Crux Capital's Gaetano on the Copper Wall Inside AI Data Centers "Somebody's Already Buying the Materials Before I Even Make the Building"
Photonics became finance Twitter's word of the month. Optical stocks are flying, traders are piling in, and very few people can explain what it actually is.Gaetano of Crux Capital has been researching the industry for over a year, positioning in it before the trend arrived. He joined TradFi Essentials, the Blockchain 100 show on Binance Square, to explain what is real and what is noise.He has no formal finance background and has never worked inside a data center. He went full-time on his Substack three months ago, with over 50,000 followers on X and more than a thousand paid subscribers.Photonics in One Sentence"Very simply, it's just the transfer of data via light."The reason it matters now is scale. Trillions are being spent building AI compute factories, and a large share of that investment goes toward one problem: getting GPUs to communicate with each other fast enough."They need to send a ton of data, they need to send it really, really quickly," Gaetano said. "So what was once done through copper electrically is now increasingly having to be done optically through photons."The industry is not new. It powered the telecom buildout that laid fiber between continents and cities. What changed is the market dynamic — the amount of optical content required by modern data centers is, in his words, "pretty extraordinary and kind of hard to comprehend."Copper Works to Two or Three Meters, and That Distance Is ShrinkingThis is the mechanical heart of the thesis."Copper works pretty well up to two to three meters, and then you can do some things to make it go a little bit further to push towards like seven meters," Gaetano explained. "But as we're getting to higher speeds and higher bandwidths, that's actually getting shorter and shorter."The line that follows is the one to remember: "Every meter shorter is more optical content that needs to be built out."Optics started at the longest distances — subsea cables linking continents — and has been moving inward ever since. "What you're seeing happening is optics is going from these really long distances and increasingly taking over copper's share shorter and shorter distances into a data center."The endpoint is optics operating within a single rack, which he identified as the next major step up in optical content.The speed progression driving it: 400 gigabits per second moving to 800, then to 1.6 terabits. Each step change lifts revenue for suppliers through better pricing structures and more content per system."If We Can Make More, We Could Ship More"Asked where the bottleneck sits, Gaetano's answer was that it sits at every layer.He summarised the current earnings season bluntly: "Almost all the companies were saying, if we can make more, we could ship more. Our customers want more and more of our content, we just can't make it fast enough."At the foundation are indium phosphide substrates — flat discs of base material most downstream optical content is built from. "There's three main players in the world that hold like 80% market share, and there's like seven total that fill out that whole 100%."Capacity cannot be bought quickly. "You can't just throw money at it and overnight just be able to make much more of it. It takes time to bring these things online."The next layer down is equally constrained. Companies converting wafers into laser chips lack sufficient fabrication facilities. "Basically every step is constrained."His most striking line on the supply-demand gap: "Before I even make a building to create these materials, somebody's already buying the materials."Capacity scheduled to come online in 2028 is already being reserved. "We don't have any visibility on when that's really going to come back and balance."China Controls the Bottleneck, and Washington Has NoticedThe substrate concentration carries a geopolitical dimension."A good amount of it is in China, and China restricts the permits of shipping out these indium phosphide substrates to the rest of the world," Gaetano said. "So China kind of controls this bottleneck a bit."The Trump administration has proposed banning Chinese-made optical content and transceivers. His read on the market implication was direct: "If the US really wants to ban a bunch of Chinese supply, that is really bullish for non-Chinese supply, especially like the US-based companies."On the broader competitive framing: "The AI race is kind of like a cold war. Nobody wants to fall behind."That extends to why these companies command the valuations they do. "They are necessary for the entire AI data center buildout. Without them, we would lose the race to China. And the US can't really afford to do that."The Question Every Investor Has to AnswerThe core framework Gaetano offered was a single trade-off."Do you position yourself before proof and take on more risk? Or do you pay a higher price later on for more proof?"His own answer: "I personally lean a little bit more towards waiting for more proof before I put up any meaningful amount of capital."The reason is a gap he sees repeatedly in the sector. "Just because you have laser technology that works in a lab doesn't mean that you can commercialize it at scale. Having the tech capability is one thing, but can you actually bring it to market and commercialize it and bring it to scale at good yields is a different conversation."Asked which signal is hardest to fake, he named one: shipped orders followed by repeat orders. "Once they get those orders in hand and they ship them and the customer receives them and places follow-on orders, that's a really good sign that the customer sees value in this enough to want to buy more."He noted the frustration built into that discipline. "We saw a lot of these bottleneck stocks go up thousands of percent within months. And the actual companies aren't