The Memory Shortage Behind 2026's GPU Price Increases

What's Actually Happening

CNBC reported in January 2026 that AI memory is effectively sold out, calling it an "unprecedented surge in prices." The three companies that make almost all the world's high-bandwidth memory (Micron, SK Hynix, and Samsung) are all seeing the same thing: demand from AI accelerators has outpaced their ability to build HBM capacity.

The numbers behind that are big. Micron's stock is up 247% over the past year and its net income nearly tripled in its most recent quarter. Samsung expects its own quarterly operating profit to nearly triple. SK Hynix has been considering a US listing as its stock price climbs. When all three major suppliers are seeing that kind of move at the same time, it's not one company's story. It's the whole market repricing around HBM scarcity.

Every wafer a memory maker puts toward HBM is a wafer it isn't putting toward standard DRAM. That tradeoff is why this shortage doesn't stay contained to AI chips. It pushes through the whole memory market.

Why This Hits Some GPUs Harder Than Others

Not every GPU carries the same amount of HBM, so not every GPU carries the same exposure to this shortage. Current-gen chips built for frontier training pack dramatically more memory per GPU than the generation before them.

A B300 or MI355X carries 288GB of HBM3e. That's 3.6x the HBM sitting on an H100. If HBM is the scarce, fast-appreciating input, the GPUs with the most of it are the ones with the most cost exposure when memory prices move. That doesn't mean every current-gen GPU gets more expensive by the same amount. It means the ones with more memory have more room for their price to move when memory does.

What This Means for Your Budget

If you're planning spend around B300, MI355X, or anything else in that memory class, a flat year-over-year cost assumption is probably wrong. The safer approach is to model a higher-than-historical price growth rate for high-memory GPUs specifically, separate from whatever assumption you use for older, lower-memory parts.

That's exactly the kind of scenario our TCO calculator is built for: run your own cluster numbers under a few different price-growth assumptions instead of betting the whole budget on one.