Forget the Artificial Intelligence (AI) Capex Bubble: SpaceX CEO Elon Musk Says the Memory Shortage Will Drive Costs Even Higher

AI hyperscalers are reportedly increasing capex as memory chips become the binding constraint for AI infrastructure: SpaceX CEO Elon Musk said memory output is rising ~20%/year while demand is rising ~200%/year (possibly higher). The article argues this supply-demand imbalance should support continued memory price strength and margin expansion for Micron (MU), framing MU’s sell-off after a late-June all-time high as valuation/positioning noise rather than a demand reversal.

Analysis

This is a near-term pricing-power story, not a clean “AI demand” story. If memory is truly the bottleneck, the first-order beneficiaries are the suppliers with the tightest high-end product mix and the lowest incremental cost of volume; the second-order winners are the equipment vendors that enable capacity expansion, but only after pricing visibility is strong enough to justify orders. The key equity implication is that MU can keep earning upgrades even if broader AI sentiment cools, because the market is buying a shortage, not just a growth narrative.

The consensus risk is that investors treat this as a straight-line capex supercycle. In practice, a memory shortage tends to self-correct: contract prices rise, customers dual-source, and the industry’s usual response is to overbuild with a lag. That means the best window for MU is typically 1-3 quarters before supply catches up; the longer-dated risk is that HBM profitability attracts aggressive capex from SK Hynix/Samsung, which eventually caps margin expansion and compresses the multiple.

For NVDA, the read-through is more nuanced: memory inflation can pressure system BOM costs and delay some deployments, but it also reinforces that AI demand is real. So this is not an obvious short on AI semis; it is more likely a relative-value opportunity where MU’s earnings revisions outrun sentiment while broad semis trade on multiple compression risk if hyperscaler capex growth normalizes. The thesis is falsified if HBM/DRAM pricing stops rising for two consecutive quarters, or if supplier capex announcements point to a faster-than-expected supply response.

Key Decisions for Investors

• Buy MU on pullbacks or into earnings revisions; prefer a staggered entry over chasing strength. Target 1-3 month upside from continued ASP expansion, with the thesis invalidated if HBM pricing flattens or MU guides down gross margin.
• Use a MU / SMH pair trade to isolate the memory bottleneck from broad semiconductor beta. Long MU, short SMH or a basket of non-memory AI winners if you want exposure to pricing power rather than capex sentiment.
• For higher-conviction exposure, consider MU call spreads 3-6 months out instead of outright stock. This captures upside from continued scarcity while limiting damage if the market starts discounting supply normalization.
• Set a watch item on SK Hynix and Samsung capex commentary plus HBM lead times. If those data points turn positive for supply faster than expected, reduce MU exposure immediately.
• Avoid making NVDA a short from this alone; if anything, this is a relative headwind to deployment economics, not a fundamental bear case. A better hedge is broad semis or hyperscaler capex-sensitive names if you want to neutralize AI beta.