The stock is down from $461.7 to $380.7 as of March 25, 2026. That’s not a crash — it’s a 17% pullback from a peak hit less than three months ago. But the price action is doing something interesting: it’s telling a different story than the income statement, and the gap between those two things is where the actual trade lives.
Here’s the primary signal. Micron’s capex-to-revenue ratio jumped from 33.4% in FY2024 to 42.4% in FY2025. That’s not incremental. That’s a company making a structural bet that the window to lock in AI memory capacity is short, and that being late is worse than being overleveraged. Revenue hit $37.4 billion in FY2025 — up 48.9% from $25.1 billion the year before — and instead of pocketing the windfall, Micron funneled $15.9 billion of it back into capex. Operating income recovered sharply, from $1.3 billion in FY2024 to $9.8 billion in FY2025, so there’s cash to burn. They’re choosing to burn it.
The question is whether that’s a rational capital allocation decision or a fear-driven one. Those are different things with different outcomes.
SK Hynix just committed $8 billion to top-end ASML equipment. That’s the move Micron is responding to. The HBM3E market — the architecture that makes AI inference at scale even possible — is being carved up right now, and whoever gets the yields, the supply agreements, and the installed base locked in over the next 18 months probably holds that position for years. This isn’t like DRAM commodity cycling where you wait out a downturn and come back. The AI infrastructure buildout is creating sticky relationships between chip designers, memory suppliers, and hyperscalers. Micron knows that. The capex number says it loud.
What’s worth flagging is the R&D line. Micron’s R&D spending as a share of revenue dropped from 13.7% in FY2024 to 10.2% in FY2025. In dollar terms, R&D came in at $3.8 billion against $15.9 billion in capex. That ratio — roughly 1:4 — tells you what phase of competition this is. It’s not about invention right now. It’s about throughput, packaging, and physical manufacturing scale. The foundational research either already happened or it’s being deferred. The bet is on execution, not discovery.
That makes this structural, not cyclical. A cyclical capex story would look like a company catching up to demand it missed. This looks more like a company trying to define the cost curve before a competitor does.
The stock’s three-month range — $284.8 at the January 2026 low, $461.7 at the February 2026 high, $380.7 on March 25 — maps onto something real. January was late-cycle DRAM anxiety, Fed noise, and general tech de-risking. The February spike to $435.8 came as AI infrastructure spend from hyperscalers got re-confirmed in earnings calls. The pullback since then isn’t macro — the Nasdaq Composite on March 25 sat at 21,931.1, barely off its three-month range floor of 21,522.8, so broad market pressure exists but it’s moderate. Micron’s drop from $461.7 to $380.7 is more specific. It’s investors pricing in execution risk on a capex cycle this large, with this much visibility into what could go wrong.
SK Hynix is also reportedly weighing a US share listing this year. If that happens, institutional money inside the memory segment doesn’t grow — it reallocates. Some of what’s currently in Micron finds a closer proxy to the HBM supply chain leader. That’s not a thesis-killer for Micron, but it’s a liquidity variable that doesn’t show up in any of the financial ratios.
The one variable that flips everything here is yield efficiency on HBM3E production. Specifically, wafer-level packaging yields — how many usable chips come off a wafer after stacking and integration. SK Hynix’s $8 billion ASML purchase buys equipment. It doesn’t buy yield. Neither does Micron’s $15.9 billion capex commitment. Both companies are placing enormous capital bets on a manufacturing outcome that the public market cannot see in real time. If Micron achieves a structural yield advantage — even a modest one, 5-8% better per wafer — the cost per HBM3E unit shifts enough to change the competitive dynamic materially. If SK Hynix gets there first, Micron’s capex this cycle looks like it bought capacity that can’t compete on margin. The weakest assumption running through this entire thesis is that HBM3E demand stays linear from here; if hyperscaler buildouts plateau or shift architecturally, both companies are overbuilt regardless of yield. The financial statements won’t show which way this is going for at least two to three quarters.
The revenue surge is real, the operating income recovery is real, and the capex commitment is clearly intentional. But the stock sitting $81 below its February peak while the income statement looks this good is the market saying: we see the bet, we’re just not sure you can cover the spread. Worth noting — every time this market has discounted a capex-heavy memory leader mid-cycle over the past decade, it’s been wrong within 12 months more often than it’s been right.
Two companies are spending billions to find out who can turn silicon into something a computer can think with — and neither one will tell you if it’s actually working until it’s too late to change your position.
Tags: Micron, MU, SK Hynix, HBM3E, AI Memory