MSFT, META, GOOGL CEOs All Flag Memory Constraints on Earnings; MU Still Trades 7x

A remarkable convergence of earnings commentary from the largest tech companies has exposed a glaring valuation mismatch in semiconductor supply. In back-to-back earnings calls in May, CEOs from MSFT, META, GOOGL, AMZN, and AAPL each independently highlighted that memory (both DRAM and NAND) remains the binding constraint in their AI infrastructure buildouts. This is not a cyclical observation; all five signaled that the supply crunch is structural and will persist for months or longer as their data center capex accelerates.

Micron Technology, the dominant DRAM and NAND supplier, reported strong earnings but trades at a valuation multiple that bears no relationship to the structural undersupply narrative being broadcast by its five largest customers. At 7x forward earnings, MU is priced as though memory is a commodity with ample supply. In contrast, other semiconductor names benefiting from the AI buildout (NVDA, AVGO, AMD) command much higher multiples, reflecting the market's implicit assumption that supply is either abundant or immaterial to the long-term investment case.

For investors, this setup suggests either a significant repricing of MU upward, or a capitulation by the mega-cap tech CEOs that memory constraints will ease sooner than they publicly indicated. The first scenario would lift MU and other memory suppliers (SK Hynix, Samsung) in tandem with continued AI infrastructure acceleration. The second would require demand to moderate or new supply to come online faster than expected, both of which would be bearish for the broader AI capex narrative that has driven mega-cap tech to record valuations.

Bears on the memory shortage thesis argue that capacity announcements from SK Hynix, Samsung, and others suggest the shortage is already beginning to ease; if new fab capacity comes online in H2 2026 and 2027, MU's current valuation may actually be defensible. However, the CEOs' comments suggest that even with announced capacity, the delivery timeline is still years out, and in the interim, memory will remain the bottleneck that forces triage of data center expansions and model training.