Ignition Portfolio Update #5
What a week!
Saturday, August 1, 2026 · Five trades this week.
On Monday morning, CXMT went public in Shanghai.
Which means it happened Sunday evening here.
Shanghai opens at 9:30 in the morning, and 9:30 in the morning in Shanghai is 6:30 the previous evening on the West Coast, so the largest semiconductor listing in the history of mainland China took place while it was still light out and I hadn’t even cooked dinner yet.
The kind of listing that gets an article in the WSJ and then disappears — filed under Asia, filed under IPO, filed under interesting-but-not-about-us.
But it closed up 466%.
Wow!
Priced at 8 yuan. Closed at 49. 8.5 billion dollars raised, which made it the biggest mainland semiconductor offering ever, past SMIC’s in 2020. A $488 billion market value by the end of trading, which made a company most Americans have never heard of the most valuable listed business in China.
There were 9.4 million individual retail orders. The public portion was oversubscribed 212 times over.
I read that last one three times. Two hundred and twelve. For every share available, someone had asked for two hundred and twelve of them.
Anyway. Those are the numbers that got reported.
The number that mattered was on the page nobody screenshots. Use of proceeds.
CXMT said it would spend the money mainly on mass-producing memory wafers.
A memory maker’s IPO being 212x oversubscribed doesn’t sound like the memory trade is over — in fact, far from it.
But there’s no natural law saying it has to be Micron’s memory being bought en masse. And there’s something worse hiding in that oversubscription number, which is that the enthusiasm is the supply. Nine million people didn’t just express an opinion about memory demand on Monday. They wrote a check for the fabs that will eventually satisfy it.
That is what stage three looks like when it’s working. The demand for the stock is the funding for the wafers.
I had spent the last three weekends writing that memory was in the third stage of a very old cycle — undersupply, high prices, supply response, lower prices — and that Micron looked cheap at four times forward earnings only because the fourth stage hadn’t arrived yet.
Four stages. It is not a sophisticated framework. It is the kind of thing you could draw on a napkin, and in fact I did draw it, badly, before I made the actual chart, and the actual chart took two hours because I could not decide whether the fourth stage should be the same red as the first one. It should. It’s the same cycle. That was sort of the point.
Here is the part I want to be blunt about, because I was too gentle with it for three weeks.
Four times forward earnings was never a valuation. It was a countdown.
A cyclical company trades at its lowest multiple at the exact moment its earnings are highest, because that is when the market is most certain the earnings are temporary. The multiple is not telling you the stock is cheap. It is telling you what the market thinks the E in that ratio is worth as a description of the future — and the market’s answer, at 4x, was not much.
I knew this. It’s in every framework I’ve published. And I still spent three weekends explaining the fourth stage while owning a position that only worked if the fourth stage stayed hypothetical.
Monday, the supply response stopped being a forecast and started being a company with a ticker.
I sold all of my Micron at $914.34.
There is something absurd about that price. Two years ago Micron was a $100 stock that people used as a punchline about cyclicals. Now it has cents on the end of a three-digit number and I’m typing it into a sell order like it’s a normal thing to do.
The order filled. Nothing happened. Nobody called. The screen updated, the position went to zero, the cash line got bigger, and I went and did something else, and that was it.
That is the whole thing, really. Not a chart level. Not a stop. An event that converted an argument I’d already made into an argument I could no longer politely postpone acting on.
1. Forecasts vs. Events
There is a gap in this work that I don’t think gets talked about enough.
You can be finished with the analysis and still not be finished with the position. The two things feel like they should happen at the same moment. They almost never do.
I had written the fourth stage down. I had published the framework. I had defended it in the comments. And I still owned the stock, because the fourth stage was a thing I expected rather than a thing I could point at, and a thing you expect is very easy to keep expecting while you collect another few weeks of gains.
CXMT’s listing didn’t teach me anything I didn’t already believe. It just removed the last excuse for not acting on it.
