What HBM actually is
High-bandwidth memory, or HBM, is a type of DRAM, short for dynamic random-access memory, that is stacked in layers and connected to a processor through a much wider path than ordinary memory chips. Instead of sitting on a circuit board a few centimeters away, HBM sits directly beside the graphics accelerator, sometimes on the same silicon package, and delivers data across thousands of parallel connections at once.
The point of all that stacking and proximity is bandwidth. A modern GPU performs an enormous number of calculations in a single clock cycle, and if it has to wait for data to arrive from memory, those calculations sit idle. HBM widens the pipe so the processor spends less time waiting and more time computing. That is why the acronym has become shorthand for the entire memory layer of the AI hardware stack.
Why every new GPU generation carries more
The defining trend of the last several years is that each new GPU generation carries more HBM than the last. The reason is straightforward: the models that run on these chips keep growing, and a larger model needs more fast memory to hold its weights and intermediate results. A training run that once fit in a few gigabytes of memory now spans terabytes, and the only way to feed a processor that fast is to stack more HBM beside it.
This is not a subtle trend. NVIDIA still controls roughly 80% of the AI-accelerator market, and its roadmap has pushed memory capacity and bandwidth higher with every release. The memory suppliers have ridden that same wave, because every GPU sold carries memory that someone else had to design and manufacture.
The three companies that make it
HBM is not a product you can buy from just anyone. The market is a three-way oligopoly: Micron Technology (MU), SK Hynix, and Samsung. These three firms control essentially the entire DRAM and HBM supply chain, and that concentration is one reason the memory business carries so much cyclical weight.
Micron is the only major US-based manufacturer of DRAM and HBM, vertically integrated from chip design through fabrication. SK Hynix is the pure-play rival that led the early HBM generations and has been a frequent NVIDIA partner. Samsung is the diversified third leg, with a memory division attached to a much larger electronics and foundry business.
Whoever builds the accelerator, one of these three supplies the memory beside it. That is the structural fact behind the whole 10X Project teardown, which resolves the pitch’s “silent partner” to Micron.
Where the growth is heading
Industry estimates the pitch leans on put the HBM market at roughly $33 billion by 2027, up from a much smaller base just a few years ago. The current margin story is the transition from HBM3e, the standard shipping today, to HBM4, which is already being sampled. Each generation carries more capacity and higher bandwidth, and each generation also carries a higher price, which is where the margin expansion comes from. Our explainer on HBM3e memory walks through that transition in more detail.
The cyclical catch
The honest caveat is that memory is a cyclical business, and HBM does not escape that history. DRAM has spent decades swinging between oversupply and shortage as the three competitors ramp capacity, undercut each other on price, and then pull back. When all three expand at once, prices fall; when demand outruns supply, prices spike.
HBM’s stacking and packaging make it more complicated to build than standard DRAM, which has so far helped keep supply tighter and pricing stronger. But the underlying business is still the same three firms with the same incentive to add capacity, and the boom-and-bust pattern has not been repealed.
HBM versus the memory it replaced
Before HBM, the memory beside a GPU was ordinary DDR-style DRAM mounted on the board, and it quickly became the bottleneck as accelerators grew hungrier. HBM solved that by moving the memory physically closer and widening the connection, at the cost of a more expensive and more complicated packaging process. That cost is why HBM is reserved for the highest-bandwidth workloads, and why only three firms have the process to build it at scale.
What it means for investors
For an investor, GPU HBM is a way to think about the memory chokepoint inside the AI buildout without buying the accelerator makers themselves. The trend is real: every GPU generation consumes more HBM, and the market is consolidating around three suppliers with real pricing power. The risk is the cycle, plus the concentration itself, because a downturn in AI demand hits all three at once.
The useful frame is to treat HBM as a growth story riding on a cyclical industry, and to size the position accordingly. The pitch that points you at Micron is directionally reasonable on the memory thesis, even if its “hidden supplier” framing has not matched reality for years. See our piece on SK Hynix stock for the pure-play side of the same trade.
Ready to see the research? Click here to access the Near Future Report.
NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.