Photonic Interconnects: The Next Constraint

Transcript
[0:10] Hello and welcome to the AI Supercycle Podcast, where we follow the capital

[0:14] flows and physical constraints shaping the largest industrial build-out of our generation. From semiconductors, AI factories and hyperscale data centers, to energy markets and power grids, critical materials, orbital compute, and embodied intelligence.

AI Supercycle Begins
[0:29] We track where the capital is being deployed, where the binding constraints are emerging, and what it means for traders, investors, and the broader economy. Today is Tuesday the 21st of July at the time of recording and this episode will be published on Wednesday the 22nd. I've moved the podcast from a Monday release to Wednesdays because producing each episode is a much bigger job than it probably appears. Between the research, the writing, recording, editing, mixing and mastering, it was leaving me working right through the weekend. Now I really do enjoy making these episodes but I also enjoy having the occasional weekend off after a busy week in the markets. So a quick update on the podcast itself. Over the coming months, we'll be welcoming guests onto the show, including hedge fund managers, chief investment officers, capital allocators, and senior executives from across the AI infrastructure ecosystem. Now those conversations will add another perspective to many of the themes that we've been discussing. For now though, we're still laying the foundations. Before we can analyze the winners and the losers of the AI supercycle, we first need to understand the physical constraints, the capital flows, and the bottlenecks that are shaping it. I'm Tim Hardwick, founder of Quantum Field's Market Intelligence, and your host for today's episode.

[1:49] Okay, so on with the show, and over the last five episodes, we have worked through what I think of as the physical stack of the AI supercycle. The chips that do the computing, the power that runs them, the materials that build them, and last week, the possibility of taking some of this infrastructure into orbit. Each episode has been about a different layer of what I call the AI continuum, from the mind to the antenna, from geological time to orbital mechanics. This week we had something that connects all of those layers together. It's not a new layer, it's more like a nervous system running through every layer that we have already discussed.

[2:28] We are talking about photonic interconnect, moving data as light.

The Photonics Bottleneck
[2:33] And it is becoming one of the most important and least understood constraints in the entire AI build-out. Here is the simplest way I can put it. For the last few years, the AI industry has been obsessed with compute. More chips, faster chips, bigger clusters. And that made sense, right? Because compute was the bottleneck. But something has shifted. The chips are now so powerful that the connections between them cannot keep up. The data cannot move fast enough to feed the processors. The motorway cannot handle the traffic. And here is the thing that surprises most people when they first hear it. GPUs are no longer the most constrained components of the AI build-out. For the first time, entire data centers are waiting for power, transformers, networking equipment and cooling infrastructure before they can operate Computers move down the constraint hierarchy Data movement has moved up, That shift is what makes photonics one of the most important investment themes of the next 3-5 years And this is what we're going to explore today.

[3:38] Let me start with the problem, because it's worth spending a few minutes on this, because it's not obvious until someone explains it. Think about what happens inside a large AI data center. You have tens of thousands of GPUs, all working together on the same training run. And to do that, they need to constantly share data with each other.

[3:58] Every chip is passing results to every other chip, continuously, in real time. Now, inside a single server, the distances are tiny. Copper wiring handles that fine. Between servers in the same rack, copper still works, but between racks, between clusters, between buildings, that is where the copper runs into trouble.

[4:19] Think of it this way. You have a thousand supercars, each capable of 200 miles an hour, but they are all trying to drive on the same single-lane motorway. It does not matter how fast each car can go, the speed limit is set by the road, not the engine, and no amount of engine tuning fixes a traffic problem. That is the I.O. wall. I.O. stands for Import Output and the wall is simply this. You cannot push data through the chip faster than you can get it in or out. The processor sits idle, waiting for data that the connections cannot deliver quickly enough. Jensen Huang has spoken about this directly. He has said that compute can keep scaling but the bandwidth connecting processors together is not keeping pace. Every new generation of GPU makes the gap wider. The engine gets faster. The motorway has stayed the same. And if you are thinking that better copper might solve this, it won't. This is not an engineering problem, it's a physics problem. As you push more data through copper at higher speeds, the signal degrades. The heat builds up. The power consumption rises. The effective range shrinks. You are fighting the fundamental properties of electrical transmission.

