Stage 9 of 14Extreme concentrationHard to substitute
Chapter 9: Memory and HBM
SK hynix, Samsung and Micron make nearly all high bandwidth memory (HBM), the stacked memory beside AI chips. SK hynix took 63.2 percent of HBM revenue in 2025.
Sharma, Yashvardhan. “Memory and HBM.” Chip Supply Chain: How AI Chips Are Made. Updated 29 October 2026. https://chipsupplychain.org/memory-and-hbm/
A twelve-high HBM stack, pulled apart and cut open: a logic die at the base, shown in copper, and twelve memory dies above it. The finished part is 11 by 11 mm and 0.72 mm tall. Each memory die is ground to about 30 micrometers, thinner than a human hair. The footprint is to scale, but the heights are exaggerated about fifteen times.
Parts, as numbered on the drawing
- 1Memory dies: SK hynix (KR), Samsung (KR), Micron (US)
- 2Base die: SK hynix (KR) through HBM3E, TSMC (TW) for SK hynix’s HBM4
- 3TSVs (through-silicon vias): etched and filled with copper by the memory maker
- 4Microbumps: joined by the memory maker, see How a stack is built
In plain terms
, or HBM, is where an AI keeps the numbers it is working on, and it is built to hand them over fast. A memory maker tiny holes into each memory , fills them with copper and grinds the wafer until the copper shows through its back. It tests each of the wafer’s chips and stacks up to sixteen good ones into a cube, with every layer wired to the ones below. The cube then sits beside the processor, the chip that does the calculating. Stacking is the hard part because each chip is ground thinner than a human hair and the finished cube has to stay within a height limit of less than a millimeter. One chip that fails after stacking can scrap the whole cube. Only SK hynix, Samsung and Micron make HBM . One US export rule covers HBM from all three, including the two based in South Korea.
In short
US-based Micron went from 5.8 percent of revenue in 2024 to 23.1 percent in the first quarter of 2026, according to IDC figures in SK hynix’s US prospectus1. Over the same period, Samsung’s share fell from 37.8 percent to 20.5 percent1.
Since December 2024, a US license has been required, unless an exception applies, to export to China any HBM moving more than 2 GB per second per of its area2. The Institute for Progress reported in October 2025 that China’s CXMT planned to produce HBM3 in 20263. Even so, working from TrendForce’s estimate of world HBM shipments, the institute projected that producers outside China would make 70 times as much HBM as China in 20263. In June 2026 SemiAnalysis judged that CXMT was still struggling to stabilize supply of eight-high HBM34.
An moves numbers from memory to its arithmetic units and back. The US office wrote in December 2024 that a processor cannot reach its full capabilities unless its memory capacity and bandwidth rise with its speed2.
- Who leads
- Where it is made
- KRSouth Korea SK hynix in Cheongju and Icheon, Samsung in Pyeongtaek and Hwaseong
- USUnited States Micron headquarters and R&D, and an SK hynix packaging plant under construction in Indiana
- TWTaiwan Micron DRAM and HBM production
- JPJapan Micron Hiroshima, 1-gamma DRAM
- SGSingapore Micron HBM advanced packaging plant, adding capacity from 2027
How HBM works and why AI needs it
Each layer in the cube is a , the ordinary kind of computer memory. Each layer is ground thin and bonded to the one below through tiny metal called microbumps. The whole stack sits on a , a chip that manages it. The copper-filled holes are called , or TSVs. They let a signal reach any layer without a wire running around the outside.
An cube has a 1,024-bit interface, meaning it exchanges data with the processor over 1,024 wires at once8. That is sixteen times wider than a standard DDR5 memory module, the memory stick in a PC8. HBM4 doubles that to 2,048 wires9. Ordinary memory cannot be made that wide, because the pins would not fit on a . The stack can, since its wires run down through the TSVs and across a to the processor beside it, with no board in the way8. The interposer is a slice of silicon carrying fine wiring, set under both the memory and the processor (see Advanced packaging).
Inside a memory stack

- Base chip. A memory stack sits beside the GPU on a shared silicon base. It starts with a base chip, which controls the memory chips above it.
- Thin. Each memory chip is ground to about 30 micrometers, thinner than a human hair.
- Stack. Twelve chips are joined by dots of solder, and copper-filled holes run straight down through all but the top one.
- Height. The stack has to fit under the cooling plate. Twelve chips reach 0.72 mm, the old limit. HBM4 raised the limit to 775 micrometers and fits up to sixteen chips.