Cyclical businesses don’t turn on the day the supply arrives. They turn on the day the market can see the supply arriving. Monday, the market could see it — an $85 billion pre-listing valuation for the world’s fourth-largest DRAM producer, priced in a market that had spent all of July worrying about exactly this, and the buyers turned up anyway. Nine million of them!
Micron’s numbers didn’t change on Monday. The multiple that the market is willing to pay for those numbers changed on Monday. That’s the fourth stage. It arrives in the multiple first.
1.1. Cycles
Every memory cycle ends the same way, so the bull answer to CXMT writes itself: they hold something like 7–8% of DRAM, they’re years behind on HBM, they can’t buy the lithography tools to close that gap quickly, and Samsung, SK Hynix and Micron still hold roughly 89% of the market between them.
All true. None of it is as comforting as it sounds, for three reasons.
One: the new entrant does not need a return on capital. CXMT is a national project. Its purpose is domestic self-sufficiency, and its shareholders just demonstrated that they will fund it at 212 times the available allocation regardless of what the return looks like. In a commodity, the price is set by the marginal producer — and the marginal producer here answers to an industrial policy, not a hurdle rate. A competitor who doesn’t have to earn its cost of capital is the single worst thing that can happen to a commodity business, because they will keep the line running through prices that would shut yours down.
Two: qualification, not capability, is the thing to watch. Reports this month said Apple has begun testing CXMT’s DRAM for devices sold in China.
Wait, what?
That is the leading indicator, not HBM production. You don’t need to be the best memory in the world to destroy pricing; you need to be good enough to be a second source. The moment a buyer has a credible alternative, they stop negotiating on availability and start negotiating on price. Commodity margins don’t die when a rival matches you. They die when a rival becomes a bargaining chip.
The 2nd largest American company in the world is considering CXMT for memory supply. Not Micron, Sk hynix, or Samsung!
Three: the moat has a clock on it. HBM is where Micron’s margins actually live, and CXMT is not competing there yet. But Samsung has just unveiled HBM4, and the fight for that socket is between three well-capitalized incumbents who all built capacity into the same forecast. Conventional DRAM gets commoditized from below by China while HBM gets competed from the side by Korea. Both ends compress at once, and the segment that’s supposed to be the escape hatch is the one where the incumbents are hitting each other hardest.
The last memory downcycle was a supply-and-demand accident. This one has a sponsor.
And memory as a whole is on a cycle, but Micron is in a cycle within the memory cycle.
1.2. Leopold
The other thing that happened this week is that a $45 billion hedge fund blew up. Well, it was only a $20 billion hedge fund in January of 2026, but still.
Leopold Aschenbrenner published Situational Awareness in June 2024 — 165 pages arguing that machine intelligence was arriving faster than almost anyone understood, and that the physical consequences would be enormous: chips, memory, data centers, power. He was 22. It became required reading in Silicon Valley. He left OpenAI, raised money against the thesis, and built one of the most watched funds in the market.
This week his brokers sold mostly the entire portfolio of public equities to Citadel at a discount. Assets under management went from $45 billion at the start of July to roughly $10 billion by Thursday.
I want to be careful here, because there’s a version of this story that’s just someone being unkind about someone else’s bad month, and I’m not interested in writing that one.
Here is what I actually think happened.
He was right about the destination. Reported leverage of up to 400% meant he could not survive the route.
When a leveraged portfolio falls, the equity cushion underneath it shrinks faster than the portfolio does. The brokers ask for more collateral. Raising collateral means selling. Selling pushes the same names lower, which shrinks the cushion again, which means more selling. A painful drawdown becomes a deleveraging spiral, and the spiral does not care what you think about superintelligence in 2030.
On July 24 he wrote to investors that the fund had not been immune to the decline, and called it one of the best buying windows since early 2025. A postscript invited clients to add fresh capital on August 1.
He may well be right about that too. Because on Thursday — the same day the liquidation completed — the exact names he’d been forced out of went vertical. Nebius up 27%. CoreWeave up 24%. IREN, Core Scientific, Riot all up 20% or more.
The seller was gone, so the selling stopped, so the prices went up. Stocks 101.
