What Micron makes, and why AI needs so much of its memory

Micron makes the memory and storage that an AI computer uses to keep information available while its processors work. Its products serve data centres, PCs, phones and other devices. The AI connection is straightforward: an expensive processor cannot do useful calculations with information that has not reached it.[1]

You may know the company through Crucial, its consumer memory and storage brand. In December 2025, Micron announced that Crucial consumer-channel shipments would end in February 2026, while warranty support would continue. It attributed the decision to supplying larger strategic customers amid AI-driven demand. The announcement concerned the branded consumer business, with Micron enterprise products continuing through commercial channels.[5]

Three products with different jobs

Dynamic random-access memory, or DRAM, holds information during active computing. NAND flash retains information when power is removed. A solid-state drive, or SSD, uses flash storage; adding storage capacity does not make it interchangeable with working memory. Both are useful because a computer must retain information and bring the required parts into active use.[3]

A saved model and its training data may live in storage; running the model needs fast access to its parameters and current working information. Micron supplies several places along that route. This is why an AI installation can increase its demand for flash storage, server memory and accelerator memory together.[1]

HBM is a form of DRAM

HBM stacks memory dies vertically and connects them through electrical paths inside the silicon. Micron’s HBM4 12-high product provides 36 gigabytes per stack and more than 2.8 terabytes per second of bandwidth. Capacity measures how much it holds; bandwidth measures data transfer per second. Those are the manufacturer’s component specifications, rather than the speed of a complete AI application.[2]

That distinction reveals a useful detail: Micron’s HBM4 12-high has the same 36-gigabyte capacity as its previous HBM3E 12-high, but more than twice the stated bandwidth. A memory upgrade can therefore improve data movement without increasing the amount of information held in each stack.[2]

A large business beyond one AI component

The filing also recorded about $134 billion allocated to undelivered contract obligations at 3 September. That measure used committed volumes and minimum pricing, excluded certain contracts and was neither revenue already earned nor a revenue forecast.[4]

The manufacturing trade-off matters. Micron’s annual report says HBM requires more wafers and cleanroom space than conventional DRAM for the same number of bits at the same technology node. Capacity devoted to HBM therefore affects the supply of other memory; a reversal towards conventional DRAM could increase supply and put pressure on prices.[4]

A factory opening needs a precise description

Micron’s February 2026 opening in Sanand, India, illustrates another distinction. It is an assembly and test facility that converts DRAM and NAND wafers from the company’s global manufacturing network into finished products. Such a site adds a different production capability from a plant fabricating the wafers themselves. Both stages must function for useful memory products to reach customers.[6]

For readers following AI, Micron makes the physical supply chain easier to understand. Ask what a new product changes: the amount of data held, the rate of transfer, the power required or the number of finished components available. Each improvement can help a different part of an AI system. A faster processor is most useful when the rest of that system can keep it supplied.

Report an error or suggest a correction