- Width. An HBM3E stack talks to the processor over 1,024 wires at once, sixteen times as many as a PC memory module. HBM4 doubles that to 2,048.
A chip needs the width most when it answers a prompt. Training reuses each stored number across a huge multiplication, so a chip can fetch little and compute a lot. But to generate one word of an answer, the chip reads the whole model from memory, plus the whole record of the conversation so far. Then it does a little arithmetic and waits for the next read. That record grows with the length of the conversation and the number of users served at once. So memory size and speed usually set how many people one chip can serve.
HBM3E and HBM4 compared
| Generation | Interface | Bandwidth per stack | Capacity per stack |
|---|---|---|---|
| HBM3E | 1,024-bit, 16 channels (32 pseudo-channels) | over 1.2 | 24 GB (8 chips), 36 GB (12 chips) |
| HBM4 | 2,048-bit, 32 channels | over 2.0 TB/s, up to 3.3 TB/s | up to 64 GB (16 chips, 32 Gb ) |
HBM3E
Micron’s HBM3E comes in two sizes, 24 GB with eight chips stacked and 36 GB with twelve10.
HBM4
All three HBM makers run their HBM4 faster than the 8 Gb/s per pin in the industry standard. The standards body published JESD270-4 in April 2025, setting the 2,048-bit interface and 2 TB/s per stack at 8 Gb/s per pin9. SK hynix says it was first to complete HBM4 development, in September 2025, at over 10 Gb/s per pin11. Samsung says it began and shipped the first commercial HBM4 in February 2026, at 11.7 Gb/s per pin, with a 4 logic base die12. It gives that HBM4 a maximum of 3.3 TB/s per stack12. Micron’s parts run above 11 Gb/s for more than 2.8 TB/s per stack13. Micron says it began volume production of its 36 GB twelve-high HBM4 in fiscal 2026 and expects to volume of HBM4E, the next generation, in calendar 202714.
Through HBM3E, SK hynix made its own base dies, the chips at the bottom of each stack15. For HBM4 it chose TSMC’s advanced logic process, so the base die can hold more functions15. As a result, TSMC, which does not make HBM, makes the base die for SK hynix’s HBM4.
As of April 2026, HBM4 reaches up to 3,300 GB/s per stack, against 2,000 in the JEDEC standard
Bandwidth per HBM stack
| HBM3E, 1,024-bit | 1,200 GB/s |
|---|---|
| HBM4, JEDEC standard | 2,000 GB/s |
| HBM4, advanced configurations | 3,300 GB/s |
Sharma, Yashvardhan. “Figure 09.2: Bandwidth per HBM stack.” Chip Supply Chain: How AI Chips Are Made. Updated 29 October 2026. https://chipsupplychain.org/memory-and-hbm/#fig-memory-and-hbm-hbm-bandwidth. Data: Siemens EDA, HBM3E and HBM4 IC design guide, April 2026.
How an HBM stack is built
- and thin. Etch vertical holes, the through-silicon vias, into a memory , fill them with copper, then grind the wafer from behind until the copper comes through.
- Stack and join. Makers join the chips through metal bumps by methods such as and MR-MUF17. In MR-MUF, short for mass reflow molded underfill, the chips are first connected through microbumps and stacked, and a liquid underfill is then injected to fill every gap at once and cured17. SK hynix uses its own version, Advanced MR-MUF, which stacks the chips, injects a liquid protective material between them and hardens it11. SK hynix says the process controls warping and dissipates heat well11. Samsung said in February 2024 that its twelve-high HBM3E uses TC-NCF, with a thin non-conductive film between the chips18. Samsung expected TC-NCF to help limit warping as chips get thinner18.
- Stay under the height limit. The package specification for HBM4 raised the height allowed for a finished cube from 720 micrometers to 775, still under a millimeter17.
is the next step, but it is not ready. It would join the chips copper to copper and to insulator, with no solder and no filler17. A tech note on SK hynix’s newsroom gives about 20 micrometers as the typical spacing between microbumps, against less than 1 micrometer for hybrid bonds17. It says full-scale adoption in HBM is most likely at HBM4E or HBM5, where stacks are projected to pass twenty layers17. Taller stacks trap more heat, and removing it becomes a major challenge at 16 to 20 layers17. The same tech note says hybrid bonding manages that heat more efficiently17.
