I’m long this complex. Long the equipment names, long the memory names, long through leveraged vehicles that punish me twice whenever I’m wrong. So discount everything below by whatever factor you apply to a man defending his own book. I’m not going to argue the chips are cheap, either. After the half they just had, that argument doesn’t survive contact with a price screen.
The claim I’m defending is narrower than that. It’s about what Friday actually said.
Nobody sold SanDisk down eleven percent because they’ve concluded Alphabet’s CFO can’t read a demand curve. Not partly. Not in some diluted, half-conscious way. The competence of the buyers was never the question on the tape, and reading the selloff as a verdict on their judgment is the fastest way to misunderstand what got repriced.
Ask Which Direction the Number Moved
Here’s the test that settles it. If the market had decided AI infrastructure demand was rolling over, the trigger would have been a capex cut. Somebody would have guided down. Somebody would have said the buildout is ahead of the revenue and we’re pausing.
Nobody said that. The opposite happened. Alphabet reported Wednesday night with revenue up twenty-four percent, cloud up eighty-two percent, a backlog north of half a trillion dollars, and quarterly capital spending of $44.9 billion — double last year. Then it raised full-year capex guidance from a $180 to $190 billion range to $195 to $205 billion, and told everyone 2027 would be higher still.
The customer said it wants more. Two days later its suppliers were the worst-performing assets on three continents. SanDisk closed at $1,436.56, down almost eleven percent. Kioxia lost nine and a half in Tokyo. The KOSPI dropped six percent. Whatever that is, it isn’t the market concluding that nobody’s going to buy the memory.
The Servers Are Already Paid For
The second test is even less forgiving. Ask who’s holding the cash.
In June, Alphabet sold roughly $49.6 billion of equity and $20.3 billion of senior notes, roughly doubling its long-term debt, and said plainly that the money was for AI infrastructure and compute. Amazon has raised tens of billions in the bond market this year for the same purpose. Meta has locked in something on the order of $238 billion in non-cancelable purchase obligations across GPUs, capacity, and energy, and stopped buying back its own stock to do it.
A company that has already sold the paper does not cancel the order the paper was raised to buy. That’s not optimism, it’s mechanics. The 2026 order book isn’t a forecast anybody has to believe in — it was settled in the credit markets months ago, and the wafers are committed. Every dollar of that $205 billion leaves Alphabet as a cost and arrives somewhere else as revenue: at TSMC, at Micron, at SK Hynix, at the packaging line, at the transformer manufacturer with a two-year lead time. That’s the part the capex-panic framing keeps losing. Capex is a transfer. It’s a cost on one income statement and the top line on another, and the market spent Friday selling the side that collects.
So What Actually Got Repriced
Rates and crowding, mostly, wearing a capex costume.
The odds of a July hike tripled in a week. The ten-year sat near 4.68 percent. Crude was still above ninety after Thursday’s move. New tariffs are landing on exactly the supply chains that run through Seoul and Taipei. And Bank of America’s sentiment gauge hit its most extreme bullish reading since 2021, which is the market’s way of saying everyone who wanted to own this was already in it.
Look at the internals rather than the headline. The Dow finished Friday up almost half a percent. The S&P closed green by a hair. The Nasdaq lost two-thirds of a percent and the technology sector dropped 2.37 percent. That’s not a market that lost faith in the economy. That’s a market moving out of the most crowded, most operationally geared, longest-duration thing it owns and into things that pay it sooner.
Intel is the cleanest read of the day. It printed its best quarterly beat in years and fell six and a half percent. When a beat gets sold, the argument isn’t about the number. It’s about what you were paying for the number. Memory just happens to be the highest-beta way to own that argument, so memory takes the first and largest hit every single time — which is also why the week didn’t feel stable so much as postponed. The Alphabet print knocked seven percent off Google on Thursday and the oil move took the rest. The memory names simply held out one day longer.
The Part I Won’t Wave Away
The bears have one argument that doesn’t depend on anyone being stupid, and I think it’s correct.
It isn’t that the spending stops. It’s that it doesn’t stop, and the income statements find out on a delay. The four largest hyperscalers bought something like $434 billion of property and equipment in the four quarters through March while recognizing roughly $149 billion of depreciation against it. Two-thirds of that spend is on short-lived assets — GPUs on three-to-five-year schedules, obsolete before they’re fully written down. That gap doesn’t care what AI revenue does in the meantime. It arrives in 2027 and 2028 by the arithmetic of accounting, and the market has not priced it.
And there’s the harder version. One of Google’s founders has said he’d rather go bankrupt than lose this race. That is not a return-on-capital statement. It’s a survival statement, and survival statements are how arms races get funded. Every one of these executives can be individually rational — spending exactly what a rational competitor should spend given what the others are spending — and the aggregate can still be an overbuild. That’s what an arms race is. So no, I don’t think the market believes these people don’t know what they’re doing. I think a slice of it believes they know exactly what they’re doing and that the sum of their correct decisions is still too many data centers.
Fine. Believe that. It’s a defensible position and I hold a version of it myself for 2028.
What I reject is the idea that Friday had anything to do with it. A depreciation wall two years out doesn’t get discovered on a Friday afternoon in July by way of an eleven-percent gap in a NAND stock, on the same week the buyer raised its order. Nothing about the long-run capital-allocation question changed between Thursday’s close and Friday’s.
Friday repriced a risk. It didn’t discover one.