"Compute for Sale" Reality Check

CONFIDENTIAL — AI INFRASTRUCTURE COMMITTEE — DO NOT FORWARD

RE: Q3 Capacity Planning, Post-"Compute for Sale" Reality Check

Team,

We need to talk about what happened to the Philadelphia Semiconductor Index this week, because I don't think everyone in this room has fully internalized what just got confirmed in public markets. Micron down double digits in a session. AMD off nearly 7%. Even Nvidia — the company that has spent three years being immune to gravity — slipped. TSMC raised its own spending guidance and got sold anyway. That last part is the tell. When the foundry says "we're doubling down" and the market says "sell," the market isn't questioning the chip. It's questioning the buyer.

We are the buyer.

Let's be honest about what triggered this, because burying it in a footnote won't make it less true. The Meta Compute unit — a division built to resell excess GPU training capacity — is not a curiosity. It's a confession. You don't build a resale marketplace for something you're worried about running out of. You build it when the closet is full and someone finally admits it out loud. For two years the entire capex arms race ran on a single unexamined premise: demand for compute is effectively infinite, so overbuild now, monetize later, worry never. Meta just told the market it has more than it needs. That's not a rounding error in a footnote about utilization rates. That's the scarcity thesis, the one every DCF on every semiconductor name has been quietly built on, getting shot in the foot by the company that was supposed to be the buyer of last resort.

And it's not the only crack. Cerebras and OpenAI are building custom silicon specifically to route around the Nvidia toll booth. Amazon has been shipping its own accelerators for a while now and clearly isn't stopping. This is the part I need finance to model correctly instead of hand-waving: every hyperscaler in this room has now demonstrated, in production, that it would rather spend engineering hours building its own chip than keep writing nine-figure checks to one vendor. That is not a hedge. That is a strategy. When Samsung posts a profit increase of roughly 1,800% year-over-year and the stock still drops on the news, you are not looking at a market that's skeptical of Samsung. You're looking at a market that has finally stopped grading AI infrastructure names on a curve.

Here's my problem with how we've been talking about our own budget internally. We keep citing the hyperscaler capex commitments — the $750 billion-plus in aggregate spending everyone loves to wave around on earnings calls — as if the number itself is the bull case. It isn't. Spending is not demand. Spending is a bet on demand, and bets can be wrong, and we are now three weeks into the market pricing in the possibility that this particular bet got ahead of itself. Kevin Warsh's Fed hasn't even needed to say a word about rate policy to accelerate this repricing — though don't think for a second that persistent inflation prints and a Fed chair who leans hawkish aren't sitting in the back of every portfolio manager's head as a second reason to trim duration-sensitive growth names right when the AI story wobbles. Two headwinds arriving at once rarely stay separate for long. They compound.

I want to be careful here, because I'm not writing this memo to tell you the AI capex cycle is over. It isn't. TSMC didn't raise guidance because it's nervous — foundries with multi-year backlogs don't do that. Earnings season is tracking a second straight quarter of profit growth above 20% across the index, and that is real, not manufactured by multiple expansion. But there is a difference between "the trend is intact" and "the trend justifies paying any price for any name with an AI ticker attached to it," and the market spent the better part of eighteen months refusing to make that distinction. This week it started making it. Painfully, unevenly, and probably not for the last time.

What I need from each of you before Monday: stop presenting utilization projections that assume our current build-out plan is the floor rather than a peak scenario. If the company that arguably has the best visibility into real-world GPU demand is now advertising leftover capacity to the open market, we owe our own board a version of this deck that doesn't quietly assume we're the exception. Maybe we are. But "maybe" used to not be a word anyone bothered including in these decks, and its sudden reappearance this week is the whole story.

One more thing, and I say this without editorializing on markets generally: when a man who has spent sixty years buying wonderful businesses at fair prices goes on television to say the tape currently rewards gambling over investing, that is not noise to filter out. That is a data point from someone with no capex budget to defend and no vendor relationship to protect. Weigh it accordingly.

More Monday.