Who makes HBM
Three firms took more than 90 percent of the revenue from ordinary computer memory, or DRAM, in the first quarter of 2026, according to IDC figures in SK hynix’s US prospectus1. HBM comes from the same three, SK hynix, Samsung and Micron1. No other company has both DRAM and a stacking line of its own. SK hynix says only a few firms have the expertise and resources to design, make and package memory chips1. It says HBM also needs advanced production processes, such as TSV packaging, and must pass the rigorous testing and approval processes its customers set1.
SK hynix says it was the first company to mass-produce several generations of HBM, including HBM3 and HBM3E1.
SK hynix and Micron are both adding HBM capacity. SK hynix broke ground in August 2026 on an plant in West Lafayette, Indiana, an investment of over $4 billion19. Micron is modernizing and expanding its DRAM and HBM production capacity in Taiwan14. It has also broken ground on an HBM advanced packaging plant in Singapore to expand its packaging capacity from early 202714.
SK hynix led HBM revenue in 2024, in 2025 and in the first quarter of 2026, and in that quarter Micron passed Samsung for second place. In 2024 SK hynix took 56.4 percent, Samsung 37.8 and Micron 5.81. In 2025 the three took 63.2, 19.3 and 17.4 percent1. In the first quarter of 2026 they stood at 56.4, 20.5 and 23.1 percent, which put SK hynix back at its 2024 share1. The split is public because SK hynix registered shares in the United States, and its US prospectus credits the research firm IDC1.
SK hynix took 63.2 percent of global HBM revenue in 2025
Global HBM revenue by supplier, 2025
| SK hynix (KR) | 63.2% |
|---|---|
| Samsung (KR) | 19.3% |
| Micron (US) | 17.4% |
Sharma, Yashvardhan. “Figure 09.3: Global HBM revenue by supplier, 2025.” Chip Supply Chain: How AI Chips Are Made. Updated 29 October 2026. https://chipsupplychain.org/memory-and-hbm/#fig-memory-and-hbm-hbm-suppliers. Data: SK hynix Form 424B4 prospectus, 10 July 2026, on IDC data.
Micron sees HBM reaching $100 billion in 2028, from an estimated $35 billion
HBM market, Micron’s estimate for 2025 and forecast for 2028
| 2025 | $35B |
|---|---|
| 2028 | $100B |
Sharma, Yashvardhan. “Figure 09.4: HBM market, Micron’s estimate for 2025 and forecast for 2028.” Chip Supply Chain: How AI Chips Are Made. Updated 29 October 2026. https://chipsupplychain.org/memory-and-hbm/#fig-memory-and-hbm-hbm-revenue. Data: Micron prepared remarks, fiscal first quarter 2026, 17 December 2025.
As HBM grows, memory makers have less factory capacity left for traditional DRAM. In December 2025 Micron put the HBM market at about $35 billion in 2025 and forecast that it would reach around $100 billion in 20285. Micron says HBM needs more wafers than traditional DRAM, the kind used in PCs and phones, to make the same number of bits14. SK hynix says the industry’s shift of production capacity to HBM has left traditional DRAM undersupplied1. The shortage, it says, helped drive a significant price recovery for that DRAM from the third quarter of 20251.
HBM export controls and China’s efforts
The leading accelerators made outside China use HBM from SK hynix, Samsung or Micron. A US export rule of 5 December 2024 covers every HBM stack that was in production when it was published, including stacks made outside the United States2. China’s strongest DRAM maker, CXMT, was still struggling to make HBM in steady supply by June 2026. That month SemiAnalysis judged that CXMT had yet to stabilize supply of eight-high HBM3 and faced even greater challenges at twelve-high4.
Nvidia’s carries 20 TB of HBM3E at up to 576 TB/s20. Each Nvidia Rubin carries up to 288 GB of HBM4 at up to 22 TB/s21. AMD’s MI355X carries 288 GB of HBM3E at 8 TB/s22.
The (BIS), the US export control office, published the rule on 5 December 20242. It added an export classification, .c, for “High bandwidth memory (HBM) having a ‘’ greater than 2 gigabytes per second per ”2. Its technical note defines that density as the memory bandwidth in gigabytes per second divided by the area of the package or stack in square millimeters2. The rule also covers stacks made outside the United States, through a , which applies US controls to goods made abroad with American technology2. BIS stated in the rule that all HBM stacks then in production exceeded that threshold2.
