[2026 Latest] What is HBM (High Bandwidth Memory)? "Why do the stock prices of SK Hynix, Samsung, and Micron keep rising?"
"If the GPU is the brain of AI, HBM is the aorta that carries the blood."
Currently, there is no doubt that NVIDIA is attracting the most enthusiastic attention in the global stock market. However, the cutting-edge AI GPUs manufactured by NVIDIA (H200, Blackwell, and the upcoming Vera Rubin in late 2026) are nothing more than lumps of iron that emit heat without "a certain component."
That component is HBM (High Bandwidth Memory).
As of 2026, HBM has become the "most difficult to procure, highest margin, and most technically challenging" component in the global semiconductor industry.
The more NVIDIA GPUs are sold, the more the demand for HBM skyrockets exponentially. Just by understanding this extremely robust and simple chain (supply chain), the resolution of your investment in the three memory manufacturers—SK Hynix, Samsung, and Micron—will increase dramatically.
This article is a "complete preservation edition" guide that thoroughly dissects the technical advantages of HBM, the market power balance, and the latest trends as of 2026 for investors.
Building on our previous article, "What is Memory Semiconductor," we will take a step further and dive deep into "What is HBM." Please read along with me.
1. What is HBM—"From a single-story house to a high-rise apartment"
First, let's understand the fundamental difference between the "ordinary memory (DRAM)" in our PCs and smartphones and "HBM."
Conventional DRAM (e.g., DDR5) = "A single-story residential area"
In conventional PCs and servers, black square memory chips are arranged "planarly (side-by-side)" around the CPU or GPU.
Data exchange is carried out through wiring (buses) drawn on the board. To use an analogy, this is like "transporting cargo from a single-story warehouse on a large plot of land to a factory (GPU) via narrow public roads."
Because the width of the road is fixed, there is a physical limit (about 50-100 GB per second) to the amount of data (bandwidth) that can be transported at one time.
HBM (High Bandwidth Memory) = "A high-rise apartment directly connected to the factory"
On the other hand, HBM has fundamentally changed the concept.
Instead of lining up chips side-by-side, "DRAM chips shaved to extreme thinness are stacked vertically in layers of 8, 12, or 16" (this is called 3D stacking).
Furthermore, instead of connecting the stacked chips with external wiring, countless microscopic "vertical holes" invisible to the naked eye were drilled through the very center of the chips and filled with copper to skewer them.
This miraculous technology is called TSV (Through-Silicon Via).
Then, this memory, which has become a high-rise apartment, is placed "right next to (a few millimeters away)" on the same board (interposer) as the GPU itself.
Normal DRAM: [GPU] <——(thin and long wiring)——> [DRAM]
HBM: [GPU + HBM (stacked vertically)] * A single integrated package
Why does AI need an "ultra-wide road"?
Let's compare AI training and inference to an "enormous kitchen." NVIDIA's latest GPUs are "world-class genius chefs" capable of performing tens of trillions of calculations per second.
However, no matter how talented the chef is, if the passage to transport ingredients (data) from the refrigerator (memory) is narrow, the chef will be left idle, and the productivity of the entire kitchen will drop.
Large Language Models (LLMs) like ChatGPT constantly move vast amounts of parameters in and out. HBM has expanded this passage into an "ultra-wide highway" that is 100 times faster than conventional memory (4.8TB to over 8TB per second). This is a level of speed that allows for reading and writing the data equivalent of 2,000 high-definition movies in just one second. Without HBM, modern generative AI would be physically impossible.
2. The "Deep Moat" of Entry Barriers and the Oligopoly Market
One of the most important concepts in investing is the "economic moat" that keeps competitors at bay. The HBM market is a market where this moat is as deep as the Mariana Trench.
Market Size: The Shock of a 40% CAGR
According to estimates by major Wall Street investment banks, the HBM market size is projected to surge from approximately $35 billion in 2025 to about $55 billion in 2026, and reach $100 billion (approximately 15 trillion yen) by 2028.
While the CAGR of the highly popular obesity drug (GLP-1) market is said to be around 20%, the HBM market continues to expand at a speed that surpasses it.
Why can only "three companies" make it?
Currently, the only three companies in the world capable of supplying HBM areSK Hynix, Samsung Electronics, and Micron Technology.
Chinese companies and others are in hot pursuit, but they are no match when it comes to cutting-edge HBM for AI. Why can't other companies enter the market?
The yield barrier of TSV (Through-Silicon Via): The technology to drill microscopic holes in chips and connect them vertically takes over 10 years to master. If there is even the slightest misalignment or thermal distortion, the entire stack of chips (worth tens to hundreds of thousands of yen) must be discarded as waste.
Extreme precision in packaging: The technology to stack as many as 16 chips with a precision of less than 1 micron (1/1000th of a millimeter) while creating a structure that dissipates heat is now in the realm of traditional craftsmanship.
Astronomical capital investment: Building a production line dedicated to HBM requires investments on the scale of hundreds of billions to trillions of yen over several years.