[5:35] Light does not have any of those problems. A fibre optic cable can carry a signal across hundreds of metres or hundreds of kilometres with almost no loss. Almost no heat. And it is completely immune to the electromagnetic interference that plagues copper at high data rates. That is why the industry is transitioning from copper to light. Not as a nice upgrade, as a necessity. Now, photonics does not replace copper everywhere. That is an important point, and I want to, you know, be specific about that. To understand where photonics fits and where it does not, I want to walk you through four physical distances. From the smallest scale inside a chip, all the way out to the laser links between satellites in orbit. At each distance, the physics changes, the technology changes, and the investment thesis changes with it.

Inside the Chip Limits
[6:28] The first distance is inside a single chip, transistor to transistor.

[6:33] Modern AI chips are built at scales that are almost impossible to visualize. The most advanced processors today use transistors just 2 to 3 nanometers across. A nanometer is a billionth of a meter. At that scale, the connections between transistors are made using copper and aluminium wiring, embedded directly into the silicon. And photonics does not compete here. Not because it's not fast enough, but because light is physically too large for the job. So let me explain that. Light travels away with a physical size of several hundred nanometers, The components you need to guide and detect that light need to be at least that large. But a modern transistor is a tiny fraction of that size. Trying to fit an optical component into the space around a transistor is like trying to park a lorry in a bicycle space. It's not going to work, is it? It simply does not fit. And even if it could fit, the energy cost of converting electrical signals into light and back again over such tiny distances would outweigh any benefit. Copper is fast enough, efficient enough and small enough at this scale. There is no problem to solve. I start here because it sets the boundary on the thesis. Photonics does not replace copper everywhere. It replaces copper where copper fails. And inside the chip, copper does not fail.

[7:59] The second distance is larger, from one chip to another inside the same processor package. A modern AI processor is not a single chip. It is a collection of chips working together inside one housing, the GPU itself, the high-bandwidth memory stacks that sit alongside it, and increasingly, specialized chips for networking and data movement. All of these live inside the same package, connected to each other. Right now, these connections are made using silicon interposers and micro-bumps. An interposer is essentially a thin layer of silicon acting as a bridge between chips. Micro bumps are tiny solar connections that join one chip to another. It works, but it is reaching its limits.

[8:45] The problem is heat. As data rates increase, those copper connections generate more and more heat inside the package. And heat is the enemy of performance. When a processor gets too hot, it has to slow itself down to avoid damage. The industry calls this thermal throttling. More data, more heat, slower chip. It is the opposite of what you want.

Co-Packaged Optics Arrive
[9:09] This is where co-packaged optics come in, and I want to explain this clearly because it's one of the most important transitions in AI hardware right now.

[9:18] Instead of copper pathways carrying data between the GPU and the memory, you integrate a small optical engine directly into the processor package, right next to the chips. This engine converts signals into light almost immediately. The data travels as light across tiny channels etched into the package, then converts back at the destination. The copper journey and the heat it generates almost disappears. NVIDIA and Broadcom are both moving towards this approach, though at different points in the journey. Broadcom has already deployed co-packaged optics at the network switch level. The GPU package transition is still ahead. NVIDIA has indicated it plans to bring co-packaged optics into the GPU interconnects in the 2027 to 2028 timeframe. When that happens, optics will no longer be a networking component sitting at the edge of the system. It will be embedded inside the processor package itself. This is a significant moment.

[10:19] So companies to watch here. AR Labs, which is developing optical technology that embeds directly into chip packages. LightMatter, which is building systems that use light not just to move data but to compute with it. And Marvell, which acquired a company called Celestial AI specifically to accelerate this transition. This is the layer where most value will be created or destroyed over the next three years. It is early, it is real, and it is moving faster than most investors realise.

Transceivers Power the Network
[10:51] The third distance is the one most people are familiar with, even if they don't know it. Server to server, rack to rack, building to building. Those yellow cables you see running overhead in photographs of data centres are fibre optics. They have been there for years. At distances beyond a few metres, copper has already lost the argument. Fibre carries data as light with almost no loss, almost no heat, over distances that copper could not manage, The device that does this conversion is called a transceiver It takes an electrical signal from a processor, converts it into pulses of light, sends it down the fibre and converts it back at the other end, Think of it as an interpreter standing at the border between the electrical world and the optical world, The transceiver market reached roughly $24 billion in 2025 It is growing fast Current speeds are running at 800 gigabits per second 1.6 terabits per second is now moving from planning into real deployments this year, And the roadmap already points beyond that, To put 1.6 terabits per second in context that is the equivalent of streaming roughly 40,000 high-definition films simultaneously through a single cable.