BIS reviews license applications to ship controlled HBM to or a destination such as China according to where the buyer is headquartered23. It presumes it will deny a license when the buyer or its ultimate parent is headquartered in one of those places, and presumes approval otherwise23.
Two provisions narrow the control. , at 740.25, applies to exporters headquartered in the United States or a Country Group A:5 ally7. It lets them ship HBM below 3.3 GB per second per square millimeter without a license, including to Macau and Country Group D:5 destinations such as China7. HBM sent to those destinations must be bought directly by the designer of the chip it will be packaged into7. Every shipment under the exception must go straight to the packaging site7. The rule also excludes chips that combine HBM and logic and whose main function is processing, which other chip controls may cover instead2.
The think tank said in December 2024 that the rules still allowed HBM2, a generation introduced in 2016, to go to China, with limits on who could buy it and what for24. It also said Nvidia’s H20, whose HBM3 is packaged with the processor, was then still exportable under the performance limits for AI chips24.
In December 2024 CSIS called CXMT one of only two Chinese firms with a credible path to producing advanced HBM, though it did not name the other24. The Institute for Progress reported in October 2025 that CXMT was planning to produce HBM3 in 20263. It also estimated that China would make virtually no HBM in 20253.
Huawei plans to use HBM of its own design in its Ascend 950 AI chips, but when it announced the designs in September 2025 it did not say who would make the memory6. The two designs are called HiBL 1.0 and HiZQ 2.06. Huawei said the 950DT, due in the fourth quarter of 2026, would carry 144 GB of HiZQ 2.0 at 4 TB/s6.
Flashcards
Recall each answer, then turn the card.
What is HBM?
Memory chips stacked into a cube beside the processor, wired straight down through the stack so they can hand over data fast.
Why does an AI chip need such fast memory?
To write each word of an answer, it reads the whole model from memory.
What is the hard part of making HBM?
Stacking. Each chip is ground thinner than a hair, the cube must stay under a millimeter tall, and one bad chip can scrap the whole cube.
Who makes HBM?
SK hynix and Samsung of South Korea and Micron of the United States make nearly all of it.
Four things to remember
- HBM is memory stacked into a cube beside the processor and wired to pass data fast.
- An AI chip reads its whole model from memory for each word, so memory speed limits it.
- SK hynix, Samsung and Micron make nearly all HBM.
- Since December 2024 US controls have covered the HBM stacks then in production, with two exceptions. The Institute for Progress estimated that China made virtually no HBM in 2025.
Sources (24)
Source tiers A company filings, governments, standards bodies, universities and research institutions. B SemiAnalysis, Reuters, the Financial Times and Nikkei Asia.
- ASK hynix, prospectus on Form 424B4 dated 10 July 2026
- AForeign-Produced Direct Product Rule Additions, and Refinements to Controls for Advanced Computing and Semiconductor Manufacturing Items
- AShould the US Sell Blackwell Chips to China?
- BChina’s CXMT Is Set to Challenge DRAM Incumbents
- AFiscal Q1 2026 Earnings Call Prepared Remarks
- AGroundbreaking SuperPoD Interconnect: Leading a New Paradigm for AI Infrastructure
- A15 CFR 740.25 -- License Exception High Bandwidth Memory (HBM).
- AHigh-bandwidth memory (HBM)
- AJEDEC® and Industry Leaders Collaborate to Release JESD270-4 HBM4 Standard: Advancing Bandwidth, Efficiency, and Capacity for AI and HPC
- AMicron, HBM3E product page
- ASK hynix Completes World’s First HBM4 Development and Readies Mass Production
- ASamsung Ships Industry-First Commercial HBM4 With Ultimate Performance for AI Computing
- AMicron, HBM4 product page
- AMicron Technology, annual report on Form 10-K for the fiscal year ended 3 September 2026
- ASK hynix Partners with TSMC to Strengthen HBM Technological Leadership
- AHBM3e and HBM4: IC design guide for next-generation high bandwidth memory
- A[Tech Note] Hybrid Bonding: Evolving into a Foundational Technology for Improving Semiconductor Performance
- ASamsung Develops Industry-First 36GB HBM3E 12H DRAM
- ASK hynix breaks ground on $4 billion advanced packaging production facility in Purdue Research Park
- ANVIDIA GB300 NVL72
- AInside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI
- AAMD Instinct™ MI355X GPUs
- A15 CFR 742.6, Regional stability
- AUnderstanding the Biden Administration’s Updated Export Controls