A honeymoon relationship with NVIDIA: HBM is not a commodity product; it is a semi-custom product tailored to the GPU's architecture. It takes 1 to 2 years to pass NVIDIA's rigorous quality qualification process.
This fact that "it is extraordinarily difficult to make" guarantees a solid oligopoly by these three companies and "extremely high profit margins" (said to be over 50-60% gross margin).
3. The "Current Status" of the Three Major Memory Companies and Stock Analysis
Now, let's dissect the current standing of the three companies dominating this golden market from an investor's perspective.
1. SK Hynix — The overwhelming champion, NVIDIA's "best partner"
Estimated market share: Over 50%–60%
Strengths: Technological leadership, strong alliance with TSMC
Investment rating: ★★★★★ (Maximum exposure to HBM)
South Korea's SK Hynix is the undisputed "champion" in the HBM market. The reason they won is simple: "they never stopped investing in HBM even during the winter period when other companies were hesitant to develop it".
As a result, when NVIDIA sparked the AI revolution, SK Hynix was the only company capable of stably supplying large quantities of HBM3 and HBM3E. For the next-generation "HBM4" standard in 2026, SK Hynix is also forming a powerful tag team with TSMC (the world's largest foundry) and is poised to monopolize the initial share for NVIDIA's next-generation "Vera Rubin" GPU.
If you want to purely enjoy the "benefits of HBM," there is no company that surpasses SK Hynix.
2. Samsung Electronics — The awakened giant, a scenario for a comeback
Estimated market share: 20%–30%
Strengths: Overwhelming financial power, undervalued valuation
Investment rating: ★★★ (For contrarian value investors)
Although Samsung is the world's largest memory manufacturer, it suffered the humiliation of "falling behind SK Hynix" when it came to HBM. Due to issues with heat generation and yields, it struggled to pass NVIDIA's HBM3E certification, and its stock price remained sluggish.
However, the situation is changing drastically from late 2025 to 2026. Through a management overhaul and massive R&D investment, they have overcome these issues and are significantly reclaiming supply slots for NVIDIA in the "HBM4" generation.
Samsung's greatest strength is that it can "complete everything in-house (turnkey)," from memory design and manufacturing to advanced packaging. The current stock price does not fully reflect the benefits of AI, making it an attractive level for contrarian investors.
3. Micron Technology (MU) — The assassin of the US market, a highly efficient niche leader
Estimated market share: 10%–20%
Strengths: High power efficiency, US-listed (easy to buy from Japan)
Investment rating: ★★★★ (Top candidate for the new NISA growth investment quota)
Headquartered in Idaho, USA, Micron is in 3rd place in terms of market share, but it is operating with extreme strategic focus. They skipped the development of older generation HBM and poured all their resources into the cutting-edge "HBM3E" at once.
The biggest feature of Micron's HBM is "power efficiency (energy saving),." As data center power consumption becomes a global social issue, Micron's HBM, which "costs less in electricity for the same computing power," is receiving enthusiastic support from many hyperscalers (such as Microsoft and Google).
HBM production for 2026 is already "Sold Out," and stable revenue growth is expected.
4. The "Wafer Trap (Cannibalization)" that investors absolutely must know
Here, I will share the "biggest catalyst" (stock price growth factor) in memory stock investing that is most important and yet overlooked by many individual investors.
It is the structure where "when HBM sells, the prices of ordinary memory unrelated to AI (DRAM for PCs and smartphones) also skyrocket".
In the industry, this is called "wafer cannibalization".
As mentioned earlier, HBM involves stacking chips vertically, and the testing process is also complex. Therefore, to produce the same data capacity of memory, it consumes about three times the silicon wafers (materials) and production capacity of ordinary DRAM.
What happens when SK Hynix and Samsung switch factory lines exclusively to HBM to increase production of the exceptionally high-margin HBM? Naturally, they lose the capacity to make "ordinary DRAM" for PCs and smartphones.
As a result, even in the PC and smartphone markets where demand is not exploding, the supply volume drops sharply due to the convenience of the supply side (factories), so the price of general-purpose DRAM is forcibly raised.
In other words, the "magic wand" that is HBM functions as a powerful lever that raises the overall profitability of memory manufacturers. The old conventional wisdom that "you can't buy memory stocks because PC sales are poor" has completely collapsed with the advent of HBM.
5. The tectonic shift in late 2026: "HBM4" and Vera Rubin
If you want to look into the future as an investor, you must not miss the transition to "HBM4," which will begin full-scale mass production in the second half of 2026.
NVIDIA will deploy a next-generation AI architecture called "Vera Rubin" as the successor to the current Blackwell. It is HBM4 that will support this Rubin.
The evolution of HBM4 goes beyond mere "capacity increase."
Doubling of the interface: The data path will double from the conventional 1,024 bits to 2,048 bits.
Standardization of 16-high stacking (16-Hi): Stacking up to 16 layers, where 12 was previously the limit, achieving a crazy capacity of up to 64GB per package.
