# The Real Bottleneck in the Optical Interconnect Cycle Is InP ### AI infrastructure's next round, and what the market hasn't priced in yet [Damnang](https://substack.com/@damnang) Apr 22, 2026 88 10 7 The Optical Investment Map I published last time drew more response than I expected. It was an article written from the perspective of an optical investor, laying out the full picture of the optics space and sharing my own take on where the next round of investment should go now that the obvious optical names have already run hard. Readers who had already been deep in the optics space gave me positive feedback that this single piece gave them a lot of insight. But I still got a lot of follow-up questions, because optics is a very wide domain. With seven layers and dozens of vendors inside each layer, the most common questions were “where do I start” and “what do I actually buy.” I felt strongly that I needed to do follow-up work on which company inside the optical stack gets re-rated first, which weaknesses get resolved and when, and what the specific trigger is. Starting with this article, each installment will pick one theme the market hasn’t yet fully reflected in price and analyze it in detail. If you haven’t read my Optical Investment Map yet, I strongly recommend going through that first before continuing here. **Today’s theme is InP.** Three things to unpack. First, technically, why InP is the real bottleneck in optical Interconnect that nothing else can replace. Second, why InP is structurally going to stay short on both demand and supply over the next three to five years. Third, a quantitative comparison of the six companies sitting in that real bottleneck section of the cycle, across five investment factors, with each company’s attractiveness, risk, and my own investment take. **Disclaimer** _Every company analysis in this piece is a thinking framework for information purposes, not a buy or sell recommendation. The per-company scores reflect my analytical criteria; they are not universal answers. You need to decide based on your own portfolio context, risk tolerance, and additional research._ ## 1. Why InP, why now ### Why nothing else works Optical Interconnects, at the end of the day, is the process of converting an electrical signal into light, then converting that light back into an electrical signal on the other side. Optical fiber sits in between. The structure looks simple, but the hardest part in practice is generating that light stably. The core device that generates the light is the laser. For the high-speed 1310nm and 1550nm optical modules used in AI data centers, the EMLs and the many CW lasers that serve as the light source are effectively built on the InP platform. More precisely, the devices are fabricated by growing an epitaxial layer on top of an InP substrate. On the receive side, certain InGaAs-based photodetectors and specific long-reach, high-performance receiver structures also depend on the InP ecosystem. In silicon photonics transceivers, Ge-on-Si photodiodes are the commercially deployed platform. But on the transmit side, where light actually has to be generated, InP sits in a structural bottleneck position. Moving into the CPO era makes the dependency on InP one level stronger. CPO brings the optical engine right next to the ASIC to improve power efficiency. The problem is that location is very hot. Lasers are temperature sensitive, so putting them directly inside that environment is difficult. In practice, this means pulling the laser outside the package and running it as a separate ELS, or External Laser Source. And the core of that ELS is again an InP-based CW laser. This is where the concept of lanes becomes important. In the previous pluggable optics era, the number of lasers needed per module was relatively limited. But in a CPO structure, a single switch ASIC has multiple optical engines attached to it, and each optical engine contains multiple lanes. Laser demand no longer scales with module count. It scales with lane count. The implication is big. A market that used to be sized by counting modules is now a market that has to be sized by counting lanes. InP laser demand isn’t just increasing. The unit of measurement for the market itself is changing. ## 2. The demand case ### A market that shifts by an order of magnitude The question now isn’t why InP is needed. It’s how fast that need is going to grow from here. The core of the story, the way I see it, is that this is more than simple growth. It isn’t that the optical components market is getting bigger. It’s that the weight of optical components inside AI infrastructure itself is shifting. AI data centers move from GB200 and GB300 through Vera Rubin to Rubin Ultra, and with each generation the number and value of optical components per rack goes up together. By Goldman Sachs’ estimates, combined scale-up and scale-out optical component TAM grows close to tenfold going from GB300 NVL72 to Rubin Ultra NVL576. This isn’t just market expansion. It means the importance and dollar content of the optical components needed to build a single rack are moving to a level completely different from the previous generation. The shift happens on two axes simultaneously. The first is speed. Transmission rates move from 800G to 1.6T, then to 3.2T. Even if the number of links stays the same, each link demands higher component performance and commands a higher price. A 1.6T module carries a meaningfully higher ASP than 800G. Link count can stay flat and the market still grows. The second axis is connection density. As clusters scale and network topology gets more complex, the number of optical modules attached per GPU goes up. The industry tracks this as attach rate. Goldman Sachs’ estimates put GPU-to-optical-module ratios at around two to three in the GB200 and GB300 era, moving to four to six starting from Vera Rubin. GPU port speeds are climbing, and larger clusters need denser