On July 8, 2026, Lumentum CEO Michael Hurlston warned at the RAISE Summit in Paris that the supply shortage of indium phosphide (InP) could become more severe than that of DRAM or NAND. According to Tom's Hardware, which reported his remarks that same day, Hurlston counted the company's InP sites as "five." However, Lumentum's publicly disclosed existing mass-production infrastructure spans four sites, with a fifth—the Greensboro, North Carolina facility in the US—currently under renovation and scheduled to begin mass production in mid-2028. In the same July remarks, Hurlston explained that the company's shipments were falling more than 30% short of customer demand, and that the volumes requested by NVIDIA and hyperscalers had jumped from the hundreds of units typical of traditional telecom customers to hundreds of millions of units. This figure—a supply-demand gap exceeding 30%, limited to EML and pump lasers—was the same number presented during the May earnings call.
This is a strong statement, but the evidence for a tightening supply-demand balance is mounting. In March, NVIDIA invested $2 billion each in rival companies Lumentum and Coherent. Both agreements include multi-year purchase commitments worth billions of dollars, as well as rights to access future production capacity. The fact that a GPU company simultaneously funded two competing laser companies illustrates the risk that optical components could delay the buildout of AI clusters.
However, interpreting the phrase "worse than memory" as a price-increase forecast misses the actual situation. The InP problem lies in the narrow supply chain stretching from raw materials to lasers, and in the structural reliance on InP for light sources even as the industry shifts toward silicon photonics. While measures to increase supply have begun to move, the timing of when they will take effect varies widely, ranging from 2026 to 2028.
Behind 90% Revenue Growth: A Company-Wide Supply-Demand Gap Exceeding 30%
Lumentum's revenue for the third quarter of fiscal year 2026 exceeded $800 million, up more than 90% year-over-year. Of this, revenue from components such as lasers exceeded $500 million, up 77% year-over-year. Revenue from 200G EML more than doubled from the previous quarter, and pump laser shipments grew 80% year-over-year. Even so, during the May earnings call, Hurlston stated that the supply-demand gap for EML and pump lasers exceeded 30%.
EML (Electro-absorption Modulated Laser) integrates a distributed feedback laser that emits light and an electro-absorption modulator that carries the signal onto a single InP chip. It is the core component carrying up to 200G per wavelength in 800G and 1.6T pluggable optical transceivers. The reason the shortage rate hasn't narrowed even as Lumentum increases production is that customer-side adoption of 800G and the transition to 1.6T are progressing even faster.
NVIDIA's contracts are designed to secure this volume shortage through capital and long-term orders. The agreement with Lumentum is non-exclusive, but beyond the $2 billion investment, it includes purchase commitments and access to future production capacity. Since NVIDIA invested the same amount in Coherent on the same day, it has effectively secured supply from two major players in advance while avoiding dependence on a single vendor. This goes beyond short-term component procurement—it represents a reservation of production capacity that will underpin optical interconnect architectures for years to come.
InP Light Sources Remain on the Bill of Materials Even After the Shift to CPO
Silicon is well-suited for mass-producing circuits that guide, split, and modulate light, but it is difficult to achieve efficient light emission using standard silicon photonics processes. This is why InP-based lasers, which are direct-bandgap semiconductors, are combined for the light source. In the current configuration where optical transceivers plug into the front of switches, EML and CW (continuous wave) lasers are used; even with CPO (Co-Packaged Optics), which places the optical engine near the switch ASIC, high-power CW lasers remain necessary.
NVIDIA's CPO switches separate the laser from the heat-generating ASIC, consolidating it into a front-mounted, replaceable external laser light source. According to NVIDIA, this reduces the total number of lasers to a quarter compared to conventional configurations. Meanwhile, the maximum configuration of Spectrum-X Ethernet Photonics features 512 800Gbps ports, reaching a total bandwidth of 409.6Tbps. NVIDIA's own materials alone don't reveal how much the laser reduction offsets the increase in port count and bandwidth. What is certain is that even with CPO, InP light sources don't disappear from the bill of materials—rather, demand shifts in form toward high-power CW lasers.
What CPO changes is not whether InP is needed, but which type of device is required. In pluggable transceivers, EML—which directly modulates data—continues to grow, while CPO increases demand for high-power CW lasers that distribute light to multiple optical engines, along with external laser modules. Sumitomo Electric estimates that the share of CW lasers in the volume mix of in-datacenter optical devices will rise from 24% in 2024 to 69% by 2028. During this transition period, demand for old and new architectures is likely to overlap.
