AI optics: a $43.5b market waiting on InP
AI is the fastest build-out the semiconductor industry has seen. Hyperscalers are racing to train models on clusters of tens of thousands of GPUs, and these clusters now stretch across multiple data centres. In these systems, computing capacity is no longer the only limiting factor. Moving data between chips, racks, and sites has become just as critical.
As lane rates climb from 100G to 200G and 400G per laser, copper runs out of reach and power budget, and light takes over: first between racks, then inside the rack, and eventually right next to the GPU through co-packaged optics. The lasers and modulators powering these links work at 1310 nm and 1550 nm, and at those wavelengths, InP is the material of choice. GaAs VCSELs at 850 nm still serve the shortest links, but they cannot follow AI beyond 100 m.
Yole Group’s latest report, 'Photonics InP and GaAs for AI 2026', puts AI optics bare die revenue at $43.5b in 2031, up from $2.3b in 2025, a 63 percent CAGR, with almost all of that growth from InP. The question is no longer whether demand will come, it says, but whether the supply chain can physically deliver it.
Acording to Roy Dagher, technology and market analyst, Compound Semiconductors, Yole Group: "The value in AI optics is moving away from the classic pluggable. By 2031, scale-up links inside the rack will capture 83 percent of bare die value, and three InP technologies, CW-DFB lasers, DWDM laser arrays, and IQ modulators, will take 88 percent of it. The winners will be those companies that master these devices at volume."
The InP chain has consolidated into a few vertically integrated players over recent years. Lumentum absorbed NeoPhotonics and Cloud Light, while Nokia took over Infinera for $2.3b in 2025. But since 2024, the playbook has changed from buying to building. Lumentum acquired Qorvo’s Greensboro site as its fifth InP fab. Nokia acquired NXP’s Chandler 6-inch fab in 2026 for an undisclosed amount. Coherent brought its first 6-inch InP fab into production across Sherman and Järfälla and raised $632m specifically for InP.
NVIDIA’s recent $2b investment in Lumentum and Coherent shows how strategic InP capacity has become. In addition, Nokia’s subsidiary Infinera purchased in February 2026 an 82,000-square-foot facility in San Jose, which will be dedicated to photonics, and SMART Photonics broke ground on a 6-inch pilot fab in Eindhoven, targeting full operation in 2028. Meanwhile, Chinese players are moving in, with Vertilite announcing an around $750m (RMB 5b) InP laser chip base in Changzhou. Yole Group estimates a total of $21b in worldwide investments since 2021 across InP and GaAs (acquisitions, funding, fab expansions, etc.)
Substrate makers are moving just as fast. AXT is doubling capacity in 2026 and again in 2027; Sumitomo targets three times its fiscal-2024 output by fiscal 2028; JX has committed up to around $760m (JPY 120b) to grow seven to ten times, and Vital Materials’ Quzhou project adds 3 million substrates a year by 2029.
Despite this investment, InP wafer supply is the real constraint. Capacity looks roughly twice the demand on paper, but once off-spec wafers, export-license delays, and volumes locked by long-term agreements are removed, sellable supply will only cover a portion of the 2026 demand, with a two-to-three-year gap to close.
Since February 2025, every Chinese InP shipment needs government approval, which has impacted wafer prices: for example, 6-inch InP wafer prices increased about 70 percent between 2024 and 2026 to roughly $4,000. 3- and 4-inch wafer ASPs follow the same trend in 2026, as AI demand continues to grow. Nevertheless, demand keeps climbing, with the substrate market reaching $1.25b by 2031, an 80 percent CAGR.
Ahmad Abbas, technology and market analyst, Compound Semiconductors said: "On paper, InP substrate capacity is about twice the demand. In practice, once you remove off-spec wafers, export-license delays, and volumes locked in long-term agreements, sellable supply covers only part of what the market needs. The remaining gap will take two to three years to close."
6-inch wafers will comprise 93 percent of InP substrate value by 2031, up from 36 percent today, and whichever of JX, Sumitomo, AXT and Vital ramps 6-inch yields first will capture most of the market. Coherent already produces on 6-inch, yet most laser dies still run on 3-inch and 4-inch. In 2026, 80 percent of lasers are grown on 2, 3 and 4-inch InP wafers.
The hard part is crystal growth: InP is brittle and prone to dislocations. The 2.25x area on 6-inch wafers only pays off if regrowth, passivation, and testing yields hold through.
The AI optics boom is real, but it will be decided at the wafer level. For the next two to three years, the market will be limited less by demand than by the supply of sellable, device-grade InP wafers. Device makers that secure allocation early and qualified second sources will ship. Others will wait. Behind them, the race to 6-inch is on, and whichever substrate maker ramps 6-inch yields first will take most of the market.