in that much of a different place than they were a few months ago."His Biggest Mistake: Hunting the UndiscoveredGaetano was direct about what he got wrong, and the numbers are stark.Around September of last year, he judged the established optical leaders already expensive and went hunting for undiscovered names instead."One of my mistakes was putting too much capital towards those unproven, more highly speculative stocks, whereas if I put more of my capital towards the more established leaders, I actually would have performed much better," he said. "Those higher speculative plays might have gone up 50 to 100%, where the other ones have gone up 1,000 to 2,000%."He extended the lesson into a general warning about chart psychology. Investors look at a vertical chart like Lumentum's and conclude the value has been extracted, then go looking for a downtrend they believe the market has mispriced."People spend too much time trying to find these undiscovered plays where the discovered ones are going up for a reason," he said. "If this stock is only going down, there's probably a reason for that. The market could be wrong, but more than likely the market is telling us that it's just not worth the investment."His conclusion: "Learning to buy charts that are going up."He also flagged trimming too aggressively around core positions — holding roughly 80% of capital in a name, cutting into strength, then failing to buy back before the stock runs on without him.Optical Shared Memory Is a 2029 StoryAsked about optical shared memory, Gaetano was careful to size it correctly."It's very interesting, and it is early. If anything, it might be a 2029 story."The problem it addresses: only so much fast-access memory can be packaged alongside a GPU. "The idea is how do we have memory that is farther away from compute that we can access quickly enough? And that's where we're seeing these photonic fabrics come into place, because you can't get that information quickly enough using copper."Marvell, following its acquisition of Celestial AI, sits at the forefront. But he placed it as a sub-sector rather than a driver: "It's more of a smaller sub-sector. That could be big, but we still don't have that much visibility on it yet."The architectures actually carrying the investment thesis are co-packaged optics, near-packaged optics and optical circuit switching. "Those are all going to unlock new markets that right now are like no revenue, and they're going to be potentially tens or hundreds of billions of dollars in revenue over the coming five years."His marker for taking optical memory seriously: more discussion on Marvell's earnings calls and broader industry proof.The Two Bear CasesGaetano recommended starting with what could go wrong. He named two things.Hyperscaler capital expenditure and its return on investment. "The entire AI trade is dependent on these hyperscalers increasing their capital expenditure, which we've seen so far. But there's another layer — does the market believe that it's sustainable and can they make it worth their investment?"He cited Microsoft's recent earnings call as a positive signal: spending heavily while remaining cash flow positive and expecting to stay there. But the risk is structural. "A lot of these companies are factoring in massive growth in 2027, 2028, 2029. If there's less visibility that their customers are going to be spending into that, they will rerate lower because their multiples aren't going to be rewarded anymore."Architecture timing slipping. Co-packaged optics unlocks a market that does not currently exist, and its arrival date is load-bearing for valuations."If that timing gets delayed further and further out because it's becoming too much of a challenge to integrate — or copper, we can figure out ways to make copper push for longer at higher bandwidth — then that can cause a rerating in the stocks."He noted this has already happened in cycles through the year, with CPO delay news knocking stocks before the market digested it and recovered.The Speculative Name He Avoids DiscussingGaetano named one early-stage company he deliberately underweights in his public writing: Aluma Technologies, which he says could largely bypass the indium phosphide constraint."If they can do that, they create another pathway for these optical components to be made that isn't reliant on the thing that is most capacity constrained. So it's hopeful, but it's very early stage and it's very risky."His reason for saying little: "People hear that it could create a great investment return and they put a lot of money into it. But the most likely outcome is that it's probably not going to pan out the way that the bull case implies."He drew a crypto comparison for the risk profile — riskier than established assets, though "not quite like a lot of the influx of meme coins we've seen, not to that extent where it most likely will get rugged."He mentioned Sivers Semiconductors as an example of the earlier hunt for undiscovered European and Japanese optical suppliers, a small-cap Swedish laser maker in a category where "we don't have enough lasers and everybody wants more of them.""You Can Just Do Things"The closing advice came with a number attached."For those first four months I was posting on Twitter, I was getting like 20 views per post. Nothing was happening. I could have very easily been like, forget it."What he did instead was keep publishing while reassessing method. "I might think I'm doing a good job, but if my metrics of success are telling me otherwise, then I should probably reconsider what I'm doing."His advice to anyone in their twenties: "Stick with something longer than they might want to just because it's not panning out, and then reassess constantly and try to figure out how to leverage your skill sets to make what you're delivering more valuable."The phrase he and his wife use, and the line that closed the episode: "You can just do things."