The Chain Reaction of Raw Materials, Export Licenses, and 6-Inch Substrates
The supply chain for InP lasers is narrow even further upstream from the finished-product factories. According to the US Geological Survey (USGS), global refined indium production in 2025 was estimated at 1,100 tons, with China accounting for 760 tons, or 69.1%. Indium is primarily recovered as a byproduct from zinc ore processing residue, making it difficult to simply increase mining output in response to rising laser demand alone.
| Stage | Supply Constraint | Requirements for Increased Production |
|---|---|---|
| Refined Indium | China accounted for 69.1% of 2025 production | Zinc smelting and recovery capacity |
| InP Substrates | China's export licensing, crystal growth, diameter expansion | High-quality 4–6 inch substrates |
| Epi Wafers | Growing application-specific compound semiconductor layers on substrates | Epi growth equipment, customer qualification, long-term supply contracts |
| Laser Chips | EML and high-power CW require different processes and designs | InP fabs, yield, assembly and testing |
On February 4, 2025, China added InP substrates to its export control list. AXT's Chinese subsidiary Tongmei obtained licenses for some European and Japanese customers on June 11 of the same year, but the company states it cannot predict the review timeline for applications going forward. Even if raw materials are secured, there's no guarantee that substrates can be exported across borders.
Companies are also expanding long-term contracts covering upstream production capacity. AXT signed a development and supply agreement with Coherent for 6-inch InP substrates dated June 25, 2026, with a contract term of three years. Major epi wafer manufacturer IQE and Tower Semiconductor also announced a multi-year contract on June 15, involving minimum purchase and minimum supply volumes. Plans to expand laser fabs will only translate into actual shipments once substrate and epi wafer supply can scale at the same pace.
The 2026 Diameter Transition and the 2028 New Factory
One means of increasing supply from existing manufacturing infrastructure is the shift to larger-diameter InP wafers. Coherent explains that moving from 3-inch to 6-inch wafers quadruples the number of devices obtainable per wafer and can lower chip costs by more than 60%. According to the company, the 6-inch process has already entered mass production, with yields exceeding those of the 3-inch process. The company plans to double InP production by the end of 2026 relative to the end of 2025, and more than double it again by the end of 2027.
Because this leverages existing sites, it may contribute to supply faster than new factories would. Coherent is expanding 6-inch production in Sherman, Texas (US) and Järfälla, Sweden, with further deployment planned in Zurich, Switzerland. Sumitomo Electric is also expanding its 4–6 inch substrate and InP optical device production capacity toward 2028. LightCounting's market forecast that "the shortage could be resolved by the end of 2026" doesn't directly cite these individual company plans as its basis. However, if the diameter transition and yield improvements materialize as planned across companies, the resulting supply increase would align with that forecast.
By contrast, the 240,000-square-foot factory Lumentum acquired in Greensboro, North Carolina, is converting a GaAs process line to 6-inch InP, with mass production scheduled to begin in mid-2028. This is a large-scale project involving investment of several hundred million dollars and maintaining or creating over 400 jobs, but it won't be ready in time to address the 2026 shortage. What will determine near-term supply-demand dynamics is expanded production at the existing four factories, while the new factory will be the facility that absorbs demand once CPO adoption accelerates from 2028 onward.
The Scope of the "Worse Than Memory" Warning
In terms of price movement, the current memory shortage is far clearer. TrendForce forecasts that general DRAM contract prices will rise 90-95% quarter-over-quarter in the first quarter of 2026, with NAND rising 55-60%. In the second quarter, DRAM is expected to rise another 58-63% and NAND another 70-75%. InP lacks a contract price index with the same level of transparency, making it impossible to compare the two markets solely based on shortage rates.
Research firm LightCounting estimated in April that optical transceiver demand exceeds supply by 30%, while also suggesting the shortage could be resolved by the end of 2026. In June, while estimating that first-quarter optical transceiver and AOC revenue grew more than 90% year-over-year, it also reported that supply constraints have begun to ease. If GPU supply grows, the optical component shortage could persist into 2027; but if GPU supply tightens, transceiver demand would also be suppressed. Looking at InP alone doesn't determine the timeframe of the supply-demand imbalance.
What underpins Hurlston's warning is not the magnitude of price increases but the difficulty of substitution. Memory supply can be increased through massive capital investment in bit output, but InP must simultaneously pass through zinc byproduct extraction, substrates, epi wafers, and laser qualification. Conversely, if the shift to 6-inch wafers achieves the fourfold increase in device count and higher yields the companies are planning, the shortage could ease without waiting for new factories. Whether the warning becomes reality will be determined by the race between the rate of demand growth and the ramp-up of the 6-inch process.
The EML and CW laser shipment volumes Lumentum reveals in its next earnings report will serve as the first indicator. If the supply-demand gap exceeding 30% narrows, confidence in the scenario of easing shortages by the end of 2026 will increase. If the gap remains wide, the reasoning behind NVIDIA's $2 billion investments in each of the two rival companies will carry even greater weight—well before the 2028 new factory comes online.