The AI Memory Trap: Why the “Cheapest” Chip Stocks Are the Most Dangerous
The picks-and-shovels names powering the AI build-out look unstoppable. History says the moment to fear them is exactly when they look cheapest.
The short version (read this even if you read nothing else)
Every dollar the tech giants — Microsoft, Amazon, Google, Meta — spend building AI data centers is a dollar of revenue for a cluster of suppliers: memory makers like Micron and SanDisk, connectivity chip designers like Credo and Astera Labs, networking and power companies. The market has fallen in love with these suppliers. Some are up several hundred percent.
Here’s the trap, in one sentence: these stocks don’t depend on how much the giants spend — they depend on how much the giants spend growing. And the giants’ spending mathematically cannot keep accelerating forever.
When the AI build-out stops accelerating — not stops, just stops getting bigger every quarter — these suppliers’ growth collapses from triple digits to flat. That is the same moment the giants’ own cash flow finally recovers (because they’ve stopped ramping spending). So the event that makes a Microsoft shareholder cheer is the event that breaks a Micron shareholder. The “good news” and the “bad news” are the same news.
And the most dangerous names are not the obviously expensive ones. They’re the ones that look cheap — the memory stocks trading at single-digit price-to-earnings ratios. That cheapness is a warning label, not a bargain.
The rest of this piece explains why, with real numbers from real history.
Part 1 — The “second derivative” problem (the core idea)
Think of three different things:
How much the hyperscalers spend (the level). This drives the suppliers’ revenue.
How fast that spending is growing (the first derivative). This drives the suppliers’ revenue growth rate — and therefore their valuation multiple.
Whether that growth is speeding up or slowing down (the second derivative). This drives the suppliers’ stock momentum.
The suppliers’ stocks are priced today as if growth will keep accelerating. So spending can stay enormous — a record level — and these stocks can still get cut in half, because “enormous but flat” means growth equals zero, and a stock priced for acceleration cannot survive zero.
This is why “the data centers are still being built” is not the comfort it sounds like. The build-out doesn’t have to reverse to break these stocks. It only has to stop getting bigger.
The reflexive twist: the AI giants are collectively spending hundreds of billions a year. A large share flows to Nvidia and the supplier basket. So when spending growth flattens, it doesn’t hit one supplier — it hits all of them at once, plus Nvidia, plus the debt-funded “neocloud” GPU landlords that borrowed against it. Owning a basket of these names is not diversification. It’s the same bet, levered.
Part 2 — Two ways these stocks die (and why “cheap” is the trap)
The suppliers split into two camps with opposite kill mechanisms.
Camp A — the commodity cyclicals: memory (Micron, SanDisk). Memory chips (DRAM and NAND) are commodities. When demand outpaces supply, prices and profit margins explode upward. When supply catches up, prices collapse — not soften, collapse. These businesses swing from extraordinary profits to outright losses within a few quarters.
The trap: at the top of the cycle, these stocks show low price-to-earnings ratios — because earnings are at a peak that the market knows won’t last. A single-digit P/E on a memory stock at the top is the market screaming “these earnings are about to fall off a cliff.” The risk isn’t the multiple — it’s that the “E” in P/E evaporates.
Camp B — the high-multiple secular names: connectivity (Credo, Astera Labs, Marvell). These design specialized chips for moving data inside AI racks. Their margins won’t collapse like memory — the products are differentiated intellectual property. But they trade at jaw-dropping valuations — many times annual revenue, and earnings multiples that only make sense if growth stays near triple digits indefinitely. The risk here isn’t the earnings — it’s the multiple, which compresses violently the instant growth merely decelerates.
Same trigger. Opposite mechanism. Both can halve.
A few names sit in safer territory — Broadcom (a large recurring software business cushions the chip cyclicality) and Arista (sticky networking software, real cash flow). And the most fragile is a debt-funded GPU-rental company like CoreWeave, which borrowed heavily against rapidly-depreciating chips — if demand slows, its revenue and its collateral fall together. That’s the canary in this coal mine.
Part 3 — The history is not a guess. It happened. Twice.
People treat “AI is different this time” as a reason the memory cycle won’t turn. Maybe HBM (the premium high-bandwidth memory used in AI chips) is structurally tighter than past cycles — that’s a real, defensible argument. But here is what actually happened to Micron the last two times the cycle rolled, and these are facts from company filings, not models:
The 2018–2019 downturn: - Micron’s revenue fell from $30.4B (fiscal 2018) to $23.4B (fiscal 2019) as memory prices softened. - Earnings per share dropped to about $5.51 — still profitable, but the market had already turned. - The stock bottomed near $32, roughly 1.1 times its tangible book value.
