"InP Shortage Will Be More Severe Than Memory"! Lumentum and SemiAnalysis Warn: Optical Interconnects Become New AI Bottleneck

Indium phosphide (InP) is becoming the weakest link in the chain of AI infrastructure expansion. As data centers see an order-of-magnitude growth in demand for optical interconnects, this irreplaceable compound semiconductor material is facing a supply gap even more severe than that of memory.

Lumentum CEO Michael Hurlston issued a clear warning at the RAISE Summit held in Paris this month: The shortage of indium phosphide will exceed the crisis currently experienced in the memory market. He stated that Lumentum currently operates five InP wafer fabs, but shipments remain over 30% below customer demand, with this gap widening further compared to the previous quarter. "I believe that even if you combine the two of us, we cannot meet the demand placed on us by Nvidia and other customers," Hurlston said. "The shortage of indium phosphide will be more severe than the situation in the memory market."

Market research firm SemiAnalysis also recently issued a similar warning, pointing out that the supply of continuous-wave distributed feedback (CW DFB) InP lasers for near-package optical interconnects is particularly tight. Industry wafer sizes generally remain at the 2-to-4-inch stage, far behind the 12-inch process of silicon-based CMOS, creating structural constraints on capacity expansion.

These warnings have attracted significant market attention. In March this year, Nvidia invested $2 billion each in Lumentum and Coherent to secure supplies from these two vendors, which together account for the majority of global high-speed data communication laser capacity. Despite this, related stocks have suffered heavy selling pressure in the past month—Lumentum (LITE) fell 22.3%, Coherent (COHR) dropped 37.6%, substrate supplier AXT (AXTI) plunged 39.9%, and transceiver manufacturer Applied Optoelectronics (AAOI) fell 35.7%—with stock price trends clearly diverging from fundamentals.

Silicon Photonics Cannot Bypass InP

The irreplaceability of InP stems from its material physical properties**. Silicon has an indirect bandgap structure, which can only guide, split, and modulate light, but cannot emit light on its own. InP has a direct bandgap of approximately 1.34 electron volts, enabling efficient conversion of electrical energy into photons. Therefore, whether it is Nvidia, Broadcom, Marvell, or Cisco's silicon photonics platforms, InP lasers must be equipped as light sources within the package.**

The evolution from pluggable transceivers to co-packaged optics (CPO) architecture changes the installation location of lasers rather than removing them from the bill of materials. In fact, the transition to CPO pushes laser types toward harder-to-manufacture high-power continuous-wave light sources and external laser modules to drive multiple optical channels. Coherent's cooperation agreement with Nvidia covers this type of high-power CW laser, external laser source modules, and fiber array units.

Nvidia stated that compared to equivalent pluggable solutions, its photonic switches reduce the number of lasers per port by four times, improve power efficiency by 3.5 times, and enhance network resilience by ten times. However, these savings are calculated on a "per-port" basis, and it is the explosive growth in the total number of ports that is driving the steep rise in the demand curve. The high-end configuration of Spectrum-X Photonics supports 512 800Gb/s ports, with a total switching capacity of 400Tb/s.

Order-of-Magnitude Leap in Demand Scale

Hurlston used a set of comparative figures to intuitively present the intensity of this demand revolution: telecom customers purchase lasers in the "hundreds," while Nvidia and hyperscale cloud providers demand them in the "hundreds of millions." "Scaling from thousands of wafers to millions, and on non-silicon materials no less, is by no means easy," he said.

Lumentum recorded revenue of $808.4 million in the third quarter of its fiscal year, a 90% year-on-year increase, with component business revenue reaching $533 million and pump laser shipments growing 80% year-on-year. Even so, the supply-demand gap continues to widen. Hurlston stated on the earnings call that the degree of supply-demand imbalance had further expanded from the 25% to 30% disclosed in the previous quarter to "over 30%," with the supply tightness of pump lasers being "even worse than expected," and core components "effectively sold out for the foreseeable future." The company's guidance for fourth-quarter revenue is between $960 million and $1.01 billion, with an expected operating margin of 35% to 36%.

Coherent also recorded record quarterly revenue of $1.8 billion, a 21% year-on-year increase, with the proportion of revenue from data center and communications businesses jumping from about 41% a year ago to 75%. The order backlog extends to 2028. CEO Jim Anderson stated that the equipment yield of the company's 6-inch InP production line has surpassed that of the traditional 3-inch line. Internal capacity will double by the end of the June quarter, one quarter ahead of schedule, and will more than double again by the end of 2027.

Duration of Shortage: Key Difference from Memory Crisis

The tight supply of InP is not only a capacity issue but is also constrained by structural limitations of raw materials and geopolitical risks. Indium is recovered as a byproduct of zinc smelting, meaning its production cannot expand independently of the economic logic of zinc, regardless of how strong laser demand is.

Despite persistent warning signals, there is still divergence in market judgments regarding the duration of the shortage. LightCounting's market forecast in April 2026 showed that current transceiver demand exceeds supply by about 30%, consistent with Lumentum's data. However, the agency also expects the shortage to subside by the end of 2026 and lowered its growth expectation for Ethernet transceivers this year from 82% in 2025 to 65%. Coherent's capacity doubling one quarter ahead of schedule also points in the same direction.

The key difference from the memory crisis lies in this: The solution path for InP is a wafer size migration that is already running on production lines with yields leading older process nodes, rather than new wafer fabs that require three years to build from scratch. This supports relatively optimistic timeline judgments to some extent, but Hurlston's warning also has its realistic basis—the valuation of his company depends to a considerable extent on the duration of this shortage.