China’s indium phosphide export controls squeeze global AI infrastructure

China’s restrictions on indium phosphide have emerged as a fresh choke point for AI buildouts, hitting the part of the stack that moves data between chips rather than the chips themselves. Barely a week after Coherent warned of a shortage on its May 6 earnings call, chief executive Jim Anderson was on a plane with a U.S. business delegation accompanying President Donald Trump on a trip to China.

The problem starts with a material that rarely draws attention outside photonics circles. Indium phosphide is used in high-speed optical communications and photonic components, including the lasers and related parts that power optical interconnects inside data centers. As AI clusters grow larger, those links have become just as important as processors and power. If the supply of indium phosphide tightens, networking gear can slip, even when GPUs themselves are available.

The pressure is already showing up in company forecasts. AXT Inc., the Fremont, California-based compound-semiconductor wafer maker with manufacturing in China, said fewer export-control permits from China’s Ministry of Commerce forced it to cut its fourth-quarter 2025 revenue outlook to $22.5 million to $23.5 million. Later, Beijing Tongmei Xtal Technology said it had received export permits in August 2025 to resume shipping indium phosphide substrates to certain additional customers, a sign that approvals were possible but selective and slow-moving.

Those curbs were part of a broader Chinese export-control move that took effect on February 4, 2025, covering items related to tungsten, tellurium, bismuth, molybdenum and indium. The indium piece has since become especially important because photonics supply chains are international by design, tying Chinese materials producers to U.S. and global equipment makers, substrate suppliers and cloud customers. That makes the bottleneck hard to isolate and harder to replace quickly.

Coherent’s scale shows why the issue matters. On its fiscal third-quarter 2026 call, the company reported revenue of $1.8 billion and said it was continuing a 6-inch indium phosphide ramp. It expected to more than double internal capacity by the next quarter, then more than double again by the end of 2027. That kind of expansion underlines both the strength of demand and the strain in the supply chain.

For hyperscalers and infrastructure providers racing to expand AI capacity, the stakes are clear: delays in optical interconnects can slow the deployment of new clusters and push up costs across the next wave of data centers. The contest over AI infrastructure is no longer just about chips and electricity. It is increasingly about the materials, permits and photonics components that let those systems talk to each other fast enough to work.