppliers are building regional capacity to serve overseas customers. Such sites can reduce tariff and logistics exposure, but they do not automatically remove dependence on Chinese corporate supply chains or upstream sourcing. 17

The bottleneck is moving upstream

As module assembly expands, shortages are increasingly likely to appear in the inputs that are hardest to scale. Four chokepoints deserve particular attention:

1. EML lasers, used in some high-performance 800G and 1.6T designs.
2. Silicon photonics, which offers a path to greater density and power efficiency but requires specialized fabrication, packaging and laser integration.
3. High-speed DSP and analog chips, supplied by a relatively limited group of advanced vendors.
4. Indium-phosphide substrates, which support many high-speed optical devices and laser technologies.

The InP constraint is especially notable. Industry estimates cited in recent reporting put 2026 demand at roughly 2.6–3 million two-inch-equivalent wafers versus effective capacity of about 750,000 wafers. 34 TrendForce has likewise described InP supply as a critical constraint on AI-data-center optical-interconnect expansion. 33

These numbers should be treated as estimates, not audited global totals. They nevertheless show why adding module assembly capacity may not solve the problem. If qualified substrates or laser capacity remain scarce, the downstream factory cannot run at its advertised output.

The market is expanding beyond pluggable transceivers

AI infrastructure is also broadening the opportunity beyond short-reach pluggable modules. The larger optical stack includes:

• coherent optics for high-capacity connections;
• data-center interconnect, or DCI;
• optical transport networks and AI-oriented OTN;
• switches and optical engines; and
• co-packaged optics, or CPO.

This expansion reflects the geography of AI computing. Training clusters need fast links within a facility, but operators also need high-capacity connections among campuses and regions. The result is a demand cycle that reaches from GPU fabrics to long-haul and metro optical networks.

Ciena’s fiscal second quarter of 2026 provides one demand signal from the transport layer: revenue reached $1.57 billion, up 40% year over year, while its backlog reached $7.7 billion and Optical Networking revenue grew 42%. 31 Those results do not establish that every order is AI-related, but they are consistent with strong reinvestment in cloud and optical infrastructure.

What determines the next winners

The next phase of competition will not be decided solely by which company announces the fastest optical module. The stronger position may belong to suppliers that secure long-term access to the entire production chain: InP substrates, lasers, DSPs, silicon-photonics capacity, packaging lines and qualified testing.

China’s advantage is therefore best described as manufacturing-system scale, not complete technological self-sufficiency. Its companies are deeply embedded in the process that turns advanced optical components into deployable AI-network hardware. At the same time, their position remains exposed to upstream technologies and materials controlled by a broader international ecosystem.

That interdependence is the central investment and infrastructure lesson. AI data centers cannot scale on GPUs alone. They also need a reliable optical supply chain—and the companies that control its scarce upstream inputs and high-yield manufacturing capacity may shape how quickly the next decade of AI infrastructure can be built.