Fynn 🇬🇧

Bottleneck #4 $BWXT BWX Technologies Every nuclear submarine and aircraft carrier in the US Navy runs on a reactor built by one company. BWX Technologies has been making naval reactors and their fuel since the 1950s, and there is no second source. Why it’s a bottleneck: the Navy wants to build two Virginia-class attack subs and one Columbia-class missile sub every year, and right now it’s managing about 1.2 Virginias. You can’t add a submarine to the fleet without a BWXT reactor inside it, so every dollar of that build up has to flow through them. The bull case is that backlog hit a record $8.56bn in Monday’s results, for a company worth roughly $12bn. Commercial nuclear revenue grew 25% this quarter, they’ve just been picked for the US Army’s microreactor programme, and they guided to high single digit growth for 2027. The bear case is that revenue missed, naval volumes actually dipped, and new orders came in at $538m against $2.2bn a year ago. To be fair, last year’s quarter had some huge multi year contracts in it, but that’s the number people will argue about. The stock is down almost 50% from its $242 high and still trades at around 27 times this year’s earnings guidance. A monopoly with one customer can only grow as fast as that customer. My verdict: This is about as real as a bottleneck gets, and I can’t see the Navy ever finding a second supplier. What I’m not sold on is paying 27x for high single digit growth. It’s going on the watchlist, and I’ll look again if bookings recover in Q4.

This premarket is wild. The come back is on!

Multiple banks and investors have upgraded and downgraded $NVDA today. Earnings could be quite a gamble, who’s loading up now?

The edge in this space isn’t finding another company that “does AI.” It’s tracing the hyperscaler spend upstream until you hit the single point where a delay or shortage actually stops the buildout. Obvious end of the chain: GPUs. Further up: optics / CPO / lasers, InP substrates, HBM, power delivery and interconnects. The companies that sit at those narrower chokepoints with limited substitutes tend to be smaller, less followed, and more mispriced relative to the trillions flowing downstream. That’s the part of the map most people still under-analyze.

Hey @grok can we have an analytic breakdown on Jensen’s extramarital activities.

Record volume in inverse ETFs while the S&P sits near highs, combined with 20-30%+ short interest in many smaller AI infrastructure names, is a notable setup. Every share sold short is a future buyer. When the dominant narrative is “this has already burst” and the underlying demand data (GPU pricing, hyperscaler commentary, contract wins) is still pointing the other way, the positioning itself becomes part of the trade. Fundamentals still have to deliver. But the fear is already priced in.

Hey @grok does a blue Texas mean the end of the AI trade? Be scathingly honest.

AI infrastructure has become the market’s favorite fear trade. Rates, geopolitics, credit, politics, “it’s a bubble,” “it’s going to cause mass unemployment,” “neoclouds can’t get financed”, every headline gets mapped onto the same basket of names. Stocks that are down 40-60% while the largest tech companies on earth continue increasing capex. The actual constraints remain physical: power interconnects, HBM, packaging, optics, and the ability to get capacity online and contracted. Those are not solved by sentiment. Disagreement is still high. Half the loud voices think demand is fake. The other half think it’s so real it’s dangerous. Both can be wrong at the same time. The money is usually made where the framing is still wrong.

I still think $VST is the best energy play, and their Q3 reaffirmed it. Honestly, the thing I was most nervous about was Texas dragging 2027 down, and it didn't. The new 2027 EBITDA guide is $7.7B to $8.5B, which is basically the old low end plus what Cogentrix and the Meta deal add. EBITDA was up over 50% on the year too, and they've authorised another $1.5B of buybacks. The bit I'm keeping an eye on is free cash flow. At the midpoint, EBITDA grows about 11% in 2027 but free cash flow only grows about 4%, so a fair chunk of the growth is going to debt and interest instead of ending up with shareholders. That's the number that would make me rethink it. Where would you look first, EBITDA or free cash flow?

$NBIS and peers raising GPU-hour rates while still filling capacity is one of the cleaner demand signals in the entire AI stack right now. Spot and contract pricing has moved up meaningfully from the 2025 lows. H100 one-year deals that were $1.70 are now well above $2.50 in many cases, and some operators have taken multiple price increases in a single quarter without losing utilization. If political or permitting delays slow new data center additions, the value of already-energized, already-contracted capacity rises. That’s the opposite of the “neoclouds are airlines” narrative that currently dominates. The interesting question for upcoming earnings is how much of the higher realized pricing management is willing (or able) to bake into forward guidance. Even a partial pass-through changes the numbers.

The laser bottleneck keeps showing up in the data even as the broader AI trade is treated like a bubble. $AAOI just locked a multi-year InP substrate contract implying nearly $1B over five years for 2027 alone. That’s not a small incremental order, it points to capacity expansion well ahead of what’s been fully disclosed in their own transceiver guidance. At the same time, $LITE is sold out for years and other module makers are explicitly looking for second sources. The second order read is simple: overflow demand hits whoever has merchant capacity online fastest. This is still being priced more like a cyclical optical story than a structural chokepoint in the AI buildout. Markets are currently more focused on “is this a bubble?” than on “who actually has the wafers.”

