Lin
Engineer & Venture Investor, The Fullstack Investor
All high-bandwidth data center interconnects will be optical within a few years, making photonics one of the most important AI infrastructure themes. The opportunity spans seven layers: materials and wafers such as AXT and IQE, tools such as Aixtron and Applied Materials, lasers such as Sivers and MACOM, foundries such as Tower and Fabrinet, testing such as Aehr and Teradyne, optics such as Lumentum, Coherent and POET, and networking such as Credo, Marvell and Ciena. x.comÂ
Optics and photonics should be among AI's biggest winners as faster GPU networks need optical connections that suppliers cannot expand quickly enough. Lumentum says some components are sold out through early 2029; Sivers is investing $30M to target more than 100M lasers a year, with operations expected in Q4 2027. Potential US restrictions on Chinese transceivers could reshape supply in 2027. Applied Optoelectronics and Sivers are the favorites. x.com x.comÂ
Coherent's pivot to AI optics is paying off as copper gives way to optical links inside large AI clusters. Quarterly revenue topped $2 billion for the first time, up 42%, with datacenter and communications up 59% to 79% of sales. Multi-year contracts with AI leaders including Nvidia extend order visibility into 2028. It launched its PhotonLink interconnect platform at ECOC 2026, and its indium phosphide, VCSEL and 6-inch wafer strengths give it a moat in 800G and 1.6T transceivers. x.comÂ
Credo's interconnect opportunity grows as campus-scale AI clusters require more links to validate and integrate. The OCP 2026 meeting takeaway is that complex requirements put interconnect at the center, with both copper and optical links set to benefit. x.comÂ
Lumentum's Raman pump lasers could benefit as more inter-site AI network spending shifts to optical line systems. Cignal AI forecasts those systems taking 57% of this spending by 2030, up from 29% in 2026, and identifies Raman amplification as a major cost driver and pump lasers as one of the tightest supply-chain components. Lumentum offers those lasers and the amplifiers built around them. x.comÂ
Tower Semiconductor is becoming to photonics what TSMC is to GPUs: a foundry building the next AI stack. It fabricates silicon photonics chips that combine lasers, modulators and detectors on one platform, and is integrating Lightwave Logic into that platform right after massive AI optical spending. As AI networks scale to 800G and 1.6T, photonics becomes the next bottleneck, with limited supply and exploding demand. x.comÂ
Networking is the next AI hardware trade after memory, whose share of rack value rose from 15% to 35%. Optics is moving inside the server, as Nvidia's Feynman architecture builds in co-packaged optics and silicon photonics. Scale-across networking could grow 30% a year just linking existing data centers, as inference moves closer to users and training spreads across sites. Supply is tight: Lumentum's CEO said it can meet only 30% of expected 2027 demand and sees optics tighter than memory. x.comÂ
AXT makes the substrates underneath the AI optical stack, InP, GaAs and germanium, and the market is finally noticing. InP feeds the lasers in 800G and 1.6T transceivers: no InP, no high-speed optics. Only a handful of companies produce these wafers at scale, and as hyperscalers race to upgrade AI networks, the materials layer is getting tight. Sitting at the overlooked bottom of the stack, it could be a 10x play. x.comÂ
AI data-center spending flows beyond GPUs and operators into networking, optics, power and cooling. Nvidia, AMD, Broadcom and Marvell supply chips; Arista, Nokia and Credo connectivity; Applied Optoelectronics and Lumentum optics; Vistra, Constellation and Bloom power; Vertiv and Modine cooling. Meta's $140B spending this year illustrates the scale, per a commenter with whom the point is agreed. x.com x.comÂ
Applied Optoelectronics makes optical transceivers for data-center networks, which AI scaling needs. x.comÂ
Optical networking's broader rollout will be slow because scale-up and scale-out require orders of magnitude more connections than today's limited optical distance links. Supply-chain discussions echo that slow pace. NPO/CPO starts in switches and reaches XPUs and GPUs much later, with new challenges. Optical backplanes will be necessary as racks move to optical, though copper will remain in many implementations. x.com x.com x.com x.comÂ