The 2022–2023 downturn (the brutal one): - Revenue halved, from $30.8B (fiscal 2022) to $15.5B (fiscal 2023). - Gross margin went negative — about −9% for the full year, with even worse individual quarters once inventory write-downs hit. A company that had been wildly profitable was suddenly losing money on every chip. - Full-year earnings were about −$5.34 per share — a loss. - The stock fell from roughly $98 into the low $50s, a drawdown around −45% to −50%, and again bottomed near 1.1 to 1.2 times tangible book value.
NAND flash — SanDisk’s entire business — is more commoditized than DRAM. In that same 2023 downturn, the flash operations that became SanDisk ran negative gross margins (around −10%). Commodity NAND has, in deep troughs, traded below tangible book value.
The pattern across both cycles is the single most useful fact in this entire piece: Micron bottoms at roughly 1.0 to 1.3 times tangible book value, with earnings at zero or negative. Not at “10 times trough earnings” — there are no trough earnings to multiply. The floor is the book value of the factories, not a price-to-earnings ratio.
Part 4 — How to actually stress-test these stocks (the method)
Forget peak earnings. Forget the seductive low P/E. Here is the disciplined way to size the downside on a memory name, and you can run it yourself:
Step 1 — Find tangible book value per share. Take total shareholder equity, subtract goodwill and intangible assets, and divide by shares outstanding — any financial data site lists all three. (One caution on SanDisk: it carries several billion dollars of goodwill from an old acquisition, so its tangible book is meaningfully below its stated book. For a cyclical, always use the tangible figure.)
Step 2 — Apply the historical trough multiple. Multiply that tangible book value by the multiple the stock has actually bottomed at in past cycles: roughly 1.1 to 1.3 times for Micron, and lower — about 0.7 to 1.0 times — for a pure NAND name like SanDisk, which has overshot below book in deep troughs. The result is your estimated downside floor.
Step 3 — Assume earnings provide no support. At the trough, earnings are negative. There is no “the P/E will hold it up.” The floor is the balance sheet, full stop.
Step 4 — Compare that floor to today’s price. The gap between the current share price and the book-value floor is your real downside in a genuine down-cycle. The higher the stock has run above book, the larger that gap — which is why a name that has already tripled is not safer for having tripled.
The uncomfortable output of this method: a memory stock that has tripled is not “safer” for having tripled. It now has further to fall to the same historical floor. The book-value floor rises slowly over time as the company retains earnings — but the multiple (1.1 to 1.3 times tangible book) is the constant the cycle keeps returning to.
The one caveat worth taking seriously: if AI demand has structurally raised the floor of this cycle — if HBM’s three-player supply discipline genuinely prevents the usual glut — the next trough could be shallower than 2023’s. That is the bull’s best argument, and it is not crazy. But “shallower than a −50% wipeout” is still a large drawdown, and the burden of proof sits with the people claiming the most cyclical industry in technology has abolished its cycle. It hasn’t yet, in 40 years of trying.
Part 5 — What to watch (so you see it before the crowd)
The inflection will not be announced. It shows up here first, roughly in this order:
The hyperscalers’ capital-spending growth rate in their quarterly updates — not the dollar amount, the rate of change. The first quarter where guided spending growth steps down while the absolute number stays high is the signal.
Memory spot and contract prices. When DRAM and NAND prices stop rising, the memory cycle is peaking — regardless of what management says about being “sold out through next year.” They always say that at the top.
Shortening lead times and falling book-to-bill ratios at the equipment and component suppliers.
“Digestion” language on earnings calls — any hint that customers feel they have enough capacity.
Credit stress at the debt-funded GPU landlords. When their refinancing gets more expensive, the cycle is turning before the chip income statements show it.
The bottom line
The AI suppliers are a leveraged bet on the acceleration of data-center spending, not the level. Spending can stay massive and these stocks can still break.
A low P/E on a memory stock at the top is a warning, not a bargain. The earnings are about to fall; the multiple is “cheap” precisely because the market knows it.
The honest way to size the downside is tangible book value times the historical trough multiple (about 1.1 to 1.3 times for Micron, lower for NAND) — not a multiple of peak earnings, which vanish at the trough.
The whole basket is one correlated trade. The event that finally lets the tech giants generate free cash again — spending growth flattening — is the same event that breaks their suppliers. Plan for both sides of that to happen on the same day.
This is analysis, not investment advice. The historical figures (fiscal 2019 and fiscal 2023) are drawn from company filings; all current prices and multiples are marked for verification before publishing. Cyclical industries reward people who respect the cycle and punish those who declare it over.