He was right about AI infrastructure. He was right that the sell-off was overdone. He is not going to get paid for either like he was in June, because the position and the thesis are two different objects and only one of them has a margin requirement.
What’s left of the fund, incidentally, is mostly its private holdings — the largest being a stake in Anthropic. The AI hedge fund survives as an AI holding company.
And there’s a second lesson underneath the leverage one, which is about what those names actually are. His largest disclosed positions were memory and neoclouds — SK Hynix, CoreWeave, Micron, Nebius. Not a random selection. He built the portfolio the thesis implied: if intelligence scales, buy the physical bottleneck. The trouble is that a bottleneck is only a bottleneck while it’s scarce, and everything on that list is a business whose entire margin structure depends on staying scarce.
Leverage killed the fund in July. Commoditization was going to be the slower problem.
1.3. Same Fear, Two Directions
Last Monday I published Who’s Afraid Of Open Source? One argument from it was that the wave of open-weight Chinese models — DeepSeek V4, then Moonshot’s Kimi K3 on July 16 — aren’t the disaster for American compute that the market treated it as. Free models get run more. But they still need compute. And on the other side, American companies must steeply decrease compute costs if they are to compete with open-source models. On all sides, more compute is needed, and fast.
But all compute is not created equal, and I was invested in the layer where it is most equal.
That was the honest resolution to the tension a few of you had flagged: how do you own Micron while calling the model layer a commodity? Answer: commoditizing the model increases demand for the thing underneath it.
I still think the logic holds. What I’m unsure of now is that people are seeing “demand goes up” as though it were the same sentence as “margins go up.” It is not. It has never been. Demand for DRAM has gone up for forty years and the industry has still managed to lose money in roughly twenty of those years.
Because look at what July actually delivered, in order.
July 1: Meta announces it will sell surplus compute to enterprises. Neoclouds drop 12–15% in a session. The customer became a competitor.
July 16: Kimi K3 lands. Worst week for U.S. chip stocks in over a year. Nvidia briefly loses the most-valuable-company title.
July 27: CXMT lists, and the proceeds are earmarked for wafers.
Those look like three separate stories. They’re one story told three times. Every layer of this buildout that can be commoditized is being commoditized, and the commoditization is arriving faster at each layer than the previous layer’s investors had priced.
China is doing both ends at once. Free the model, then build the memory that runs it. CXMT is where the open-source fear and the supply-response fear turn out to be the same fear wearing different clothes.
The demand is real. I was right about that. But there is nothing in “inference demand is real” that says the memory serving it has to be Micron’s, at Micron’s margins, at Micron’s multiple.
1.4. Two Layers Up
The neocloud version of this argument is the one I have been slowest to accept, because I own(ed) one.
A neocloud borrows money to buy a depreciating asset, rents it by the hour to a small number of customers, and reports the resulting bookings as evidence of durable demand. Every part of that sentence is fine while capacity is scarce. None of it is fine afterward.
Almost like a rental car service for GPUs, for those unfamiliar with the term “neocloud.”
Consider what July 1 actually established. Meta has enough surplus GPU capacity to sell it, and has decided to. That’s not a new competitor entering a market — that’s the largest buyers discovering they built an oversupply and choosing to monetize it. Their marginal cost is already sunk. Their capacity is already energized and permitted. They can price an hour of compute at whatever recovers something rather than nothing, and a company whose entire business is that same hour has no answer to that.
Then a look at how the demand gets financed. Nvidia holds a reported 9.3% of Nebius, stemming from a $2 billion partnership. Nebius signs a billion-dollar compute agreement with an AI lab that is itself venture-funded. The chipmaker funds the cloud that buys the chips to serve the labs that are funded by the same investors buying the chipmaker. Every link is a real contract. The circle is still a circle, and a demand signal that you partly financed is not an independent demand signal.
Put it another way.
I owe Savannah $10. She owes Nico $10. Nico owes me $10.
I pay my debt to Savannah, she pays hers to Nico, and he pays his to me.
Then we each report $10 of revenue.
Now scale it up to billions of dollars.
