✦ HBM4 & the Vera Rubin RampNarrow moat

The most important fact about HBM4 is not its speed; it is its order book. Micron entered high-volume production in the first quarter of calendar 2026, a quarter ahead of its own schedule, with the parts designed into Nvidia's Vera Rubin platform¹ — and every unit of 2026 capacity was already committed under multi-year agreements, with 2027 supply being negotiated while the first wafers moved². Commodity products do not sell out a year forward at fixed terms. Enterprise software does.

The product itself earns the queue. The 36-gigabyte, twelve-high stacks Micron began sampling in June 2025 deliver more than 2.8 terabytes a second of bandwidth on pin speeds above 11 gigabits — roughly 2.3 times its HBM3E — while cutting power more than 20%³. In an AI data center where memory bandwidth, not compute, is the binding constraint, that specification is why the buyers signed early.

What the contracts really bought Micron is visibility — the one thing a memory maker has never had. A sold-out 2026 means the supercycle's most profitable year was locked in at negotiated prices before any rival could discount it away. The test comes with the 2027 book: contracts only tame the cycle if their pricing holds when Samsung and SK Hynix's HBM4 capacity lands and supply finally catches demand. Watch the terms of the 2027 agreements as they close, and watch HBM's share of Micron's data-center revenue. If contract pricing slips more than politely, the order book was a symptom of shortage, not a moat — and shortage, in memory, has never once been permanent.