AI's Optical Backbone Is Caught in a China Supply Squeeze

Between a GPU and the fiber-optic cable that carries its output to the next rack sits a sliver of crystal most readers will never see. That sliver, a compound called indium phosphide, is the laser substrate inside the optical transceivers that convert electrical signals into light. It is also the part of the AI buildout that just started getting expensive.

Nomura-cited industry data shows two-inch indium phosphide wafers rose to as much as 880 yuan (US$130.54) each in June 2026, up from as little as 500 yuan at the end of 2025. Three-inch wafers climbed to roughly 3,200 yuan from about 1,800 over the same span, per the same analyst read summarized by the South China Morning Post. Wind's InP index, which tracks more than 20 mainland China-listed firms in the material, gained 4.2% in a single session in early August, a move steeper than lithography and broader semiconductor materials segments, the same report noted.

A 76% jump in half a year on the smaller wafer, and a single-session index move that beat the rest of the chip supply chain, would be a footnote if indium phosphide were a generic chemical. It is not. The compound is the substrate the laser diodes inside high-speed optical transceivers are grown on, and those transceivers are now standard issue in the spine of every AI cluster. Without InP wafers, the light does not leave the chip.

Three independent price signals lined up in one quarter: Nomura's wafer print, Wind's index, and a third from Japan. JX Advanced Metals, one of Japan's leading indium phosphide substrate producers, is introducing a first round of price revisions that could raise prices to two to three times previous levels, according to JPMorgan research cited in the same SCMP coverage. JPMorgan separately noted that customer inquiries to JX continue to exceed the company's planned production capacity expansion. Taiwan's Economic Daily News, also cited in the same report, said the industry is mulling further hikes of up to 10% in the fourth quarter of 2026.

An indium phosphide wafer is the platform on which a laser diode is grown. The diode inside a transceiver turns the GPU's electrical signal into a beam of light that travels down a fiber strand. Silicon chips do the computing; indium phosphide does the conversion. The substrate is grown differently, sourced from a different country mix, and priced through a different pipeline.

Lumentum, one of the largest merchant buyers of the substrate, has called the squeeze worse than the memory chip shortage, in remarks carried by Tom's Hardware. The comparison understates how different the categories are: memory is a much larger market with different inventory dynamics, and treating InP as a memory proxy is the kind of compression that gets supply-chain stories wrong. TrendForce separately called the substrate the optical-interconnect bottleneck inside the AI buildout, a category that has expanded with the size of GPU clusters.

AXT, a US-listed substrate maker, posted record second-quarter results on InP demand, per Yahoo Finance, with the company's profile page on Ad-Hoc-News laying out how the substrate became the line item doing the work.

The supply side is where the risk lives. China is a major producer of indium phosphide, anchoring the upstream end of the chain. A US policy that restricted optical transceivers could cut roughly 60% of AI data center supply that relies on Chinese indium, according to TechTimes reporting on a potential FCC move. Western substrate capacity outside China exists, including JX in Japan, but JPMorgan noted that even with planned expansion, JX cannot clear the inquiry backlog.

The falsifier is the JX transmission. If the Japanese revisions do not pass through to non-Chinese substrate buyers and the Wind index move turns out to be one-quarter noise rather than a regime change, the optical-layer story breaks. The market will tell us in Q4: a 10% second hike on top of the first would confirm the story; a flat print would suggest the spike was positioning, not shortage.

For now, the simplest way to read AI infrastructure cost pressure is the GPU, the rack, and the power. The harder, more useful way is to add a fourth: the laser substrate inside the optical transceiver. That sliver of indium phosphide, most of it grown in China, just became the most expensive sentence in the cluster.