Memory's Rally Isn't an Oracle Story. It's a Supply Story — and It Just Split the Group

Memory chip makers Micron and SanDisk rose again in pre-market trading Friday, Sept. 11, and the easy read is that Oracle's blowout quarter proved AI spending is alive. Oracle did beat: fiscal-first-quarter revenue of $19.3 billion, up 30% year over year, with cloud infrastructure revenue up 121% and more than $30 billion in new AI cloud contracts signed, against a fiscal-2027 capital spending plan of $90 billion to $95 billion. Nothing in that says the AI buildout is slowing.

But an AI demand datapoint is not what moved a group that was already trading at the top of a supply-constrained market. The conventional explanation — new demand, new confirmation — is only half the story, and it is not the operative half. Memory stocks have spent 2026 being repriced on a shortage, not on any single quarter's demand surprise. Oracle confirms demand that the market had already decided exists; the reason these names are up 241% year to date and roughly sixfold from their 52-week low is that the industry has, for a year, refused to add supply.

Pricing is doing the work, not volume

The signature of a supply-discipline-driven cycle is that revenue rises on price, not on shipments. Micron's reported numbers are a clean example. Quarterly revenue has gone from about $8.7 billion roughly a year and a half ago to about $23.9 billion in the most recent quarter; earnings per share went from about $1.9 to about $12.20 over the same stretch. Revenue is up 167% year over year with a gross margin above 70% and an operating margin near 66%. That margin profile is not what unit growth produces. It is what a seller of a scarce good produces when buyers have no alternative and no leverage.

The mechanism behind it is the two forces that define this cycle. First, Samsung, SK Hynix, and Micron spent last year holding capacity back — prioritizing technology migration and letting contract prices climb, a deliberate reversal of the old practice of adding volume. Second, the industry has been reallocating its leading-edge wafer output toward high-bandwidth memory for AI accelerators, pulling capacity away from conventional DRAM and NAND. The result is a market where those who planned ahead for memory are being rewarded, orders are stretching deep into 2027, and price hikes that began this year are accelerating again by August rather than moderating. Memory is no longer behaving like a normal consumer-electronics cycle; it is behaving like an allocation-driven shortage.

The same morning, the group split

If that supply reading is right, it predicts something the Friday headline obscured: the memory trade is not one trade, it is two, and it broke apart the same morning the U.S. names rose. While Micron and SanDisk climbed in New York, SK Hynix fell about 2% in Seoul after a report that DeepSeek's V4.1 Flash model claims to cut high-bandwidth-memory requirements by 75% and SSD needs by 87.5%. The HBM-heavy name sold off on an efficiency scare; the names weighted toward conventional DRAM and NAND — the memory that efficiency improvements have not displaced — held up.

That divergence is the two-market map in miniature. One market, HBM for AI training racks, is led by SK Hynix at roughly 50–55% share and is the most exposed if inference architecture genuinely needs less memory. The other market, conventional DRAM and NAND, is where the pricing leverage sits for Micron and SanDisk, and where DeepSeek's claim has, so far, had no effect. The investor implication is that "memory stocks rose" compresses two exposures with different risks into one sentence. An efficiency breakthrough that touches HBM demand hardest lands on SK Hynix before it reaches a conventional-memory seller — and the September 11 session showed that distinction playing out in real time.

Nothing new was required to lift these names — and that's the risk embedded in that

The uncomfortable fact about Friday's move is that it did not need Oracle to happen. A Goldman note two sessions earlier had already told clients the worst of the memory downturn may be over, and the stocks had been ratcheting higher all week. That should focus attention on what the rally is actually repricing: earnings that are increasingly the product of price, not volume, at a stock around $975 with a trailing P/E near 22 on a $1.1 trillion market cap. Pricing-driven earnings are the most fragile kind — they persist only as long as the shortage does, and they reverse hard when it does not. The July session is a reminder of the volatility: after an IBM warning that customers were shifting spending toward memory chips, SK Hynix, SanDisk, Western Digital, and Micron each fell roughly 8–10% in a single day on what looked like profit-taking, before the move resumed.

The judgment on these names therefore turns on two linked conditions rather than on any single earnings print. The bull case holds if the manufacturers keep capacity constrained and if HBM demand survives the efficiency threat — if a 75% claimed reduction in memory-per-model does not translate into a fall in aggregate memory consumed across a three-to-five-times larger AI fleet. It breaks if suppliers finally add volume into climbing prices, or if DeepSeek's engineering proves out at scale and the AI stack's memory intensity quietly declines. Friday's Oracle beat settles neither question. It only confirmed, at the top of a supply-constrained market, that the demand the scarcity was created to serve is still there.