Portfolio Review (July 2026)

+207.68% Since Inception, +127.91% Outperformance

Aug 03, 2026

∙ Paid

July 2026 Markets Review

Portfolio: -0.94%
S&P 500: -0.82%
NASDAQ: -4.15%

July was the second red month in a row for the indexes. It was a very mixed month for the market, with a few Big Tech names performing strongly: MSFT +24.58%, AMZN +13.95%, AAPL +6.76%.

The pain was largely in the AI trade, especially memory names that saw a pretty sudden and large pullback, with the reason for the speed of the pullback explained in the last couple of days, Leopold Aschenbrenner’s margin calls and subsequent forced selling. This was not the only forced selling, as Koreans also saw massive liquidations, with sources revealing that 1.2 million retail accounts hit margin calls, erasing over $2T in market value from the KOSPI since its June peak.

In my view, there were 3 drivers for these mixed results:

The AI CAPEX trade was confusing

In the past year, the rising tide lifted all boats, with every company in the AI stack benefitting from increasing CAPEX, largely fuelled by the hyperscalers. This month was a good tell that things have changed.

Frankly, from my point of view, it was very confusing. If we rank the hyperscalers by their cloud growth numbers during earnings this Q, we get GCP at 82%, Azure at 43% and AWS at 36.7%. However, when we rank them by share price reaction, we get the exact opposite.

The reason was also not CAPEX discipline, as I intuitively thought. Amazon raised full-year CAPEX from roughly $200B to $220B and reported TTM free cash flow of negative $7.6B, yet the stock went up 13%. Google raised its guide by $15B at the midpoint with free cash flow at negative $5.9B and went down 7%.

Moonshot AI released Kimi K3 on 17 July

Moonshot AI unveiled Kimi K3 at the World Artificial Intelligence Conference in Shanghai on 17 July and within hours the market treated it as a second DeepSeek moment.

Kimi K3 is an open-weight Chinese model, similar to DeepSeek. The results of the release was especially surprising, with the model benchmarking within a few points of the frontier, trailing only Claude Fable 5 and GPT-5.6 Sol, and doing it under export controls, which implied the American compute advantage was worth less than the CAPEX assumed. This meant the terminal value of the Western infrastructure buildout had to be marked down structurally.

However, when we look deeper, it is much less worrying for the infrastructure players. Kimi K3 is 2.8 trillion parameters, the largest open model ever released, up from roughly one trillion for K2, and it is priced at $3 and $15 per million tokens, identical to Claude Sonnet 5 and 3-4x what the previous Kimi cost.

4 days after the launch, Moonshot paused new subscriptions because its own GPU capacity had run out. It turned out that Kimi K3 was much bigger, more expensive and compute-constrained than what the market initially thought.

Forced Selling

In my view, this was the most important reason for the magnitude of the sell-off this month. Leopold Aschenbrenner's Situational Awareness ran roughly $225M into as much as $45B in under two years and was up 439% net YTD going into July.

His fund was long AI infrastructure and short software at nearly 4x leverage. In July both legs went the wrong way, with many infrastructure names falling 40-50% while software names rallied. This led to a complete wipeout of capital for the portfolio. Hilariously, Citadel, the very market maker that shifted its outlook on 28th July to predict a surprise Federal Reserve interest rate hike, bought his entire book. Of course, the rate hike didn’t happen, but it was enough to force Leopold to sell out.

This, coupled with the leverage single-stock ETFs that have been discussed about a lot in the Korean market, led to cascading liquidations.