Memory's "Worst Is Over" Is a Supply Call, Not a Demand Call

In mid-2026, the two symbol stocks of the AI memory boom did something their earnings never did: they fell. Micron dropped from a ~$1,214 June 25 high to about $830, a decline of roughly a third. SanDisk, the purer NAND play, slid from about $2,335 to near $1,260 over the same stretch. Both were down 25% to 42% from their highs, and the moves read to many retail investors as a bubble finally catching up with a rally that had been running for a year.

The reasons offered were almost entirely demand-side. A July 1 Bloomberg report said Meta was weighing whether to rent out its excess AI compute capacity. South Korea's financial regulator warned about leveraged memory-chip ETFs, and Micron fell more than 13% in one session on the follow-through. Rising bond yields added pressure. When Goldman Sachs then said the worst may be over for Micron and SanDisk, it was answering those worries in kind.

That debate is being held on the wrong axis. The memory rally was not built on demand, and a demand scare is not what will end it. The earnings surge of 2026 is a pricing event: prices have run up because suppliers, who control an effectively fixed stock of silicon, are being bid up by buyers with nowhere else to go. The question that actually decides "is the worst over" is not whether AI spending holds. It is whether and when the supply constraint breaks.

The earnings are price, not volume

The cleanest evidence comes from SanDisk, which Western Digital spun off and which is the closest thing to a pure NAND bet among large U.S. stocks. Its revenue rose 175% for the fiscal year with a gross margin above 71%, an adjusted profit of about $71 per share. The company attributes two-thirds of its revenue growth to higher pricing and the rest to volume.

The price path explains why. SanDisk raised NAND contract prices 50% in November 2025. Research firm TrendForce projected NAND contract prices up 70–75% for the spring quarter before cooling to 10–15% in the current one. Goldman estimated conventional DRAM pricing up about 176% in 2026, and Micron's latest quarter still carried revenue growth near 167% with a gross margin above 72%.

This is the mechanism the consensus story gets backwards. Starting around 2023, the big suppliers moved their capital toward high-bandwidth memory, or HBM, because AI accelerators paid a premium for it and margins were better. That starved conventional DRAM and NAND of new capacity. Then AI shifted from training models to running them — inference — which consumes enormous amounts of storage. Demand for plain memory grew just as the industry had stopped building it, so the makers sold roughly the same silicon at far higher prices.

The behavior confirms the constraint is being managed, not met. Micron paused NAND and DRAM price quotes for about a week in the fall of 2025. SanDisk's November hike forced memory-module makers to halt shipments while they reassessed. And the scarcity shows up in how customers commit: per Goldman, Micron holds about $100 billion of future revenue through sixteen strategic customer agreements in which buyers prepay to secure supply, and its 2026 HBM output is already sold out. When buyers pay upfront for chips that will not exist in sufficient quantity, that is scarcity, not growing unit sales.

Which is why the correction, however violent, did not touch the underlying condition. None of the triggers — Meta's lease ambitions, a regulator's ETF warning, bond yields — changed the fact that Micron, SK Hynix, and Samsung report being sold out of DRAM and HBM into 2027. Goldman's "worst may be over" is a claim that the scare passed while the shortage did not. Goldman's own supply math puts DRAM undersupplied 5.0% in 2026, widening to 5.9% in 2027 — the worst shortfall since a 4.2% deficit in 2017 — with NAND undersupplied 2.5% in 2027.

Two very different ways to own a pricing cycle

For an investor, Micron and SanDisk are not the same bet, and the split matters more as the cycle matures.

	Micron	SanDisk
What it sells	DRAM, HBM, NAND (fab owner)	NAND storage (buys wafers from Kioxia JV)
Market cap	~$1.1 trillion	~$250 billion
Trailing capex	~$25 billion	~$177 million
Free-cash-flow margin	~29%	~57%

Micron is an integrated supplier: it owns its fabs, so it carries the upside of tight HBM and DRAM, but it also carries roughly $25 billion of trailing capital expenditure and the burden of building the capacity that eventually ends the shortage. SanDisk, by contrast, does not own wafer fabs — it buys NAND from Kioxia under joint-venture supply agreements — so its trailing capex is a rounding error and nearly 57 cents of every revenue dollar drops to free cash flow. SanDisk is close to pure price leverage: each point of NAND price flows almost straight through with no fab to fund.

That steelman also defines the fragility. A company with no fabs has no way to defend its supply if prices roll over, and a company whose earnings are all price has no unit growth to soften a decline. SanDisk's pure-play boost cuts both ways.

There is a second split, inside demand. SanDisk's datacenter revenue more than doubled, to roughly $3 billion, while its consumer revenue fell 32% sequentially as buyers balked at record flash prices. The price hikes are being paid by AI data-center buyers; the consumer market is already hitting a ceiling — which is precisely why the current-quarter NAND increase has slowed to 10–15% from the spring's 70%-plus. An upcycle that depends on one class of buyer paying ever-higher prices for the same silicon is an upcycle with a visible lid.

The condition that ends it

The "worst is over" call holds only until supply discipline breaks, and the markers of that breaking are already on the schedule. SK Hynix approved roughly $38 billion for two new memory fabs, spending that analysts immediately flagged as seeding oversupply concerns for 2027 and beyond. ChangXin Memory, China's DRAM maker, lifted its global DRAM share to about 10% via a $8.6 billion IPO; it has begun risk production of HBM3E with mass production possible in 2027. YMTC, China's NAND maker, is preparing its own large IPO. A Samsung adviser has warned a capacity glut could form by late 2027, and analysts point to 2028 as the year oversupply returns.

There is a real disagreement buried inside the bullish consensus, and it is about timing, not direction. Goldman's own view is that the new capacity from Micron, Samsung, and SK Hynix will not meaningfully ramp until 2027 to 2029, which is why it sees undersupply persisting into 2028. The skeptics who see the glut forming by late 2027 are betting the announced investment becomes chips on a shorter clock.

For the retail investor, the useful correction is the one that never made the tape. This was never a unit-demand story, so it should not be valued as one at its peak, and it should not be sold on a demand scare at its trough. The decision rests on the supply side: whether the companies keep exercising discipline and delaying the capacity they have already announced — in which case price, and with it these earnings, holds deeper into the cycle — or whether new capex and Chinese chips arrive first, in which case a price rollover hits earnings before it touches the balance sheets. In a price-driven earnings stream, the margin for error is thin. That is the axis Goldman's reassurance actually rests on.