Integration of logic die: This is the biggest change. The base die at the bottom of the HBM will now have some calculation functions (logic functions). As a result, it cannot be made by memory manufacturers alone, and advanced collaboration with foundries (contract manufacturers) like TSMC becomes essential.
This transition to HBM4 means that the technical hurdles will jump even higher. As a result, it may enter a "winner-takes-all" phase, moving from a state where "three companies share the market nicely" to one where "the company that meets the technical requirements takes it all".
At this point, the coalition of SK Hynix and TSMC has the greatest advantage.
[Special Column] A blind spot in HBM investment? The threat of the "Groq (LPU)" architecture and NVIDIA's next move
I have talked about the overwhelming superiority of HBM so far, but in order to make a fair investment judgment, it is also necessary to look at the "risks that threaten the future of HBM".
The biggest threat is the rise of new AI chip architectures (LPU: Language Processing Unit) represented by "Groq", and the impact if NVIDIA itself begins to explore that direction as a dedicated inference chip.
1. What exactly is "Groq (LPU)"? Why is HBM unnecessary?
Groq is a semiconductor startup founded by former Google TPU developers. The biggest feature of their chips is that "they do not use HBM at all, but instead pack a massive amount of ultra-high-speed memory called 'SRAM' directly into the chip".
Weakness of HBM: No matter how fast it is, because it is located "outside (next to)" the chip, it takes a small amount of time to retrieve data.
Strength of Groq (SRAM): Because the memory is "completely integrated" with the calculation circuits, latency is almost zero. In text generation (inference) like ChatGPT, it achieves ultra-high speeds more than 10 times faster than NVIDIA's GPUs.
If NVIDIA itself, in order to monopolize the "inference market," begins to abandon HBM and deploy its own chips with an "SRAM-specialized architecture" like Groq's on a large scale, will the HBM market collapse?
2. HBM demand will not collapse. Rather, "market segmentation" will progress.
In conclusion, investors do not need to be overly afraid of this "Groq shock." There are three reasons for this.
1. SRAM is "extremely expensive" and "has too little capacity"
SRAM is overwhelmingly fast, but its manufacturing cost is extremely high, and its physical size is large, so only a small capacity (around several hundred MB) can be mounted on a single chip. On the other hand, the latest AI models (GPT-4 and models to come) have trillions of parameters, and the data size reaches several terabytes.
If you try to run the latest AI using Groq's method, you have to daisy-chain hundreds of chips, which results in skyrocketing costs and power consumption. "In terms of the balance between 'large capacity and ultra-high speed,' there is currently no physical optimal solution that surpasses HBM.
2. HBM is absolutely essential for 'learning' and 'multimodalization'
What SRAM-specialized chips are good at is only having an already completed AI provide answers in "text" (inference). The "learning (training)" process to make AI smarter, and the inference of "multimodal AI" that processes not only text but also video, audio, and high-resolution images instantaneously, require a vast space to temporarily store enormous amounts of data. Here, too, HBM comes into play.
3. HBM itself is evolving toward 'Groq-ification (integration of logic)'
Actually, memory manufacturers are not just standing by and watching. "HBM4", mentioned in Section 5, will provide logic functions capable of performing calculations on the base die at the bottom. In other words, it is attempting to evolve by "moving from the memory side toward the calculation circuits to bring latency even closer to zero."
3. Implications for investment
NVIDIA certainly has the potential to explore custom chips (LPU-like ones) that do not use HBM for edge or specific applications specialized for simple text inference.
However, that is merely "segmentation by use case" as the pie of the entire AI market expands. At the heart of "learning for cutting-edge models and complex multimodal inference", which is the mainstream of global AI investment, NVIDIA's flagship GPUs and the HBM that accompanies them will continue to sit firmly.
[Summary] The "AI memory supercycle" has only just begun
The memory industry of the past was a sector where it was extremely difficult to time investments, as it repeated "3-4 year cyclical booms and busts" influenced by the sales of PCs and smartphones.
However, the current cycle that began in 2024 is unlike anything in the past. This is being called the "AI Memory Supercycle". Giant IT companies (hyperscalers) like Microsoft, Amazon, and Google view their massive capital investments in AI infrastructure as a "survival race," and they are buying up GPUs and HBM regardless of economic cycles.
Furthermore, the era of "Edge AI (AI PCs, AI smartphones)" is coming, where AI will be installed not only in data centers but in PCs and smartphones themselves, and larger capacity, higher-speed memory will be required for all devices.
"Will the HBM market eventually exceed the total market size of DRAM?"—a prediction that some analysts used to make half-jokingly is beginning to be discussed as a realistic possibility as of 2026.
As the pickaxe for digging into the golden mountain that is NVIDIA, HBM will continue to shine at the center of the semiconductor industry.
Only those investors who discard the old conventional wisdom that "memory is a commodity" and realize the true value of HBM as high-value-added infrastructure will be able to reap the fruits of this supercycle.
