What is HBM, and why is AI causing a memory shortage?

A fast AI chip with slow memory is like a chef with a huge kitchen and one small cart bringing ingredients from a distant storeroom. The chef spends most of the shift waiting. Memory bandwidth, how many bytes per second can reach the processor, is often what limits how fast a model answers.

HBM deals with this by moving the storeroom next door. It takes several DRAM chips (the kind of working memory in every computer), stacks them vertically, wires them with a very wide data path (2,048 connections in HBM4, says SK hynix), and places it next to the processor on the same package. SK hynix lists six generations so far: HBM, HBM2, HBM2E, HBM3, HBM3E and HBM4. It says HBM4 doubles the bandwidth of the previous generation and runs at more than 10 Gbps per connection.

How much of it AI systems use

The amounts are large, and growing with each chip generation.

Only three companies supply this memory at scale. Nvidia's annual filing names them as its memory suppliers: SK hynix, Micron and Samsung.

Why HBM squeezes everything else

The key number is the trade ratio. Micron's chief executive said in 2025 that HBM3E consumes three times the silicon of ordinary DDR5 memory for the same number of bits. In August 2026, TrendForce reports, a Micron executive repeated the 3-to-1 figure at an investor conference and said the ratio would get worse, to nearly 4 to 1, with HBM4 and HBM4E.

Here is what that means in practice, using our own rough arithmetic and the conservative 3-to-1 ratio:

Memory factories cannot respond quickly. In a Scientific American interview, the AI policy researcher Lennart Heim said memory makers lack clean-room space, and that the few firms able to build these plants are fully booked. So when demand jumps, price moves first and supply follows much later.

What the companies are reporting

These figures come from the companies' own statements, which are written to reassure investors, so treat the forward-looking parts as forecasts.

Korean coverage says SK hynix described its HBM and other DRAM as sold out through 2029. TrendForce adds a useful qualifier: that picture counts demand still being negotiated on top of signed contracts, so "effectively sold out" is the more careful wording. TrendForce, citing Korean media, also reports SK hynix's average DRAM selling price rose by a high-20s percentage in a single quarter.

Micron's call offers one more signal. As quoted by The Next Platform, its chief executive said agreements for most of its 2027 HBM supply were done at significantly higher prices. Micron no longer breaks out HBM revenue, so any HBM-only figure you see for it is an estimate.

Memory has a history of sharp price cycles, and nobody can say from these filings when this one turns.

Will this make your next computer or phone more expensive?

Quite possibly, though the amount is hard to predict. Heim's view in Scientific American is direct: with memory going to AI chips, prices rise and next year's smartphone costs more. Device makers can absorb some of that, cut other parts, or sell models with less memory than you would otherwise get.

There is already one concrete consumer effect. In December 2025, Micron announced it would leave the Crucial consumer business, the brand many people bought memory upgrades from, with shipments through February 2026. Its stated reason was to better supply larger strategic customers in faster-growing segments, tying the move to AI data centre demand.

This is not a reason to rush a purchase. Prices vary by product and region, and the company outlooks above are not forecasts of what a store will charge. If you do need a machine that will run AI models locally, memory is the one specification that is hard or impossible to upgrade later, so it deserves more of your budget than it used to.