Thursday's tape, and why nobody in this building should sleep well tonight

TO: Risk Committee
RE: Thursday's tape, and why nobody in this building should sleep well tonight
STATUS: Not for distribution. Obviously it will be distributed.

I want to talk about the sentence "excess data center capacity" for a minute, because I don't think this committee has priced what those four words did to the entire capital stack we've built our thesis on.

Micron dropped 13% in a single session. That's not a pullback, that's $138 billion of market cap evaporating between the open and close on a stock that was up 884% over the trailing year — a stock we all, at some point in the last six months, described in an internal note using the word "inevitable." Intel went down 9%. AMD went down 7%. The SMH fell 5% coming off a 71% quarterly gain, which means the ETF that was supposed to be our hedge against picking the wrong horse in this race just proved that when the race ends, everyone loses together. That's the part I keep coming back to. We diversified across the AI infrastructure trade like diversification was going to save us from a trade that was never actually diversified in the first place. It was one trade wearing eleven tickers.

And here's the mechanism, since somebody on this desk is going to ask "what actually happened" in the 4pm call, so let me save us the theater. SK Hynix — a company that just three weeks ago was announcing a $64 billion domestic buildout of memory and packaging capacity, a number so large it was supposed to be reassuring — is now reportedly slow-walking its high-bandwidth memory expansion and eyeing a pivot back toward commodity DRAM. HBM is the component that makes Nvidia's chips actually function at the throughput hyperscalers have modeled into their five-year capex assumptions. If the company making the memory is quietly hedging against its own build-out, that is not noise. That is the single best-positioned player in the supply chain telling you, in the only language public companies are legally permitted to use, that they don't believe the demand curve either.

Meanwhile the Kospi did its now-familiar routine — down 10% intraday, circuit breaker, twenty-minute cooling period, the whole seizure — with Samsung and SK Hynix both down more than 12% at the low. I want everyone on this desk to notice that this is not the first time this quarter the Kospi has done this. It did an 8.2% drop in June and clawed almost all of it back the next day. Markets that oscillate this violently on this short a cycle aren't discovering new information every time. They're a crowd that already knows it's standing on a trapdoor and keeps testing whether the hinges hold.

Now overlay the Meta story, because this is the one that should actually keep you up. A report that Meta is exploring selling off excess data center capacity as a cloud business is, on its face, a rounding-error headline. In practice it's the single most bearish sentence anyone has said publicly about the AI infrastructure buildout all year, because it implies the thing this entire trade cannot survive contact with: that the hyperscalers built more than they need. Corning, which sells the fiber that connects all of this, fell 13.6% on the read-through. Vertiv, which builds the data centers, fell 7%. Caterpillar — Caterpillar, a company that makes literal earth-moving equipment — fell 7% because someone modeled fewer data centers getting built. When the excavator company is trading on hyperscaler capex sentiment, we have officially financialized a hole in the ground.

I'm not telling this committee to blow out of the thesis. I'm telling this committee that we have spent eighteen months treating every one of these air pockets — Broadcom's guidance miss, the smartphone demand collapse IDC flagged, now this — as a buying opportunity in a structurally intact story, and at some point that stops being conviction and starts being the sunk-cost fallacy wearing a research report as a costume. Every one of those air pockets has had a real, specific, falsifiable cause. None of them have been "the market panicked for no reason." The market has had a reason every single time. We've just decided the reason didn't count.

Warsh isn't helping. A Fed chair who spent Wednesday in Sintra insisting prices are "too high," a day before a payrolls miss that should have argued for easing, is a Fed chair who is not going to bail out an overbuilt capex cycle with a rate cut on any timeline this committee can underwrite. If the demand-side cracks in memory are real and the Fed is not cutting, the two support beams holding up this valuation — infinite AI demand and falling discount rates — are both wobbling in the same quarter. That has not happened yet this cycle. It is happening now.

Recommend: cut position sizing in the memory names by a third before Monday's open, keep the Nvidia and TSMC core exposure, and somebody please pull SK Hynix's actual capex disclosures instead of relying on the wire summary, because if that HBM slowdown is confirmed rather than rumored, the second leg down happens on a Tuesday morning before this committee has finished its coffee.