AI Memory Crisis: MSFT, AMZN, GOOGL Earnings Warn Supply Cannot Keep Up

Within a two-day span in early May, the chief executives of five of the world's largest tech firms sent the same message on their earnings calls: memory is becoming the critical constraint on AI buildout, and relief is not imminent. MSFT CEO Satya Nadella, META CEO Mark Zuckerberg, GOOGL CEO Sundar Pichai, AMZN CEO Andy Jassy, and AAPL CEO Tim Cook each referenced memory shortages explicitly, and in some cases warned that memory costs are inflating faster than they'd anticipated.

The specifics are revealing. AI infrastructure requires two distinct types of high-bandwidth memory: traditional DRAM (used in CPUs and general compute) and specialized HBM (high-bandwidth memory) stacks mounted on AI accelerators like NVIDIA's GPUs. Both are produced by a handful of suppliers, chiefly Samsung, SK Hynix, and Micron. Demand from cloud hyperscalers building out inference infrastructure has far outpaced production capacity.

Micron (MU) is the only US-listed pure-play DRAM and HBM manufacturer with meaningful capacity. Yet the stock trades at roughly 7x trailing earnings, a valuation that the tape suggests implies either consensus skepticism about durability of AI capex or structural overcapacity concerns. By contrast, NVIDIA, which is memory-constrained but not memory-producing, trades at a premium to Micron despite facing supply competition from the very memory suppliers who control Micron's upside.

The narrative tension is significant: if AI capex continues (and all five CEOs signaled it will), memory is the limiting reagent. Micron should be priced like a gatekeeper. Yet short interest and analyst downgrades suggest the street believes memory supply will normalize faster than executives are guiding, or that hyperscalers will eventually shift to custom in-house memory (as they have with chips). The arbitrage between memory demand signals in mega-cap earnings and Micron's depressed multiple is one of the year's most glaring disconnects.