Why AI Data Centers Are Driving an InP Wafer Price Squeeze

Why AI infrastructure needs more InP

AI data centres move enormous volumes of information between processors, memory, switches and storage. As electrical interconnects encounter speed, distance and power constraints, optical links become increasingly important for moving data across racks and between systems.

InP-based lasers are part of the component chain that makes those links possible. Consequently, the growth of AI clusters can increase demand not only for GPUs and networking equipment, but also for the substrates, epitaxial wafers, lasers and transceivers that connect them. Reuters reported that China’s restrictions on InP exports have become a major hurdle for companies developing fast and energy-efficient optical components for AI data centres.

This creates a form of indirect demand: an AI server order does not consume an InP wafer by itself, but the optical modules required to connect that server can. If data-centre deployment accelerates faster than optical-component manufacturing, substrate availability becomes a limiting factor further down the chain.

Capacity is concentrated—and slow to replace

InP supply is difficult to expand quickly because producers must grow high-quality crystals, process them into wafers and qualify the material for demanding optical-device applications. New facilities also need equipment, process validation and customer qualification before their output can reliably replace established supply.

That makes the market less responsive than a standard commodity market. A price increase can encourage investment, but it cannot create qualified wafers immediately. China’s export-licensing restrictions add another layer of risk by making supply outside China less predictable. Reuters reported that the average price of a six-inch InP wafer had risen 250% to $5,000 since the restrictions were introduced.

The combination of concentrated production, long qualification cycles and export controls gives existing suppliers unusual leverage. It also means that a shortage can affect optical-device makers even when the underlying material represents only one part of the total cost of an AI networking system.

Another increase could be coming

Industry reports say InP substrate prices had already increased three times since the October–December 2025 quarter, with suppliers considering a further increase of more than 10% for the October–December 2026 quarter. Epitaxial wafers, which add a semiconductor layer to the substrate, have also reportedly seen multiple increases.

A fourth increase would matter because it would indicate persistence rather than a one-off repricing. The reported sequence suggests that demand is continuing to outpace available qualified capacity, while the size of the possible next increase points to a market that is becoming more difficult for buyers to secure.

It does not guarantee that prices will rise indefinitely. AI infrastructure orders could slow, customers could redesign systems, or additional suppliers could bring capacity online. But until one of those factors materially changes the balance, the near-term pressure remains upward and volatile.

JX Advanced Metals is responding—but not instantly

JX Advanced Metals has announced up to ¥120 billion of capital investment over four years to strengthen its InP-substrate production capacity for optical communications. The company plans to add capacity beyond its existing Isohara Works, including in the Hitachinaka area.

Industry reporting describes the broader programme as targeting production capacity of roughly seven to ten times its fiscal 2025 level by fiscal 2030. JX is also discussing price revisions with customers as it works to establish what it calls a more stable supply framework.

That response is important for two reasons. First, it confirms that suppliers see data-centre optical communications as a large enough opportunity to justify major investment. Second, the timeline shows why new capacity is unlikely to eliminate the immediate squeeze: the expansion is measured over several years, while AI data-centre orders are being placed now.

Reports have also raised the possibility that JX could seek prices two to three times higher as customer inquiries exceed planned expansion. That specific multiple should not be presented as an announced company-wide price. The official JX release supports the investment plan, while the supplied industry reporting supports discussions about price revisions—not a confirmed two- to threefold list-price increase.

What the squeeze means for AI data centres

The immediate risk is not necessarily that AI data centres will stop operating. It is that the rollout of dense, high-speed optical networking could become more expensive, slower to schedule or more dependent on a small group of qualified suppliers.

For hyperscalers and equipment makers, the pressure may appear through several channels:

• higher costs for InP substrates and epitaxial wafers;
• longer lead times for lasers and optical transceivers;
• greater exposure to export licensing and regional supply disruptions; and
• stronger negotiating power for substrate and optical-component suppliers.

China’s controls intensify the concentration risk, while planned investments in Japan and elsewhere point toward a gradual diversification of supply.

The likely path from shortage to expansion

High prices create an incentive for suppliers to expand, for customers to secure longer-term agreements and for the industry to seek alternative materials or more efficient optical designs. However, those responses will take time, and not every proposed project will produce qualified commercial output on schedule.

The most defensible conclusion is therefore narrower than a prediction of permanently soaring prices: InP has become a meaningful near-term supply risk for AI optical networking. Reported wafer increases of roughly 76% to 78%, possible further increases above 10%, export restrictions and multibillion-yen capacity plans all point to a market under stress.

Whether the pressure becomes a lasting structural feature will depend on two variables: how long AI data-centre construction remains strong and how quickly new, qualified InP capacity reaches customers. Until those two forces come back into balance, the material beneath the laser is likely to remain one of the least visible—and most consequential—constraints on AI networking growth.