[12:09] The scale of data movement inside an AI data center is genuinely extraordinary. Hyperscalers are expected to spend over $725 billion on AI infrastructure this year. The networking and interconnect share of that spending is growing faster than almost any other category. This is a large, fast-growing market where demand is clearly outpacing supply. When supply is the constraint rather than demand, pricing power sits with the producers. That is a good place to be. The key names in this layer. Coherent, which produces transceivers and also manufactures the substrate material that optical chips are built on, making it one of the few vertically integrated players. Lumentum, which is expanding manufacturing capacity specifically to meet AI demand. Fabrinet, which is a contract manufacturer assembling optical components for many of the major names and InnoLight Technology, one of the largest Chinese transceiver producers whose position in the supply chain matters to the geopolitical story that we'll come to shortly. Further upstream sits the fibre itself. A large AI campus may require thousands of kilometres of optical fibre when you account for all of the runs between buildings, racks and equipment. Corning and Prismian manufacture that fibre. They are further from the headlines but firmly inside the supply chain.

[13:37] There is one more signal worth paying attention to. NVIDIA's acquisition of Mellanox in 2020 for nearly $7 billion looked, at the time, like a networking play. In hindsight, it looks like NVIDIA understood the IO wall was coming before almost anyone else did. They bought the networking capability years before it became the constraint, The smart money was positioned early.

Laser Links in Orbit
[14:03] The fourth distance is the most dramatic. Satellite to satellite, in the vacuum of space. Now if you listened to last week's episode, you'll remember the challenge at the heart of the orbital compute thesis. Connecting satellites closely enough to run AI workloads across them requires enormous bandwidth between those satellites. We are talking tens of terabits per second. And those links are optical. Laser beams carrying data through space. It is the same fundamental technology as the data center transceiver, adapted for a completely different environment.

[14:39] Instead of light traveling through glass, it travels as a focused beam through a vacuum. The physics of light are the same, but the engineering challenges are very different. Both satellites are moving at roughly 28,000 kilometers per hour.

[14:54] The optical terminals must maintain precise alignment while both platforms are in motion. And this is not a trivial problem. Pointing a laser accurately at a moving target from another moving target across hundreds of kilometers of space requires extremely precise tracking systems. Companies like Minarik are building exactly these free space optical terminals. Kepler Communications, whose relay network already carries data from orbital data center nodes, is expanding. Starlink satellites already use inter-satellite laser links, so this layer is not theoretical, it is operational. And here is the connection that ties all of these four layers together. The same indium phosphide lasers that sit inside a data center transceiver are, in a modified form, the components that generate those inter-satellite laser beams. The same material bottleneck, the same technology, from the data center floor to low Earth orbit.

Indium Phosphide Risk
[15:54] Which brings me to the horizontal constraint running across layers 2, 3 and 4, the material that underpins all of them, and the supply chain risk that most investors have not yet priced, indium phosphide. Most people have never heard of it, but it is the foundation of modern optical communications, and here's why it matters. To convert electrical signals into light at the wavelengths used in fibre optics, you need a material with very specific physical properties.

[16:23] Indium phosphide has exactly those properties. Its crystal structure generates and detects light at the right wavelengths, with extraordinary efficiency. No other commercially available material can match it at the speeds that modern AI infrastructure requires. Even silicon photonic devices, which use standard silicon manufacturing for some optical components, still need an indium phosphide laser to generate the light in the first place. You cannot get the light without the material. Now here's where it gets really interesting. China controls approximately 70% of global primary indium production. An indium does not exist as a standalone ore. You cannot open an indium mine. It is a by-product of processing zinc, lead and tin. So supply does not respond quickly to demand signals. It grows slowly, indirectly and unpredictably.

[17:17] In February 2025, China added indium phosphide to its export control list. Not an outright ban. A permit requirement. Every shipment needs government approval. The effect was immediate. The price of a 6-inch indium phosphide wafer, the standard size used in manufacturing optical chips, roughly tripled, from around $1,400 to approximately $5,000. The substrate market itself is small, roughly $200 million, but think of it as the foundation of a pyramid. The optical module market built on those substrates is 50 times larger. The data centre infrastructure built on those modules is 50 times larger again. A small bottleneck at the base creates an enormous constraint at the top.