And a constraint that decides who actually gets to grow: New York has imposed a one-year moratorium on new hyperscale data centers. The binding limit on this industry is not appetite. It is power, permits, transformers, and interconnect queues — which means the winners are whoever already controls generation and land, and that is not the company renting you an H-series hour.
And then this week’s tell. Nebius fell 13% on Wednesday and rose 27% on Thursday, and nothing about Nebius changed in between. A forced seller finished selling. That is the entire explanation. A business being priced on cash flows does not move 40% in two sessions on a change in who is holding it.
Here’s how the two halves of this letter connect, and it’s the part that surprised me.
Cheaper memory does not save the neoclouds. It helps them, briefly, on input costs — and then it helps their competitors identically, and their competitors have balance sheets and captive demand. In a commodity, an input-cost saving that everyone receives is not a margin. It’s a price cut waiting to be passed through to the customer.
Stage four in memory and stage four in compute are the same stage. The scarcity ends, the pricing power moves to whoever owns something that can’t be built by an oversubscribed IPO — and what’s left underneath is a lot of very good businesses selling something nobody has to choose them for.
1.5. What Changed
My published position going into July was that the AI boom had become a bubble in chips and neoclouds, with energy, robotics, applied AI, and materials as the exceptions.
I’m narrowing it.
I now think the only two layers of this buildout still offering an attractive risk-reward are applications and energy.
Chips, models, and neoclouds all now carry some combination of the same five problems: circular financing between the buyers and sellers of compute, real debt obligations against uncertain returns, visible fatigue with capital spending announcements, revenue promises that keep getting pushed right, and commoditization from both open weights and new supply.
The regime change is in that third one, and it happened in public this week.
Alphabet guided 2026 capital spending to $195–205 billion and the stock fell 7%. Amazon’s long-term debt rose 81% in a single quarter, to $119 billion. Microsoft’s free cash flow is expected to go negative for the first time since at least 2001. Combined hyperscaler capital spending for 2026 is now estimated somewhere around $700–750 billion.
For two years, a bigger spending number was bullish for everyone downstream of it. That is no longer true. When the market starts treating capital expenditure as a cost rather than a signal, the supplier gets repriced before the spender does — because the supplier’s whole valuation was built on the assumption that the number goes up forever.
The energy layer is different, and this week showed why. Oil rose 6.6% on Wednesday to $84.46 on the Iran escalation. The Fed held at 3.50–3.75% with three dissents in favor of a hike, and the 30-year yield went to 5.20%, the highest since July 2007. Japan spent roughly $53 billion on Thursday buying yen — likely the largest single-day intervention in its history — and the New York Fed sold euros to buy yen on Friday.
None of that is an AI story. All of it is an energy-and-funding-cost story, and the data centers are the largest new energy consumer on the planet. The demand for electrons doesn’t get commoditized by an open-weight model release in Beijing.
Before we continue:
Get access to my personal portfolio, watchlist, insights, stock picks, deep dives and more with the annual plan. Subscribe or upgrade now!
2. The Free View
15 positions, 9 themes, and — for the first time since I started publishing this — a meaningful cash balance. Roughly 17% of the portfolio is sitting in cash on purpose.
Why? Because I’m entering a major rotational phase in the portfolio, and the macro environment is making me very uneasy.
Micron is gone. Nebius is a quarter of what it was. Cameco is three quarters of what it was. ReAlloys is new. The tickers for Kraken, Adyen and Kraken’s OTC listing still cannot display real-time information, because they are non-US companies.
Power & Generation GE Vernova ($GEV), Constellation Energy ($CEG)
Nuclear Cameco ($CCJ), NuScale ($SMR)
Critical Materials ReAlloys ($ALOY), Energy Fuels ($UUUU)
Space & Defense Karman S&D ($KRMN), Rocket Lab ($RKLB)
Physical AI Nokia ($NOK), Kraken Robotics ($KRKNF)
Fintech Pagaya ($PGY), Adyen ($ADYEY)
Software & Ad-Tech AppLovin ($APP)
AI Infrastructure Nebius ($NBIS)
Biotech Recursion Bio ($RXRX)
Cash ~17%
