The Story
Micron expects memory shortages to continue into 2028 and possibly beyond, and has already sold most of the memory it will produce next year, chief executive Sanjay Mehrotra told investors on the company's fiscal 2026 earnings call.
The company reported adjusted earnings of $33.42 a share for its fiscal fourth quarter against $31.61 expected, with data centre revenue up elevenfold. It guided to around $61.5 billion of revenue and $38.15 adjusted earnings per share for the fiscal first quarter, ahead of analyst expectations of $57 billion and $35.40.
Pricing did most of the work. DRAM bit shipments rose in the mid single digits from the previous quarter while average selling prices rose in the high teens. NAND bit shipments rose about 10 per cent with prices up roughly 30 per cent. Micron's shares have risen more than 500 per cent over the past year.
The constraint is high-bandwidth memory, the stacked DRAM that sits beside AI accelerators from Nvidia and AMD. Mehrotra said HBM demand is growing faster than demand for the DRAM used in servers, and that most of Micron's 2027 HBM output is contracted at prices significantly above 2026 levels. Micron is the only United States-based maker of HBM and said it is working with Nvidia on the industry's first custom HBM implementation.
Because HBM consumes considerably more wafer capacity per bit than conventional DDR5, producing more of it reduces the capacity available for ordinary memory. Cloud Memory margin was flat quarter on quarter, as a larger HBM share offset higher prices.
The squeeze has reached consumer products, with memory costs contributing to higher prices for devices including Apple's iPads and MacBooks. DRAM and NAND production is concentrated among Micron, Samsung Electronics and SK hynix.
Why It Matters
The mechanism by which AI data centres raise the price of a laptop is worth setting out, because it is more specific than general inflation.
High-bandwidth memory is stacked DRAM, and it consumes roughly three times the wafer capacity per bit that conventional DDR5 does. A fab that shifts output towards HBM therefore produces substantially fewer bits of ordinary memory from the same silicon. Since AI accelerators need HBM and will pay almost anything for it, that shift is rational for the manufacturer and ruinous for everyone buying the memory that goes into phones, PCs and ordinary servers.
Micron's own numbers show the pattern cleanly. DRAM bit shipments grew in the mid single digits while prices rose in the high teens. The company is not selling much more memory. It is selling roughly the same memory for considerably more money.
One detail cuts against the simple reading. Cloud Memory margin was flat quarter on quarter because a larger HBM share offset higher prices, which means HBM is currently margin-dilutive relative to conventional DRAM despite being the scarce product. The wafer capacity it consumes costs more than the premium it earns. Micron expects its 2027 contracted prices to close that gap.
Which explains the forward selling. A manufacturer that has committed next year's output has locked in the upside and given away the option of selling into an even tighter market later. That is what a company does when it thinks the current price is good rather than when it thinks prices only go up.
Sanjay Mehrotra, Micron's chief executive, told investors the company has a \"strong roadmap for future HBM products\" and is working with Nvidia on the industry's \"first custom HBM implementation.\"
The Strategic Read
For India the uncomfortable part is that none of this is addressable domestically.
Tata Electronics is building the country's first commercial fab at Dholera for mature and legacy nodes between 28nm and 110nm, aimed at automotive, consumer devices, data storage and AI hardware. That is a sensible place to start and it is not a memory fab, and it is certainly not an HBM fab. Advanced DRAM and HBM capability sits with three companies, and the entry cost is measured in tens of billions of dollars and a decade of process learning.
So the memory squeeze is something India imports in full. Every laptop, handset and server assembled or sold here carries it, and the Dholera fab will not change that when it comes online. Applied Materials' $5 billion commitment and Fujifilm's materials plant are building the tooling and chemicals layer around fabs that do not make memory.
There is a second-order effect worth watching too. Indian data centre operators are building capacity into exactly this market. SB Energy and others have committed billions to facilities whose largest input cost, after power, is the servers inside them. Server DRAM prices rising in the high teens quarter on quarter changes the arithmetic of projects already financed on earlier assumptions.
Micron's own caution deserves repeating in a quarter this good. Memory has always been the most cyclical business in semiconductors, and valuing a memory company on peak earnings has historically been a reliable way to lose money. Selling 2027 output forward reduces the risk of a sudden reversal. It does not abolish the cycle.
For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.