[18:06] And making it worse, even when new substrate capacity becomes available outside China, it takes 12 to 18 months of qualification testing before it can be used in production. New supply does not arrive quickly.

[18:20] There is a new indium-phosphide fab under construction in Eindhoven in the Netherlands, which is an important development, but meaningful volumes are unlikely before 2027 or 2028. The pattern should be familiar. A critical material, concentrated supply, geopolitical exposure, Slow qualification times Demand growing faster than supply can follow We documented the same structure in episode 4 with rare earths and copper, This is the same story in a different material, And China's approach is consistent across all of them Control the upstream Prioritise domestic industry Use export controls as leverage, That is not a coincidence That is a deliberate strategy, Okay, so now for the base case, and this is where we take a position.

Investment Base Case
[19:09] The four-layer framework gives us a structured way to think about where the opportunity sits and where it does not. Layer 1, on the chip. No thesis here, copper rules. Photonics does not apply. Layer 2, in package. This is the frontier and the most exciting layer analytically. Co-packaged optics is entering early production. The GPU-level transition is two to three years away. High conviction directionally, but with real execution risk The companies developing this, AIA Labs, Light Matter, Marvel's Celestial AI division Represent the highest upside if the transition happens on schedule.

[19:48] Layer 3, System to System This is the revenue layer today The market is large, the growth is real, The constraint is supply capacity Coherent, Lumentum, Fabrinet and the broader transceiver supply chain represent the most investable part of the photonics universe right now. The indium phosphide risk is live and worth pricing into any exposure here.

[20:13] Layer 4. Orbital. Speculative. Strategically important, yes, and directly connected to the thesis from Episode 5. Minarik and Kepler are the names to watch. The timeline is longer and the revenue is smaller, but the direction is clear. Across all three active layers, the binding material constraint is indium phosphide. The framework assigns a probability of approximately 75% that supply constraints in this material become a significant factor in AI infrastructure timelines within the next 18 months. And there is a broader point here that I think deserves some emphasis. The AI industry began with compute as a scarce resource. GPUs were impossible to get. That era is passing. Today the scarce resource is system-level throughput. An AI factory is only as fast as its slowest component, and increasingly, the slowest component is not the chip, it is the connection between the chips. Photonics does not solve a future problem. It solves a problem that is constraining AI deployment right now. And the market has not fully priced that yet. Here's the indicators that we are tracking. Indium phosphide wafer pricing and allocation trends.

[21:26] Non-Chinese substrate capacity expansion, particularly the Eindhoven fab timeline. Also, co-packaged optics deployment from NVIDIA and Broadcom. Transceiver shipment volumes at 800 gigabits and 1.6 terabit speeds and any changes to China's export licensing framework for Indium. This is the base case as it reads today. If the data change is enough to warrant it, we'll update the view.

Light Connects the Stack
[21:52] So that is the photonic interconnect thesis. Four distances, one technology, one material that underpins all of it. At the smallest scale, inside the chip, copper rules and photonics cannot compete. Moving out to the package level, the transition is happening right now. At the system level, photonics already dominates and the market is large and growing. And in orbit, the same technology connects satellites across the void. The AI supercycle is not just a compute story, it is a data movement story. And data movement runs on light. We started this series underground with the copper and uranium that come out of the earth. We moved to the surface, to the power grids and the data centers. We went into orbit and now we have found the thread that connects all three layers together. Light and photons. And a material most investors have never heard of, sitting at the base of the entire pyramid, from the transistor to the satellite. That is the AI continuum. And photonics is what connects it. You can find all of our work at qfmi.substack.com. The Market Pulse and In the Spotlight articles are free and always will be. The weekly outlook, the weekend debrief, and the monthly strategic report sit behind a paid subscription. Subscribe and you get the full picture.

[23:13] Before I go, a quick update on the book that I've been researching and writing for the better part of the last year, and it's currently in the editing phase. Trading the AI supercycle gives you a cross-asset framework for capital intelligence across the entire AI industrial economy. The book follows the money through the whole value chain, hyperscaler capex, semiconductor supply chains, data center build-out, power generation and transmission, critical materials physical ai and the emerging orbital layer, it draws on historical precedents like the railroads and electrification to show where the ai cycle is likely to rhyme with the past and where it won't, i'm aiming for a september release so i'll keep you posted here and you can get insights and updates by subscribing at qfmi.substack.com thanks for listening to the ai supercycle podcast see you next week.