AI Optics Demand Forces Soitec Into Multi-Year Wafer Deals

A quiet bottleneck in AI infrastructure is becoming a contract battlefield. French semiconductor-materials maker Soitec is locking customers into multi-year supply agreements for the wafers used in silicon-photonics chips, as hyperscalers scale the optical links that move data inside AI systems.

The details were reported by Reuters, citing Soitec CEO Laurent Rémont. Customers will reserve capacity with deposits and fixed prices. They recover the deposit if they take the agreed volume, but risk forfeiting it if they fall short. The structure gives both sides more visibility in a market where demand is accelerating faster than supply planning.

Why AI data centers need more optics

Large AI clusters increasingly depend on optical connections to move data between accelerators, switches, and storage. Copper remains useful over short distances, but power consumption and signal performance become harder to manage as bandwidth and rack scale rise. Silicon photonics offers a path to higher throughput with better energy efficiency, making its underlying materials strategically important to AI builders.

Soitec supplies the silicon-on-insulator substrates beneath much of this photonics ecosystem. The company told investors that Photonics-SOI revenue should more than double this financial year from slightly above $100 million. Rémont said that more than $200 million is now an absolute floor, according to Reuters.

The infrastructure lesson for builders

This is not only a semiconductor story. It shows how AI deployment depends on layers that are easy to ignore until they become scarce. A model provider can add capacity by buying more GPUs, but the full system also needs networking, packaging, power, cooling, substrates, and qualified suppliers.

Soitec expects to cover this year and next using existing facilities, equipment additions, and production shifts before considering a new fab. A Singapore building could be equipped within six to twelve months if demand requires it. That flexibility may matter as much as raw capacity: AI infrastructure is expanding in bursts, and suppliers that can reserve output without overbuilding gain leverage.

Builder impact: AI teams planning large inference or training deployments should treat optical networking and supply-chain lead times as architecture inputs, not procurement details. Capacity reservations, multi-vendor designs, and realistic expansion windows can protect a project from bottlenecks far below the model layer.