Every AI boom needs a bottleneck, and for the past two years one of the tightest has been a type of memory chip most people have never heard of. High bandwidth memory, usually shortened to HBM, sits right next to the processors that train and run artificial intelligence models. Without it, the most expensive chips in the world would spend much of their time waiting for data. That is why memory makers such as SK Hynix, Samsung and Micron have gone from cyclical afterthoughts to some of the most closely watched companies in the market.
What HBM actually does
A computer chip is only as fast as the data it can get its hands on. Ordinary memory sits some distance from the processor and sends data along a relatively narrow path. That is fine for a laptop. It is not fine for an AI accelerator that needs to shift enormous amounts of data every second.
HBM solves this by stacking several memory chips on top of each other, connecting them vertically and placing the whole stack right beside the processor. The result is a much wider and shorter road for data to travel. Think of it as replacing a country lane with a twelve-lane motorway that runs straight into the factory gate.
Why it is so hard to make
Stacking chips is fiddly. Each layer has to be thinned, drilled with tiny vertical connections and bonded precisely, and if one layer fails the whole stack can be wasted. That means production capacity is limited and takes years to expand. Only a handful of companies can make HBM at scale, and customers tend to qualify each supplier carefully before buying.
Limited supply and surging demand is the classic recipe for high prices and fat margins. It is also why memory makers have begun signing long-term agreements with customers, sometimes saying their HBM output is effectively sold out years ahead.
Why traders care
Memory has traditionally been one of the most boom-and-bust parts of the chip industry. Prices rise, makers build too much capacity, prices crash. If you want the background, our guide to the memory chip cycle explains how that pattern usually plays out.
HBM has changed the conversation because it ties memory demand to AI spending rather than to smartphone and PC sales. When a big memory maker reports record results, it moves its home market sharply. South Korea’s index, for example, is heavily weighted to chipmakers, which is why our guide to the KOSPI is worth reading alongside this one.
The flip side is that HBM demand depends on a small group of very large buyers. If the companies building data centres slow their spending, memory makers would feel it quickly. That is why traders watch the spending plans of the big cloud companies, often called hyperscalers, as closely as the chipmakers’ own results.
What to watch
For anyone following the theme, a few signals matter most. Watch what memory makers say about how far ahead their output is booked. Watch capital spending guidance from the largest cloud and AI companies. And watch whether new capacity coming online starts to outpace demand, because that is usually how memory booms end.
It is also worth remembering that a sold-out order book today does not guarantee pricing tomorrow. Contracts can be renegotiated, rivals can catch up and technology moves on to the next generation of chips.
The bottom line
High bandwidth memory is the plumbing that lets AI chips work at full speed. It is hard to make, in short supply and currently very profitable, which has made it one of the clearest ways the AI boom shows up in company earnings. Understanding it helps explain why markets such as Seoul can move so violently on a single set of results.
Themes like this can feel like easy money until the cycle turns. If you want an honest read on how you handle concentrated, fast-moving trades, the free Samuel & Co Trading Trader Assessment is a good place to start.
