AI’s Next Supply Bottleneck Moves to Optical Interconnects

Soitec’s multi-year photonics-SOI agreements show that AI infrastructure is shifting capacity risk from compute alone toward the wafers, lasers, packaging, and optical engines that move data.

Semiconductor Supply Chain Update | September 2, 2026

The AI infrastructure bottleneck is changing. Compute remains scarce and expensive, but the ability to connect thousands—or eventually millions—of accelerators is becoming equally strategic.

As lane speeds move from 100G toward 200G and 400G, traditional copper links face higher signal-loss, power, and reach constraints. That is pulling optical technology closer to the switch and package through LPO, NPO, CPO, and silicon-photonics architectures.

For procurement teams, this creates a new category of capacity risk. A complete AI deployment can have sufficient accelerator supply and still be delayed by optical wafers, lasers, fiber arrays, packaging, coupling, testing, or network-system capacity.

01 | Capacity Is Being Locked Before the Market Fully Scales

Soitec’s contract structure is one of the strongest signals. The company is asking customers for deposits, fixed-price commitments, and minimum purchase volumes across multi-year agreements.

That structure is designed to prevent overbooking and give the supplier clearer demand visibility before adding tools and reallocating cleanroom capacity. For buyers, it means access to future optical capacity may increasingly depend on early commitment rather than short-term purchase orders.

02 | AI Networking Is Moving From Copper Toward Optics

The technical reason is straightforward: data movement is becoming a larger part of AI-system power and performance. TrendForce says the limits of conventional copper become more visible as lane speeds progress toward 200G and 400G, increasing signal-loss, equalization, and power requirements.

NVIDIA’s Spectrum-X Ethernet Photonics uses co-packaged silicon photonics and is designed for up to 409.6 Tb/s of switch bandwidth. NVIDIA says its CPO-based networking can deliver five times better network power efficiency and longer sustained AI application runtime than traditional pluggable-transceiver architectures.

03 | The Optical Supply Chain Is More Concentrated Than It Looks

The optical stack includes more than transceivers. Silicon-photonics wafers, indium-phosphide lasers, optical engines, fiber-array units, coupling processes, advanced packaging, and reliability testing all contribute to production readiness.

TrendForce has identified silicon-photonics wafers, InP substrates, and fiber-array components as concentrated supply areas. It also identifies fiber alignment, thermal management, and testing cost as key mass-production barriers for CPO.

04 | 2026 Is Moving From Adoption to Volume Production

TrendForce describes 2026 as the transition year for AI optical interconnects from adoption toward mass production. Its estimate for the global AI optical-transceiver market rises from US$16.5 billion in 2025 to US$26 billion in 2026—more than 57% year-over-year growth.

TrendForce also expects 800G-and-above optical modules to represent more than 60% of the global optical-transceiver shipment mix in 2026. Large-scale CPO procurement is expected to accelerate in 2027–2028.

05 | Singapore Becomes Part of the Capacity Question

Soitec is expanding without immediately building a new fab. It is reallocating production and adding tools inside existing cleanroom space, while evaluating whether to equip an unused facility in Singapore within the next six to twelve months.

That matters because the market is prioritizing availability over manufacturing geography. For many AI customers, the immediate question is not where a photonics wafer is made, but whether capacity is qualified, committed, and available on schedule.