Superintelligence Needs Memory: Micron Reports

Micron Technology reports fiscal fourth-quarter results after the close on September 30, with management already telegraphing a quarter that would have been unthinkable in the 2023 memory trough: about $50 billion of revenue, plus or minus $1 billion, roughly 86 percent gross margin, and $31.00 of non-GAAP earnings per share, plus or minus a dollar. Consensus has drifted slightly above that revenue line. Either way, the company is not asking permission to join the AI capex supercycle. It is closing a fiscal year in which third-quarter operating margin hit 80.4 percent on $41.5 billion of sales.

The easy headline is that superintelligence needs memory. The harder question — the one tonight’s call has to answer — is where Micron ranks among the profit machines that actually supply it.

The league table is memory-first

Strip the conglomerate wrappers and compare operating margins in the May–July window, when Micron’s fiscal third quarter overlapped calendar second-quarter prints from Seoul.

Supplier	Period (approx.)	Revenue	Operating margin
Micron	FQ3 FY26 (ended May 28)	$41.5B	80.4%
SK hynix	CQ2 2026	KRW 79.3T (~$58B)	76%
Samsung (DS / memory)	CQ2 2026	KRW 127.5T DS (~$93B)	~70% on memory
Nvidia	FQ2 FY27 (ended July)	$96.2B	~56% op. (75% GM)

Micron is smaller than SK hynix on revenue in that window and far smaller than Samsung’s device-plus-memory empire. It is also more profitable per dollar sold than either Korean memory peer — and materially more profitable than Nvidia once you move from gross margin to operating income. Nvidia’s own guide already prices memory as a cost problem: Colette Kress cited extreme memory pricing and a 300-basis-point gross-margin reset even as revenue scales toward $108 billion. Micron sits on the other side of that invoice.

The residual is not whether AI needs DRAM. It is whether the marginal dollar of AI profit pools in the GPU vendor or in the three firms that stack HBM.

SK hynix still leads the pure HBM narrative — HBM4 mass shipments, SOCAMM2 on servers, and the tightest coupling to Nvidia’s accelerator cadence. Samsung’s device division lost money in the same quarter its memory unit carried 99.7 percent of company operating profit, which makes the Korean giant a dividend-and-fab story wrapped around a memory core. Micron is the U.S.-listed, pure-play read: no handset hangover, no foundry distraction, just Cloud Memory and Core Data Center units that printed 78–83 percent operating margins in FQ3.

What the guide already claims

Management’s FQ4 outlook implies another step-up from the record third quarter — revenue up roughly 20 percent sequentially, gross margin near 86 percent, and operating expenses still stranded below $2 billion while sales approach fifty billion dollars a quarter. HBM4 is in high-volume shipment for a lead customer; HBM4E on 1-gamma is slated for calendar 2027. The product story is catch-up and co-leadership, not also-ran: Micron spent years trailing SK hynix on HBM yield, then used the AI shortage to re-price the entire DRAM portfolio, not just the stacks that attach to Blackwell-class GPUs.

The durability bet is newer. Micron closed FQ3 talking up multi-year Strategic Customer Agreements — volume and pricing visibility that management argues will smooth what has historically been the most violent cyclical stock in semis. When Seoul poured concrete into Honam, foreigners sold Samsung and SK hynix into the headline. Micron’s contracts are an attempt to keep the margin when the cycle eventually turns, not merely the revenue when it peaks.

Capex is the falsifier

The bear case is not demand. It is supply hitting the balance sheet before it hits bit growth. Micron guided roughly $10 billion of capital spending in FQ4 and about $27 billion for fiscal 2026, with management warning that every quarter of fiscal 2027 exceeds FQ4 capex — more than half the year-on-year increase from construction, not tools alone. Construction does not ship HBM next month. If pricing softens while fabs rise, free cash flow compresses faster than operating margin.

That is the same tension we mapped when Nvidia beat and the tape shrugged: record fundamentals colliding with a market that now asks what happens after the easy comparisons. For Micron, watch three lines on the call — HBM revenue run rate versus SK hynix’s HBM mix, commentary on contract duration under Strategic Customer Agreements, and whether FQ1 FY27 capex still rises when DRAM spot indices plateau.

Closing checksum

Nvidia names the era; Micron and the Koreans tax it. Tonight’s print is less about beating a whisper number than about confirming that an 80 percent operating margin was a quarter, not a glitch — and that the only U.S. equity that offers a clean bet on that tax trades at a fraction of Nvidia’s revenue scale with a higher incremental margin. If guidance holds and contracts extend, the AI trade’s second act is not another GPU launch. It is memory suppliers running profit rates that look like software — on silicon that still has to be baked in Idaho and Yongin.

Superintelligence needs memory. The market is only beginning to price memory like it already prices compute.