Most people holding Micron stock can't clearly explain how it makes money.

Last quarter, Micron, as the primary HBM chip manufacturer based in the United States, had a single-quarter revenue of $41.5 billion, higher than the annual revenue of many semiconductor companies.

One year ago, in the same quarter, Micron's quarterly revenue was $9.3 billion. After four quarters, it became $41.5 billion, an increase of 346%.

Where does the money come from? Let's break it down into three layers.

Layer 1: DRAM, $31.3 billion, accounting for 76% of total revenue.

DRAM is memory that clears when power is lost, found in computers, phones, and servers. Micron has always relied on this, but previously profit margins were razor-thin, with prices fluctuating wildly with the market. In good years, they made money; in bad years, they lost so much they questioned their existence.

But now it's different. A traditional server is equipped with a few hundred GB of DRAM, while an AI server with NVIDIA Blackwell GPUs requires 5 to 8 times the DRAM of a regular server.

NVIDIA's next-generation platform, Vera Rubin, has not yet been deployed at scale. Once it is, the memory demand per machine will climb to the next level.

This year, the combined capital expenditures on AI data centers from Amazon, Microsoft, Meta, and Google exceed $725 billion. This money ultimately flows into GPUs, networking, cooling, and storage.

DRAM is something that every server cannot do without.

Layer 2: NAND, $9.9 billion, accounting for 24% of total revenue.

NAND retains data when power is off; it's used in SSDs and phone flash memory. This category used to be even tougher, because Samsung's capacity was so massive that prices were kept low for a long time.

However, now the volume of data produced by AI is astronomical and requires storage.

Micron's enterprise SSDs sold $5 billion in a single quarter, which are data center-specific storage, and also the highest quarterly figure ever for this category. Its growth logic is the same as DRAM's: it's not driven by the consumer market, but by AI infrastructure.

Layer 3: HBM, which falls under DRAM but is the core of profitability.

HBM is a completely different architecture. Ordinary DRAM is laid flat on the circuit board, while HBM stacks multiple chips vertically, connected with copper through-silicon vias, and then attached directly next to the GPU, with bandwidth dozens of times higher than ordinary DRAM.

Every NVIDIA AI GPU must be paired with HBM; without it, the GPU cannot run.

Micron's HBM4 shipments this quarter have already exceeded $1 billion, representing the first mass-production deliveries customized for NVIDIA's Vera Rubin platform.

On the same wafer, the selling price of HBM is 5 to 10 times that of ordinary DRAM. That's why Micron's gross margin this quarter reached 84.9%, nearly double that of Apple.

All three product lines are booming simultaneously, with the driving force coming from the same place: every AI data center and every AI server is paying Micron.

And in the short term, nothing can replace it.