interconnect. AI infrastructure isn’t just demanding faster links. It’s demanding more optical connectivity at the same time. ### Silicon photonics penetration This demand expansion connects to a structural change in module architecture. In the datacom optical module market, the transition from EML-centric structures to silicon photonics based structures is moving fast. Public market data shows silicon photonics share has already climbed to meaningful levels, with some forecasts calling for it to reach roughly half the market within a few years. The key point is that even as the architecture shifts, the need for InP does not disappear. If anything, the more complex the module architecture gets, the more strategically important external laser supply becomes. Silicon photonics proliferation doesn’t eliminate the slot where InP is needed. The market structure changes, but InP stays in its core position. ### Demand beyond AI infrastructure On top of that, InP demand isn’t fully explained by AI data centers alone. AI infrastructure is the biggest growth engine, but the list of areas that need InP is wide. High-speed InP-based devices are important for high-frequency signal processing in AI RAN and 6G fronthaul. In quantum computing, InP-based devices are core candidates on both the single photon source and detector sides. In autonomous driving LiDAR, 1550nm eye-safe lasers matter, and InP-based devices are strong in that window. Beyond these, there are smaller but high-barrier markets in medical diagnostics, space communications, and military IR sensors. The key point is that these niche markets don’t run on supply chains that are cleanly separated from optical Interconnect. They compete for the same substrates, the same epi lines, similar equipment and materials. When optical Interconnects demand rips, niche capacity gets squeezed too. And when niche demand runs stronger than expected, optical Interconnects capacity gets tighter. InP looks more like a single resource that multiple industries share at once. In that kind of structure, whichever market grows fastest drives the tension in the entire supply chain. ## 3. The supply case ### Why adding capacity doesn’t actually catch up If demand is rising this fast and supply keeps pace, the bottleneck thesis weakens. But this is exactly where InP differs from a normal semiconductor. There are plenty of capacity announcements, but there are several structural reasons InP supply can’t **ramp quickly.** **The first is the qualification cycle.** Optical components go through a long and strict customer qualification process. From the moment a new epi line or device line is built to the point where it receives production approval from an actual hyperscaler customer, it typically takes twelve to twenty-four months. The announcement can come fast, but the lag until that announcement converts to real revenue is long. **The second is equipment lead time.** Ordering an MOCVD tool doesn’t mean production starts. The tool has to be received, installed, the process has to be stabilized, and yield has to be ramped. Today’s order is a signal about future supply, not a solution to today’s shortage. **The third is that manufacturing know-how itself is a barrier to entry.** With InP, bringing in the equipment doesn’t automatically produce stable yield. Aligning crystal growth, wafer uniformity, epi quality, and laser performance consistently takes years of accumulated experience. At 6-inch the difficulty goes up another notch. New entrants can put up capacity numbers, but that doesn’t mean real supply capability follows at the same pace. **The fourth is raw material and geopolitical risk.** The InP supply chain isn’t just about wafers and tools. Indium, the key raw material, has a high concentration of supply in China and is exposed to policy variables like export controls. Those risks don’t stop at the raw material layer. They propagate down to substrate, epi, and device stages. The narrower the supply chain, the more small shocks get amplified. **The fifth is that the shortage spreads across the entire supply chain.** Major device makers are aggressively expanding their own capacity, but AI data center demand is likely to grow faster. In the field, word is that a few of the major optical component IDMs are moving very aggressively to ramp their own InP capacity. When that happens, the limited pool of substrate and equipment resources gets allocated to them first, and device makers without their own fabs get pushed to external epi foundries. External foundry capacity then tightens fast. The spillover dynamic is exactly what the market hasn’t fully priced in yet. Which is why the numbers say supply is expanding, but people in the market keep feeling that it’s still short. The fact that the words you actually hear on the ground are “InP epi capacity,” “MOCVD lead time,” and “substrate supplier diversification” tells you something. Before the market confirms it in numbers, the people inside the supply chain already sense where things are clogging. The core of it is simple. Demand isn’t just rising, it’s structurally re-accelerating. Supply is rising too, but neither the pace nor the quality of that expansion is keeping up. The gap between the two is the InP cycle. The optical map splits into seven layers, but each layer splits again into multiple steps. The InP theme inside Layer 1 is the same. Substrate, MOCVD tools, epi foundry, device, IDM. The steps are different, and the timing of earnings impact is different in each step. Inside the same InP theme, some companies have already run hard and some the market hasn’t fully looked at yet. I've covered why InP matters and why this cycle is structural. Behind the paywall, I narrow it down to the six companies with the purest InP exposure in the supply chain and run a quantitative analysis across five investment factors. Why these six, what each one's investment case and risks look like, and how I view them.