AI Memory Constraints Persist; Micron Trading at 7x P/E Despite Capex Boom

A remarkable pattern has emerged across Big Tech earnings: in just two days last month, five of the largest firms independently reported that memory is the bottleneck in their AI infrastructure plans. This is not a cyclical comment but a structural constraint tied to the physics of neural network training and inference. Yet Micron Trading at 7x earnings despite having a multi-year visibility into demand from these customers suggests the market has either not priced in the thesis or is doubting that memory makers can capitalize.

The constraint is not DRAM alone but the entire memory ecosystem: HBM (high-bandwidth memory), GDDR, and the interconnect bandwidth between compute and storage. NVIDIA's H200 is explicitly designed to address this by doubling bandwidth compared to H100, but the broader supply chain (Micron, SK Hynix, Samsung) has limited capacity to retool and ramp. CFOs are explicitly budgeting for higher memory costs as part of their capex plans, signaling they expect prices to stay elevated.

Micron's valuation discount to peers suggests three narratives: (1) investors believe memory prices will fall as supply catches up, contradicting CEO guidanceCompany-issued forecasts of future financial performance.; (2) investors are skeptical Micron can execute on capacity ramps; or (3) the market is pricing in cyclical mean reversion despite multi-year visibility. Historical precedent from 2016-2017 memory super-cycles shows that holders of constraint-stage capacity can capture outsized returns if they manage execution.

The counter-thesis is that AI capex itself peaks in 2026-2027, and memory demand will normalize post-peak. Nvidia's public denial of acquisition rumors around PC makers also signals the company is confident in organic growth and not seeking to diversify away from pure chips. If memory constraints persist through 2027, Micron at 7x earnings is materially cheap; if capex peaks and rolls over, it is fairly valued.