Temple 8 Research
AI Buildout, Macro & Energy Infrastructure Research
InP substrate demand and pricing still support the AI optics supply chain. Yunnan Germanium reports higher InP volumes from AI data centers and optical communications, plus higher prices. Q3 revenue rose 134.6% year on year, near AXT's guided 136%, but also includes germanium. IQE's H1 photonics grew 45%, partly from US defense funding. AXT's $66m Q3 guide covers only licensed exports or sales needing no license; a miss would more likely reflect licensing. x.comÂ
Indium phosphide substrates, AXT's business, should never become tight despite explosive demand, because Chinese capacity comes online quickly. InP demand should grow nearly 6x from 2026 to 2030, a 56% CAGR, driven by ultra-high-power CW lasers for co-packaged optics and InP photonic chips. Yet Yunnan Germanium alone should make nearly half of all InP production by 2030, so demand never fully exceeds substrate supply, which is bad for AXT. x.comÂ
AXT sits in the optical upstream, where surging AI compute demand meets supply that cannot be built fast enough. Amazon raised 2026 capex to $220B and said even that won't meet demand, with 2027 also short. AXT posted record revenue of $47.6M, up 164% year over year, record indium phosphide revenue of $30.7M and 45% margins targeting 50%+, and is doubling capacity in both 2026 and 2027 yet still can't keep up. x.comÂ
AXT is the upstream chokepoint of the photonics buildout, supplying the indium phosphide substrates every high-speed laser needs amid a massive global shortage. About 600,000-700,000 wafers shipped in 2025 against demand of 1.5-2 million, and AXT holds roughly 36% of a market three firms control. Record revenue rose 164% to $47.6M and gross margin hit 45% from 8.2%; Lumentum prepaid $87M to reserve capacity through 2031. The real risk is China, where slow export permits caused a guidance miss. … Show more
AXT looks very bearish on the risk that China stops allowing indium phosphide exports to the US. The same export question raises doubt over whether US companies will even be able to use CXMT DRAM. x.comÂ
Veeco is a top beneficiary of AI data centers' shift to photonics, as its tools make the indium phosphide lasers inside 800G and 1.6T transceivers and equipment orders lead the chip cycle. It disclosed over $250M of orders for its IBD, MOCVD and WaferEtch systems, shipping from late 2026 and ramping through 2027. Its laser annealing is critical to yields for 3nm and 2nm logic and HBM at makers like TSMC and Micron, and its pending Axcelis merger should add synergies. x.comÂ
AXT's indium phosphide business is at China's mercy, since its expansion inside China lets Beijing decide every line item on every purchase order. Per a report shared by @wallstengine, US officials are pressing China over delayed InP export approvals as the material becomes an AI data center bottleneck, and AXT says export permits are its biggest challenge. The stock was rejected at resistance despite that news. x.com x.com x.comÂ
IQE's new multi-year indium phosphide epiwafer supply deal with Tower Semiconductor puts a 5x-10x move from March 2026 levels in the works, and IQE is the pick over Tower itself. Per Wall St Engine, the wafers support Tower's silicon photonics platforms for AI data center optics, including 200G/lane pluggable transceivers, 400G/lane modulator prototypes and optical circuit switches. The two also settled their IP dispute, with Tower granting IQE a worldwide royalty-free license. x.com x.comÂ
Jabil looks like a compelling long at $38B because markets have not priced in its 1.6T LRO pluggable transceiver business. With Sivers as the H1 2027 bottleneck, the question is how much it can make, not whether demand exists. It has massive supply chains, took over Intel's pluggable lines and looks more scalable than Applied Optoelectronics' ATM-funded laser fabs. Recognition likely comes in H1 2027; a ~40% rerating seems plausible, though at its size tripling is less likely. x.com x.comÂ
Sivers Semiconductors' $30M expansion of its Glasgow InP fab, targeting over 100M CW DFB lasers a year by Q4 2027, likely means hard dilution. As @PhotonCap notes, the plan shifts Sivers toward hybrid manufacturing, and its capacity target and customer ramps remain forward-looking rather than committed volumes. x.comÂ