Ask anyone what powers the AI boom and you will hear one word: Nvidia. The GPUs get the headlines, the trillion-dollar market cap, and the keynote applause. But the component that actually decides how fast the AI era can move is not the processor. It is the memory bolted right next to it — and it leads straight back to a handful of factories in South Korea.
That component is HBM: high-bandwidth memory. It is the quiet choke point of the entire AI buildout, and understanding it explains almost everything else — why memory prices exploded, why SK Hynix passed Samsung, and why hyperscalers spending hundreds of billions of dollars still cannot get enough compute.
Disclaimer: This is general analysis, not investment advice.
What is HBM, and why does AI need it?
A modern AI accelerator has two halves that matter: the logic (the GPU or custom chip that does the math) and the memory (which feeds it data). For years, the logic was the star. But training and running large AI models is less about raw math and more about moving enormous amounts of data to the processor fast enough to keep it busy. A starved GPU is an expensive paperweight.
HBM solves that by stacking memory chips vertically — like a tiny skyscraper — and wiring them directly beside the GPU with a very wide, very fast connection. The result is a firehose of bandwidth. Nvidia’s next-generation Vera Rubin platform moves to HBM4 at a reported 22 TB/s of bandwidth per GPU — roughly 2.75× the previous Blackwell generation. That bandwidth, not the logic, is increasingly what determines how many AI queries a chip can serve.
Here is the fact that reframes the whole industry: HBM now accounts for roughly half the bill-of-materials cost of an Nvidia Blackwell chip. Half. The “Nvidia” chip everyone is fighting to buy is, by cost, as much a memory product as a logic product.
The numbers are staggering
The demand curve for HBM looks less like a component market and more like a gold rush:
| Metric | Figure |
|---|---|
| HBM market size, 2025 | ~$7.3 billion |
| HBM market size, 2026 | ~$55 billion |
| Growth | roughly 7× in a single year |
| SK Hynix share of HBM supply | ~57–62% |
| HBM as share of a Blackwell chip’s cost | ~50% |
A market going from $7 billion to $55 billion in twelve months is almost unheard of. And unlike commodity memory, HBM is extraordinarily hard to make — it requires advanced stacking and packaging that only a few companies in the world have mastered.
Who controls the choke point? (Hint: it’s not America)
Three companies make essentially all of the world’s HBM: SK Hynix, Samsung, and Micron. Two of them are Korean, and one — SK Hynix — controls an estimated 57–62% of supply. That single fact ties together several stories we have covered:
- It is why SK Hynix overtook Samsung to become Korea’s most valuable company: it got to HBM leadership first.
- It is a big reason ordinary DRAM prices exploded and sparked a price-fixing lawsuit: the same fabs that make your laptop’s memory were redirected to far more profitable HBM.
In other words, the AI boom’s most important bottleneck is held by a tiny number of firms, concentrated in Korea. Whoever controls HBM has a hand on the throttle of the entire AI economy — and, right now, extraordinary pricing power.
Does Nvidia play favorites? It is a fair question to raise. Nvidia’s CEO, Jensen Huang, was born in Taiwan, and Nvidia’s logic chips are fabricated by Taiwan’s TSMC — while its HBM comes mostly from Korea’s SK Hynix and Samsung. Does heritage tilt the table toward Taiwan and against the Korean memory makers? In practice, almost certainly not. A supply chain this large runs on leverage and security of supply, not sentiment. Nvidia needs Korean HBM far too badly to play favorites — and it does the opposite of picking a friend: it deliberately qualifies multiple suppliers (SK Hynix, Samsung, Micron) precisely to pit them against one another on price and guarantee volume. Jensen Huang is, above all, an operator; he has publicly praised SK Hynix’s memory because it works and he cannot build AI chips without it. In a market this tight, the only loyalty is to whoever can deliver the bandwidth. Business logic, not roots, signs the contracts.
The bottleneck keeps moving down the stack
For a while, the constraint on AI was GPUs. Then the industry learned an uncomfortable lesson: solve one bottleneck and it simply moves to the next-weakest link. In 2026 the constraint has cascaded down the supply chain:
- Logic chips — still tight, but available with enough advance commitment. Every major chip is fabricated on TSMC’s 3nm node, which is running at 100% utilization with demand reportedly ~3× supply.
- HBM — the memory itself, dominated by SK Hynix, sold out well in advance.
- Advanced packaging — the step that bonds logic and HBM together is its own scarce resource.
- Power — and here is the newest wall. The binding limit in a growing number of markets is no longer silicon at all; it is electricity. The IEA warns data centers will draw more power in 2026 than all of Japan, and Microsoft alone has disclosed an estimated $80 billion Azure backlog it cannot fulfill because it lacks the power to turn the servers on.
Each new GPU generation makes this worse, not better, because it demands more HBM per chip. More bandwidth means more stacks, which means the memory bottleneck tightens exactly as the logic gets faster.
Why this matters — and what could change it
The strategic takeaway is simple: in the AI era, memory is power. The company that leads in HBM captures a huge share of the value of every AI chip sold, regardless of whose logo is on the front. That is why the AI boom has been such a windfall for Korean memory makers — and why their fortunes now rise and fall with hyperscaler capex plans that run to $725 billion in 2026 alone.
But three things could reshape the picture:
- Samsung catching up. If Samsung closes the HBM4 gap with SK Hynix, the supply crunch — and the pricing power — eases.
- The bubble question. Hyperscalers are investing something like $13 for every $1 of current AI revenue. If that ROI gap does not close, capex could stall, and HBM demand with it. (We will dig into the bubble debate in a future piece.)
- Custom silicon. Google, Amazon, Microsoft, and Meta are all building their own AI chips to reduce their dependence on Nvidia. But here is the twist: those custom chips need HBM too. Changing the logo on the processor does not change who makes the memory.
The bottom line
The story of the AI boom is usually told as a story about Nvidia and its GPUs. But look one layer down and a different picture appears: the pace of the entire AI era is being set by a stack of memory chips, made by a handful of companies, most of them Korean. The GPUs get the glory. HBM holds the throttle — and for now, that throttle is in Korean hands.
FAQ
What does HBM stand for? High-Bandwidth Memory — memory chips stacked vertically and placed right next to an AI processor to feed it data extremely fast. It is what keeps expensive GPUs from sitting idle.
Why is HBM so important for AI? Running large AI models is bottlenecked by how fast data can reach the processor, not just by raw computing power. HBM provides that bandwidth, and it now makes up roughly half the cost of a top-end Nvidia AI chip.
Who makes HBM? Just three companies: SK Hynix, Samsung, and Micron. SK Hynix leads with an estimated 57–62% of the market, which is a major reason it recently became South Korea’s most valuable company.