Going into the week, you had two institutions telling you opposite things.

On one side, Morgan Stanley said the AI infrastructure pullback was a buying opportunity, driven by technical factors rather than deteriorating fundamentals.

On the other, Citadel Securities was telling anyone who would listen that Warsh was going to deliver a surprise rate hike.

So instead of believing one or the other, I did what I always do. I went to what was actually happening, not to a narrative being sold by someone who benefits from you believing it.

What we were witnessing was what a crowded trade unwind looks like. Semiconductor ETF inflows went parabolic this year: $46B in under seven months, more than the previous nine years combined. And now it was going out the same door it came in.

For those of us who lived through or studied 2008 and 2000, this felt different. Faster. More violent. More mechanical.

Look at Goldman’s US High Beta Momentum index. This was happening at unprecedented speed. We were living through history and we did not know what hit us, but boy did it hit hard.

Then the FOMC came along and the Fed did not hike. Nine in favor, three against.

So why was Citadel so vocal about a hike?

A few hours after the release, we found out. Leopold Aschenbrenner’s Situational Awareness had sold its entire public book on July 29. The buyer was Ken Griffin’s Citadel.

I am not going to comment on whether that was the right or wrong decision from Leopold. What I care about is how the market reacted, and it reacted like we had actually bottomed. The next day, Thursday the 30th, we got the hardest rebound since the unwind began on June 22.

One thing worth sitting with before we move on. A forced seller sets the price. The seven largest positions in that book were sold at roughly half their June peak, and bounced 25% the very next day. That is not a market discovering fair value. That is a market absorbing one participant’s margin call.

Part 2 — Korea

With each passing day it became clearer that what we had lived through was a mechanical unwind, not a broken thesis. And the biggest tell was how hard it was hitting Korea.

After what felt like forever but was actually five weeks, the KOSPI was sitting at an oversold RSI and a valuation JPM openly called crisis-level.

What pains me most is who was selling and who was buying. Day after day, on X and on Substack, I tried to reach as many people as I could to either hold or rotate back in. Every pointer said the same thing: institutions were buying, individual investors were selling.

Two panels. Foreign investors printing their biggest net buy of the year. Individual investors printing their biggest net sell. Same day, same index, opposite directions.

So does that mean we are in the clear, safe to trade, heading straight back to all-time highs?

Sadly, nothing is ever that easy. History says August and September are the two worst months for average S&P returns, and this is a midterm year on top of it.

But nothing about this July has been average. So I am doing what I do best: trade inside the AI buildout and look for the next opportunity to reclaim and pass my previous high.

Part 3 — The AI buildout

An exciting week. Microsoft and Amazon reported, and for those of us actually looking, so did AXT.

Microsoft delivered one of the cleanest earnings reactions we have seen in a while. Azure growth held up, EPS came in ahead of the street, and critically, management held its capital spending outlook rather than trimming it. That last part is what the market was actually asking about.

Amazon was the best of them. Andy Jassy, in his own words:

“I’ll start with AWS, which is booming right now… What we think is going on and why we’re enthusiastic about the ROI equation. Even with heavy CapEx next few years.”

AWS grew 36.7% year over year in Q2, the fastest in 18 quarters, with the AI and chips businesses each passing a $25B run rate. Growth accelerating and capex expanding at the same time. That is the combination the bears said could not happen.

AXT is the one only those looking hard enought got to cover it. Morris Young called Q2 an inflection point with a step-function increase in revenue, and the highest quarterly indium phosphide revenue in company history, driven by data center optical connectivity demand.

AXT also signed a definitive supply and capacity-reservation agreement with Lumentum for InP substrates, on top of the three-year Coherent deal carrying $22.29 million in prepayments. When a substrate supplier starts signing multi-year reservations, that tells you where the real chokepoint sits, and it is not the module.

After all was said and done, hyperscaler capex looks healthy. To a degree. Except for Oracle. My god, Oracle.

Look at that pink line. Oracle is spending close to 100% of trailing sales on capex while the others sit around 50%. They are betting the house and everything in it.

Part 4 — My portfolio

Read more

Why share this weekly? It’s not to show you how I’m doing. It’s to help you understand where my conviction lies, not in the particular names but in which layers.

Why does that matter right now? Because the tape has been living history. These past weeks have been one of the harshest high beta selloffs on record. I won’t go into every reason; there are a lot of factors in play and I like to keep things simple. If one sector runs too hot for too long, profit-taking is coming from somewhere.

What this has meant for retail

A lot of retail has capitulated. Margin is only useful if you know how to use it properly, and if you can avoid it, avoid it. Many of you have stopped being net buyers and have become net sellers.

So does that mean your intuition is wrong? Human nature says that when there is danger, you run. And there has been danger. Look at what the tape has done this month.

You are right to feel hurt. For a lot of us money is the key to goals we want to reach as soon as possible. I’m working toward a number that means my family is taken care of regardless of anything else I do. Watching that drop this month has been painful.

So I write this as much for myself as for all of you. I sincerely only care that we succeed, if not for ourselves then for the people we cherish most.

If it makes you feel any better, the average retail portfolio looks horrid in 2026. It doesn’t make me feel better. If anything, it makes me want to push my work to a broader audience. Bloomberg’s basket of the 50 most popular retail stocks fell 13% in July alone, the steepest monthly decline for that group since 2022.

A separate Jefferies tracker of Russell 1000 names with high retail ownership is down more than 25% since June. Meanwhile, SOXX, the closest thing to an AI buildout ETF, is still up roughly 75% year-to-date after peaking near +117% on June 22.

The index is fine; the people trading it are not.

So is the top in?

Short answer: not even close.
Long answer: hyperscaler CapEx projections keep climbing. Morgan Stanley has the four majors going from $413B in 2025 to $743B in 2026, $1,159B in 2027, and $1,287B in 2028. If anything I think those are conservative, because they price in AI and agentic AI as the pillars of growth and little else.

That money comes from somewhere and it flows somewhere. It moves from the hyperscalers building the AI to the hardware it gets built on. Some of that spend is cyclical. Some of it, the way NVIDIA’s did, turns structural. We’re seeing the first signs of that in memory with long term agreements, but we need more complexity in the stack before that holds as a rule.

Look at Alphabet. This is the first quarter of negative free cash flow in the company’s history. That is not a profitability problem, Google is still enormously profitable. It is capex outrunning operating cash flow, which is exactly what it looks like when a company decides the buildout matters more than the optics of a clean cash flow statement. If that doesn’t tell you where things are headed, I don’t know what will.

And this doesn’t stop in 2028 or in 2030. AI compute, in particular inference, is becoming the new oil. Maybe not in 5 years, but in 25. AI itself is a stepping stone to whatever comes next, and all of it needs the same thing to operate: compute.

The complexity that is hard to price in

We are living through several things at once. On top of everything above, this is a midterm election year, which usually behaves differently from a normal year for the market. Read that again, because nothing about this year has been business as usual. And there are external factors adding uncertainty on top. I won’t do political commentary, but the current macro backdrop is its own level of complexity.

Read more

I am not telling you that to flex. I am telling you that because the number is the least interesting part of the story. The interesting part is that every position that produced it sits on one map, and the map was built quarter by quarter, mistake by mistake, in exactly the order I am about to walk you through.

One of the things I like doing most, at my job and in life, is teaching people how to fish instead of handing them the fish. So this is not a victory lap. This is a twelve-month autopsy of my own portfolio: what I was buying each quarter, why I bought it, what broke, what compounded, and which decisions actually made the money.

Twelve months is also just long enough that the lessons stopped being opinions and started being scar tissue.

Read more