HBM: the memory chip race that AI depends on, and only SK Hynix, Samsung and Micron can do it

Two years ago, if someone asked me about Micron, I would answer that this is a company that produces USB and memory cards for cameras. This month, Micron just announced third-quarter results with profits 15 times higher than the same period last year, from $1.88 billion to $28.2 billion, and fourth-quarter forecasts of $49 to $51 billion in revenue. This result makes them being considered the next NVIDIA. This change didn't happen because Micron suddenly got better. It happened because the world just realized that the type of chip they make: high bandwidth memory (HBM — High Bandwidth Memory), is the real bottleneck of the AI ​​race. And there are only three companies in the world that can produce it.

What is HBM and why is it more important than you think?

To understand why HBM is in this position, it helps to understand a little about how GPU AI works. When an NVIDIA H100 or Blackwell is running a large language model, it needs to continuously read and write terabytes of data every second: model parameters, intermediate results, inputs, and outputs. If the memory isn't fast enough to feed data to the computing cores, the entire GPU will sit waiting, and you're wasting tens of thousands of dollars in hardware.

SK Hynix — the ruler holding the throne

SK Hynix currently holds about 62% of the global HBM market share as of the second quarter of 2026. They achieved this position mainly thanks to a timely decision: going first in mass producing HBM3E, the generation of memory chips that NVIDIA uses in the H200 and Blackwell lines, when Samsung was still struggling with quality issues and Micron had not yet caught up.

Currently, SK Hynix is ​​also the leading HBM4 supplier for NVIDIA with a 16-layer HBM4 chip reaching a speed of 11.7 gigabits per second, the fastest in the industry to date. In the HBM4 allocation for NVIDIA's next-generation Vera Rubin platform, SK Hynix is ​​estimated to have about 55%, Samsung 25% and Micron 20%.

Samsung with the fall and the journey to catch up

Samsung is the most notable story in this race, but not in a positive way. Throughout 2024 and into early 2025, Samsung continuously failed NVIDIA's HBM3E quality tests. The reason was revealed that Samsung's chip generates too much heat and consumes power beyond standards, a serious problem when an AI GPU can attach up to 6 or 8 HBM stacks, and the overall temperature is the deciding factor in stable performance.

Micron is an outsider who suddenly rose up

Micron is the only American company among the three, and they're taking advantage of that. While SK Hynix and Samsung are Korean companies that face geopolitical pressure in selling chips to Chinese customers, Micron has the advantage of reaching US customers more easily and is prioritized by US government domestic chip production support programs.

The third quarter financial results clearly reflect Micron's rising position: revenue of 41.45 billion USD (4 times more than the same period), profit of 28.2 billion USD (15 times more), and forecast for the fourth quarter of 49-51 billion USD. Micron's HBM market share surpassed Samsung's in the second quarter of 2026 — something that has never happened in the history of the HBM market.

HBM4 and the next battle

I see HBM4 as the point where Samsung has a real chance to regain market share — if they don't repeat the quality problems of the previous generation. The industry is looking to Samsung to see if the painful lessons from HBM3E will actually change their quality control processes.

Whoever controls the HBM, controls the speed of the AI

I think there's one thing that many people get wrong when talking about the AI ​​hardware race: they only talk about GPUs. But the GPU cannot run faster than the memory that feeds it data. An NVIDIA Blackwell with slow or lacking HBM is a race car that mounts an F1 engine to a 5 liter fuel tank. This is why Meta signed a $27 billion contract with Nebius for infrastructure using NVIDIA Vera Rubin — but Vera Rubin needs HBM4 from the three vendors above. This is also the reason why Apple increased the price of MacBook and iPad because DRAM (HBM's cousin) is being depleted of supply by AI data centers. And this is why South Korea just announced a $576 billion plan to maintain its dominant position. Three companies — SK Hynix, Samsung, Micron — control a resource that the world needs but that almost no one can produce on their own in the short term. That is the definition of strategic leverage. And Micron's $41 billion in revenue in one quarter is just the tip of that iceberg.