DRAM memory modules. HBM stacks this same memory silicon vertically and bonds it right beside the GPU — which is exactly why it has become the bottleneck. Photo: Mister rf / Wikimedia Commons, CC BY-SA 4.0.

On a normal Thursday in June, three earnings calls, a Nvidia keynote, and a quiet certification notice told the same story from four directions: the hardest thing to buy in artificial intelligence right now is not a GPU. It is the memory bolted next to it.

Every AI accelerator, every Nvidia H200, B200, and the new Vera Rubin parts, is fed by stacks of a special chip called high-bandwidth memory, or HBM. There is not enough of it. It has been sold out, by some accounts, more than a year in advance. And the race to make it has done something almost no one predicted five years ago: it turned the sleepy, commodity memory business into one of the most strategic, most profitable, and most geopolitically loaded corners of the entire chip industry.

For a decade the story of computing was the logic chip, the processor, the GPU, the transistor count. Memory was the afterthought you bought by the gigabyte. That era is over. A modern AI chip can do trillions of operations a second, but only if it can be fed data fast enough, and a single conventional memory chip simply cannot keep up. The answer is to stack memory dies vertically, right beside the processor, and run an enormously wide pipe between them. That is HBM.

Because HBM sits so close to the logic and ships in such volume, it has become the true rate-limiter on how many AI servers the world can build. You can design a brilliant accelerator, but if you cannot get the memory stacks, you cannot ship the box. That single fact is why three companies you may never have thought about are suddenly among the most important in technology.

Here is the upset. The world's largest memory maker, Samsung, is not winning HBM. The leader is SK Hynix, a smaller Korean rival that bet early and heavily on stacked memory, and now supplies the lion's share of Nvidia's HBM. As of the second quarter of 2025, SK Hynix held roughly 62% of the HBM market.

The second surprise: Micron, the only American memory maker, passed Samsung for the number-two spot, at roughly 21% versus Samsung's 17%. Samsung, the giant, came third in the one product that matters most. The financial gap is just as stark, SK Hynix reported a record quarter with an operating margin of roughly 58%, the kind of number you almost never see in a business historically known for losing money at the bottom of every cycle.

Strip away the acronyms and HBM is a simple, clever idea. Instead of laying memory chips flat on a board, you stack a dozen or more DRAM dies on top of each other and drill thousands of vertical electrical connections, called through-silicon vias, straight down through the stack. You then place that tower on a silicon interposer millimeters from the processor. The result is a data pipe vastly wider than ordinary memory, with far less distance for the signal to travel.

The generations march fast: HBM2E delivered around 460 gigabytes per second per stack, HBM3 about 819, today's HBM3E around 1,229, and the new HBM4 pushes past 2,000, while widening the internal bus from 1,024 to 2,048 bits, the largest architectural jump the standard has seen. Each leap is harder than the last, which is exactly why so few companies can do it.

The market followed the difficulty. HBM revenue is estimated at roughly $37 billion in 2025, climbing toward $58 billion in 2026, with Micron projecting it could approach $100 billion by 2028. Treat that last figure as a supplier's own forecast, but even the near-term numbers describe one of the fastest growth curves in the history of semiconductors.

Here is the part that reaches into your own life. HBM is expensive to make not just because it is hard, but because it is wafer-hungry: producing a gigabyte of HBM consumes roughly three times the silicon wafers of a gigabyte of standard DDR5. As the makers pour capacity into HBM, with AI applications estimated to absorb around 20% of all DRAM wafers in 2026, there is simply less factory time left for the ordinary memory in laptops, phones, and game consoles.

The next chapter has a date. On June 5, 2026, Nvidia's Jensen Huang confirmed that all three memory makers, SK Hynix, Micron, and Samsung, had passed HBM4 qualification for Nvidia's next-generation Vera Rubin platform, with shipments targeted for the third quarter of 2026. That matters for two reasons. First, it puts Samsung back in the game at the cutting edge. Second, it means the supply base for the most advanced memory is, for now, exactly three companies, two Korean, one American, and not one of them Chinese.

Which raises the obvious question. China's leading memory champion, CXMT, has made real strides in conventional DRAM, including DDR5, but on HBM it is widely assessed to be roughly three years behind the leaders, and cut off from much of the advanced equipment needed to close the gap. For now, the most strategic memory on Earth is made entirely outside China, a fact that turns these three companies into instruments of policy as much as commerce.

For the investor
You are buying a genuine super-cycle wrapped around a historically cyclical business. SK Hynix's margins are extraordinary and may stay high while HBM is sold out, but "sold out" is a demand condition, not a law of nature. Size the position to the part of the thesis you actually believe, and respect that memory has humbled confident investors many times before.

For the supply-chain & hardware leader
If you build anything with memory in it, model tighter conventional DRAM and DDR5 supply and firmer prices through 2026 as wafers migrate to HBM. Qualify a second source now, not when the shortage bites.

For the policymaker
The most advanced memory on the planet is made by two Korean and one American company, and none in China. That is a strategic asset and a concentration risk at the same time. Resilience means more packaging and more sources, not fewer.

1. SK Hynix investor relations, quarterly results and HBM roadmap [10 min]

2. Micron investor relations, HBM share and market forecast [10 min]

3. TrendForce, HBM market share and pricing research [8 min]

4. JEDEC HBM standards, the technical spec behind the bandwidth numbers [15 min]

HBM (high-bandwidth memory): DRAM dies stacked vertically and placed beside a processor for a very wide, short data pipe. The scarce input behind AI accelerators.

DRAM: the ordinary working memory in computers and phones, historically a cheap, cyclical commodity.

TSV (through-silicon via): a vertical electrical connection drilled through stacked chips to link the layers.

Interposer: a slice of silicon that carries the ultra-dense wiring between the memory stack and the processor.

Bandwidth: how much data per second can move, measured in gigabytes per second; HBM's whole reason to exist.

HBM4: the newest generation, past 2,000 GB/s per stack, with a doubled 2,048-bit bus, qualified for Nvidia's Vera Rubin in 2026.

Qualification: the customer's sign-off that a supplier's part meets spec; for memory, Nvidia's qualification is the gate to the AI market.

Sourced facts. The market-share figures (SK Hynix ~62%, Micron ~21%, Samsung ~17%, Q2 2025), market-size estimates (~$37B 2025, ~$58B 2026, ~$100B 2028 per Micron), bandwidth-per-generation numbers, the ~3x wafer intensity and ~20%-of-DRAM-wafers figures, SK Hynix's record ~58% operating margin (Q4 2025; ~41% in Q2 2025), and the June 5, 2026 HBM4 qualification for Vera Rubin all come from company filings, TrendForce and analyst estimates, JEDEC specs, and Nvidia's public statements.

What is estimate, and what is interpretation. Market sizes and shares are analyst estimates that vary by source and shift quarter to quarter, the $100B-by-2028 number is Micron's own forecast. Calling memory's shift a "regime change" is my framing, the cyclicality I warn about is the counter-case, and it is real. Treat every forward number here as conditional on AI demand staying strong.

Silicon & Steel Intelligence Desk, Supply Chain Strategy & Semiconductor Analysis. Nothing here is investment advice. Corrections & coffee: [email protected]