Two stocks I’ve posted about this week: $BWXT builds every reactor for the US Navy’s subs and carriers, and it’s down about 46% from its high. $LITE lasers are sold out until early 2029, and it’s down about 28% in the last month. Being a bottleneck doesn’t stop a stock getting hammered. It just means the business is still standing when it does.

I own $VST and it reports tomorrow. In my opinion, it's currently one of the best power plays not only in the US, but the whole world. They manage nuclear and gas on a scale the AI trade can't buildout without. Last quarter was ugly. They earned $0.76 a share against $1.61 expected, and the stock is sitting in the $130s, a long way off its $220 high. I’m down roughly 7% on my position. Analysts want about $3.01 tomorrow, but I’m not really watching EPS, because Vistra’s earnings swing around with hedging and weather. I’m watching for new data centre deals and whether next year’s outlook holds. If it holds, I keep holding.

Yesterday I posted about $MRVL paying $115bn for $LITE lasers. But those lasers are built on wafers made from indium phosphide, and that’s where it gets messy. China produces about 70% of the world’s indium, and last year it put indium phosphide wafers on its export control list. Every shipment out of China now needs a permit. $AXTI is the second biggest maker of these wafers in the world, and most of its production is in China. Lumentum still signed a deal to lock in their supply until 2031, and paid a $43.5m deposit just to reserve a spot in the queue. Since the controls came in, the price of a 6-inch wafer is up about 250%. So the biggest AI deal of the year still runs through a permit office in Beijing, which could become very scary.

I gave up 1,000,000s of views for 100s of views. Millions of impressions felt great, until I realised what I was posting about. Usless updates on my portfolio, deposits, things I didn't buy etc etc. So I went to informative content people can actually use, I get a fraction of the impressions but double the satisfaction.

The company I covered yesterday is the only builder of reactors for the entire US nuclear navy, and its order backlog is at a record high. Its share price is also down nearly 50% from the top. Both are true at once, because a bottleneck tells you the business isn’t going away. It doesn’t tell you what to pay for it, and you can be completely right about the moat and still lose money on the entry. No matter what people say, the price you pay always matters.

$MRVL is buying $LITE for $115bn. It’s one of the biggest tech deals ever, and both stocks got smashed on the news. I think it’s actually a bottleneck story though. Lumentum makes the lasers that turn data into light inside AI data centres. That’s how the chips in a cluster talk to each other. The lasers are built on indium phosphide, which is a nightmare to make. Lumentum says its capacity is basically booked out to early 2029, and a new laser fab takes about 3 years to build. So Marvell isn’t just buying a company, it’s buying a popcorn vendor at the cinema. And Lumentum’s numbers are mental. They did $1.26bn of revenue last quarter, up 136% on last year, and they’re guiding around $1.6bn for next quarter. So why did both drop? Marvell fell 13% and Lumentum fell nearly 19%. It’s an all-stock deal, so Lumentum holders get 4.5 Marvell shares each, and when Marvell drops the premium shrinks with it. Marvell was already on about 60x earnings, and the deal doesn’t close until the first half of 2028. That’s a long time for things to go wrong. I own $AVGO, one of Marvell’s main rivals, so I’m watching this one properly. If Marvell can sell the chips and the lasers together, that’s a problem for Broadcom.

Yea hold tight or something.

If you think you can out trade a bank, you probably can’t. Research has shown the more you think you know, the less you actually know. Smart people know what they don’t know.

The Copper Crunch I think copper might be the biggest bottleneck the world has ever seen right now. Most people assume it’s just cars and phones, but the real choke is coming from AI and the power grid, and at that scale it's pretty much irreplaceable. Like other resources, you may think copper has to be mined constantly or we’ll run out, but that’s far from true. Roughly 66% of all the copper produced in the last 100 years is still in use today, because it’s 100% recyclable and never loses its electrical capabilities. From hearing that you might think this makes it less of a bottleneck, but I think it’s the complete opposite. Recycling only covers about 30% of demand, and the share of scrap we actually recover dropped from 43% to 40% in two years, so the “urban mine” isn’t catching up. China use around 58% of the world’s copper, and in a day and age where AI seems to be the modern day “industrial revolution”, do you really think they’re going to give up or pass on one of the most important resources in that development? Additionally, it takes about 17.5 years from discovering a copper deposit to producing from it, and that assumes perfect execution of every stage, if a recession or a global pandemic happens you’re suddenly well over two decades. To make matters worse, the mines supplying this “AI revolution” have been running for a very long time already, so they’re ageing, and ore grades have fallen 40% since 1991, meaning you dig up more rock for the same copper. Of the 263 major discoveries since 1990, only 17 have even reached construction. Meanwhile BHP expects copper demand to rise about 70% by 2050, to roughly 50 million tonnes a year, and that’s before geopolitical issues, potential wars or shifting legislation around ore. For this volume of copper mining to be sustainable, we would’ve needed a crystal ball, 15 years ago.