Tepper Cashed Out of the Memory Trade at the Top. His Rotation Maps the AI Cycle's Next Leg.

Tepper Cashed Out of the Memory Trade at the Top. His Rotation Maps the AI Cycle's Next Leg.

David Tepper's fund just sold the best-performing stock in the S&P 500. In its second-quarter filing, Appaloosa Management exited SanDisk entirely — the pure-play NAND flash maker that was up roughly 591% by the end of the second quarter — and cut its Micron position by more than 40% even though Micron remained its second-largest holding. The man who rode the 2026 memory mania all the way up walked away from it before July, when the stock collapsed.

Most coverage will frame this as a billionaire's stock tip — sell SanDisk, buy whatever he bought. I read it as a market-structure signal. Tepper did not sell artificial intelligence. He sold the last leg of the training-era bottleneck — memory — and rotated the proceeds into the layers of the stack that get paid in the next leg: the foundry, the inference cloud, and electricity. That is a map of where this cycle stands, and it is worth decoding on its own terms rather than as a shopping list.

The memory trade's problem is that it was never a moat

The exit at the top is not the only sign that memory peaked around the middle of 2026. Look at what SanDisk just reported. In its fiscal fourth quarter, which ended in early July, revenue came in near $9 billion — nearly three times the $3 billion of the quarter before — with earnings per share jumping from about $6 to roughly $39. The trailing-twelve-month figures are even more striking: revenue up 175% year over year, a 71.5% gross margin, a 61% operating margin, and a free cash flow margin near 57%. Those are not commodity-memory numbers; they are the numbers of a company printing money from a supply shortage.

And yet the stock crashed. SanDisk peaked above $2,300 in mid-June and by the end of July had fallen about 54% from its intraday high — on top of that blowout quarter. That is the market refusing to extrapolate. The debate now framing the trade puts the conflict in plain terms: the market is pricing a 30% margin against management's 75%. Investors are betting the windfall reverts; management says this NAND cycle is structurally different.

That skepticism makes sense. Memory is the corner of the semiconductor stack where the upcycle has always ended in oversupply, because new capacity responds to price with a lag. The price spike itself is the mechanism that funds the next supply wave. SanDisk's own balance sheet shows the tell: free cash flow of roughly $11.5 billion over the trailing year against just $177 million of capital expenditure. A company with a true structural scarcity would not leave the capacity buildout to the rest of the industry.

Line the five most relevant exposures side by side and the divergence is visible in a single glance:

Metric (trailing twelve months)	SanDisk	Micron	TSMC	Nvidia	CoreWeave
Revenue growth YoY	+175%	+167%	+24%	+71%	+115%
Revenue growth QoQ	+51%	+74%	+13%	+20%	+24%
Gross margin	71.5%	72.6%	64.4%	74.2%	69.4%
Operating margin	61.2%	65.6%	56.3%	64.0%	-2.6%
TTM capex	-$0.2B	-$25.3B	-$46.2B	-$6.6B	-$20.6B
Price/sales (TTM)	11.5x	12.1x	16.0x	20.5x	6.4x

Read it this way: the memory makers are showing the best margins of the group — record margins on a commodity — yet the market pays them the most modest price-to-sales multiples of the group, roughly 11x to 12x, because it does not trust the earnings to last. TSMC earns a higher multiple at 16x sales precisely because it spends on the future — $46.2 billion of capex over the last year and guidance between $52 billion and $56 billion in capital spending for 2026. Nvidia's 20x sales prices durable compute leadership. CoreWeave's 6.4x sales compensates for negative free cash flow and the heaviest debt load in the group.

Where the money went: the layers that get paid no matter which chip wins

The same logic sat underneath Tepper's redeployment. He lifted TSMC by 322,500 shares to a stake worth about $788 million in the $2.17-trillion foundry that sits under every serious accelerator program, Nvidia's included. He topped up Nvidia by another 53,500 shares. He opened a brand-new CoreWeave position worth about $107 million — the GPU-cloud and inference operator whose revenue grew 115% year over year, whose roughly $29.5 billion of net debt is the price of admission in a capital-intensive market, and whose reported $104 billion revenue backlog is the asset that backs it. He pushed Amazon to a roughly $1.19 billion position, his largest, and added the power producers Vistra and NRG, both beaten down hard this year.

Every one of those destinations shares a property SanDisk lacks: they earn revenue regardless of which chip architecture wins the inferencing battle.

TSMC is the cleanest expression of the transition I keep pointing to. In an up-cycle, the durable shortages are the ones money cannot cure quickly. Foundry capacity is one of them; TSMC's management has called advanced packaging capacity "extremely tight," which is why advanced chips now account for roughly 75% of its wafer revenue and why its results keep setting records even as the market argues about whether AI capex is slowing. Electricity is the other. You cannot spin up a power plant in a quarter, and the Energy Department now projects data centers could reach 12% of total U.S. electrical demand by 2028 — which is why a fund that spent the whole boom in silicon is now paying up for independent power producers at below 16x forward earnings.

For Tepper, the sequencing is the story: he took profits in the most speculative part of the trade and put them into the parts with the most boring, most durable claim on the AI buildout. Even the small new $38.5 million position in SpaceX reads the same way — a bet on the next expansion of compute, sized small enough to be optional.

The honest counterargument: this time really might be different

The memory bulls have a real case. Micron has signed 16 strategic customer agreements carrying roughly $100 billion of remaining performance obligations, with floor prices designed to hold margins above past cycle peaks. SanDisk says multiyear agreements now cover about half of its capacity, and it guides toward 75% to 80% gross margins through 2028–2030. Industry estimates point to NAND supply constraints expected to persist into the second half of 2027. If buyers are locking in three years of supply at today's prices, the crash-and-burn that ended every prior memory cycle may genuinely be muted this time.

Contractualization reduces the amplitude of the next downcycle. It does not repeal the supply response that a price spike of this size funds — and the reason the multiyear contracts exist at all is that buyers are terrified of running out, which is precisely the fear that tells producers to build more. The market's reaction to SanDisk's blowout quarter — selling the stock 54% off its high on the best results in the company's history — tells me it believes those earnings are peak-cycle even with the contracts in place.

Where the capital goes from here

I don't take a billionaire's quarterly filing as investment advice. A 13F is a two-month-old snapshot of where someone else's money already was; I deal in facts, not in following anyone's picks — Tepper would likely say the same about anyone copying him.

What the filing does is line up with the cycle read, and the cycle read is what matters. The memory leg of the AI trade did its job: the easy return — a year and a half out of a $46 spinoff into a roughly 35-bagger near $1,600 — is mostly realized. The remaining argument is about whether the windfall persists, which is a narrower bet with a less favorable payoff than the one that already paid. In my opinion the marginal AI dollar is better deployed owning the infrastructure layers than chasing a memory windfall the market already refuses to believe. Carry risk in the names with the architectural position; do not carry it in a pricing cycle.

What would break my read is specific. If NAND pricing holds through 2027 and SanDisk sustains gross margins anywhere near 70%, then the market's 30%-margin assumption was the error, management's 75% guide is right, and the memory trade gets a second wind. The trade will decide itself in that gap between the margin the market is paying for and the margin management says is structural — and Tepper, the man with the asymmetrical payoff on both sides, chose to ring the register rather than wait for the answer.