The Price War Begins

What a difference a few weeks makes!

Today markets are buzzing about Situational Awareness, the fund launched by Leopold Aschenbrenner who rose to fame on the back of an essay of the same name.

He was early and right on AI, and parlayed that exposure into a >$20 billion hedge fund. The details and numbers are filtering through so we’ll have to wait for the full story, but it looks like he was running up to 4x exposure. Coming into July, the Wall St Journal reported he was up over 400% year-to-date.

Most recently the fund disclosed large positions in neoclouds like Sharon AI, Coreweave and Nebius, and a major position in the US listing of SK Hynix.

Only a few weeks later, it’s now been reported that the entire portfolio was margin-called and purchased by Citadel.

Many of the largest holdings, down 40-50% in July, promptly bounced over 25% last night. I’m sure we all know that feeling, but not at that scale!

Apparently the deal was for the public side of the book only, and there’s still (among other private investments) $5 billion in Anthropic stock. But even that is not as secure as it looked only recently.

The price war we flagged has finally come: OpenAI cut the price of its latest 5.6 Luna model by 80%:

Which is almost certainly the right strategy to compete head-to-head with Chinese models on their primary selling point. With intelligence this cheap, why bother switching?

Note also the reference to fast mode. OpenAI already offers Cerberus for older models. Speed is the killer application that (we think) will ultimately make current generation obsolete over the next few years.

Is it over?

In our wholesale fund we were selling heavily throughout the month, and are once again at one of those situations like March last year. As I look around there are a number of companies down 50% or more, with no signs of a let-up in capex from the major hyperscaler spenders.

The stars aren’t quite aligned - a major difference from prior episodes is that this time there has been no broader market CTA/momentum unwind, the damage was focused on one part of the market. Key indices like semiconductors and the Nasdaq 100 didn’t even cross the 200 day moving average, one of the better definitions of a bull market. And hyperscaler (and certainly neocloud) CDS spreads have been rising, as the issues I wrote about in my last letter are still being digested.

Speculative length has been significantly reduced in Korea:

And the removal of Korean leverage and at least one heavily leveraged and concentrated US investor could be enough to spark a reversal, particularly given the situation was increasingly made aware to market participants who traded against it, exacerbating the falls.

The AI trade may well now enter a new regime with new winners and losers. Certainly a different set of investors.

With the liquidation of leveraged Korean retail accounts, and this newly reported sale of Situational Awareness’s listed equity positions to Citadel, there is one less player levering up into certain semiconductor stocks. And I’m sure funds and prime brokers alike have lost their appetite for this kind of thing, at least for now.

It’s a good thing we can rely on our quant risk management system to (broadly!) keep us on the right side of these trends.

Meanwhile, there are other parts of the growth space that deserve attention too, even though the action over the last few weeks has all been in semis. Companies like Sea, MercadoLibre, Nubank, Elf, Novo and Transmedics were solid throughout the sell-off.

Regardless of the direction of the next move, there are some clear themes crystallising out:

1. The Nvidia neocloud complex looks shaky, for reasons described in prior letters. The case for caution is strengthened with every new off-balance sheet deal, and these Nvidia equity investments in return for multibillion dollar GPU orders are not done from a position of strength from either party.

2. We still think there is an opportunity in sovereign AI and inference-specialised chips, and similarly:

3. We’re increasingly confident that hyperscalers and major LLM companies will vertically integrate to the maximum possible extent, and put every workflow they can on their own hardware, where they can squeeze out as much efficiency as possible in what have become extremely competitive end markets.

Often a high profile liquidation marks a turning point in markets. We’ll see if that proves the case once again.

Good luck out there.

Michael