What is HBM? The high bandwidth memory behind AI chips
HBM (high bandwidth memory) is stacked DRAM placed beside an AI chip. A guide to how it works and how to read HBM in SK hynix, Samsung and Micron earnings.
HBM (high bandwidth memory) is DRAM stacked vertically and mounted right next to an AI chip. Take Rubin, the Nvidia GPU (the workhorse chip for AI computing) that the company said in January 2026 was "in full production." Each one carries up to 288GB of HBM4, the newest standard. Memory bandwidth reaches up to 22TB/s, or 22 terabytes moved every second.
Yet none of the three companies that make this memory breaks out HBM revenue, based on the filings PRISM reviewed (earnings materials and a 10-K). SK hynix's third-quarter earnings presentation folds HBM into "Graphics (including HBM)" revenue. Samsung Electronics' Q3 investor relations (IR) deck has no HBM figure either. Micron's 10-K is the annual report US-listed companies file with the Securities and Exchange Commission (SEC). There, HBM is bundled with other DRAM in a single business unit's revenue.
As of November 7, 2026, this explainer is current. We'll update the earnings language and product specs whenever the three memory makers report quarterly results. The next updates will follow Micron's fiscal Q1 2027 results (December) and SK hynix's and Samsung Electronics' fourth-quarter results (January 2027).
What is HBM? DRAM stacked into a tower
SK hynix describes HBM as high-performance memory built by vertically connecting multiple DRAM chips. Instead of spreading houses across a lot, you build an apartment tower. One stacked bundle of DRAM is called a stack, and "12-high" means a stack of 12 DRAM dies.
The layers are wired together with TSVs (through-silicon vias). Tiny holes are drilled through each chip and filled with electrodes, so the chips above and below connect directly. SK hynix said in 2023 that it made its DRAM chips 40% thinner, so a 12-high 24GB HBM3 stack stood at the same height as its 8-high 16GB product.
At the bottom of the stack sits the base die, a logic chip that manages the DRAM layers above and exchanges signals with the GPU. Where that chip gets made differs across the three manufacturers.
"High bandwidth" means lots of data moves every second. Think of a highway: bandwidth is the number of lanes (I/O pins) multiplied by how fast each car goes (per-pin data rate). The HBM4 standard specifies 2,048 lanes at up to 8Gb/s per pin. Multiply the two (PRISM's calculation) and you get 16,384Gb (gigabits) per second. There are 8 bits (b) in a byte (B), so divide by 8 and you get 2,048GB (gigabytes) per second, or about 2TB/s for a single stack. Rubin's figure of up to 22TB/s covers all the HBM attached to one GPU, and Nvidia hasn't said how many stacks that is. Product speeds can run faster than the standard (Table 3), so you can't divide 22 by 2 to get the stack count.
The specs are set by JEDEC, the Arlington, Virginia-based semiconductor standards body. Here are the figures from its HBM3 standard announcement and HBM4 standard announcement.
Table 1. HBM3 vs. HBM4 under JEDEC standards
| Item | HBM3 | HBM4 |
|---|---|---|
| Standard | JESD238 | JESD270-4 |
| Announced | January 27, 2022 | April 16, 2025 |
| Interface width | – | 2,048-bit |
| Per-pin data rate | Up to 6.4Gb/s | Up to 8Gb/s |
| Bandwidth | 819GB/s per device (stack) | Up to 2TB/s per stack (about 2,048GB/s) |
| Independent channels | 16 | 32 |
| Stack heights | 4-, 8- and 12-high (provision for future 16-high) | 4-, 8-, 12- and 16-high |
| Max capacity | – | 64GB (16-high, 32Gb dies) |
"–" means the relevant JEDEC press release doesn't give a figure. HBM4 doubles HBM3's independent channels, the paths that carry data separately, and brings 16-high stacking into the standard.
You won't find HBM3E in the table. The name is all over earnings materials, but JEDEC's press release list (checked October 2026) has no standard announcement with HBM3E in its title. HBM3E is a product generation that memory makers developed as an extension of HBM3. According to Samsung Electronics, I/O pins doubled from 1,024 on HBM3E to 2,048 on HBM4.
When comparing HBM with ordinary DRAM, check the unit first. HBM bandwidth is quoted per stack. GDDR7, the graphics memory, is quoted per chip, while DDR5 for PCs and servers is usually quoted per module (DIMM).
Why HBM became the AI chip bottleneck
AI accelerators are chips, like GPUs, built to handle AI computation. In April 2026, IEEE Spectrum, the magazine of the Institute of Electrical and Electronics Engineers, described HBM as memory designed for AI processors. It also pointed out that Nvidia and AMD keep demanding more memory per chip. Newer accelerators have indeed raised HBM bandwidth or capacity per chip (Table 2). Among the materials PRISM reviewed, none directly explains why ordinary DRAM can't be used in HBM's place.
Table 2. HBM per AI accelerator (company-published specs)
| Accelerator | HBM generation | Capacity | Bandwidth | Stack height | Stacks |
|---|---|---|---|---|---|
| Nvidia Blackwell Ultra | HBM3E | 288GB | 8TB/s | – | Sources conflict (note 1) |
| Nvidia Rubin | HBM4 | Up to 288GB | Up to 22TB/s | 12-high | Not stated (note 2) |
| AMD Instinct MI355X | HBM3E | 288GB | Up to 8TB/s | 12-high | 8 |
| AMD Instinct MI455X | HBM4 | Up to 432GB | Up to 23.3TB/s | Not stated | 12 |
Note 1: Nvidia's technical blog post and its later correction notice describe the HBM configuration differently, so we've left stack height and stack count blank. Note 2: Nvidia says only that it uses 12-high stacks. Nvidia figures are company-published specs and may change.
Nvidia's bandwidth jumped from 8TB/s on Blackwell Ultra to up to 22TB/s on Rubin. AMD's capacity grew from 288GB to 432GB. In January 2026, Nvidia said Rubin was "in full production," with partner products due in the second half of 2026.
HBM sits in the same package as the GPU, right beside it. The best-known technology for that pairing is TSMC's CoWoS (Chip on Wafer on Substrate). It mounts the compute die and several HBM stacks side by side on a silicon base, called an interposer, so they work like a single chip. TSMC's CoWoS page explains: "By integrating multiple SoCs and high-bandwidth memory stacks, CoWoS® empowers products with high-performance compute and memory bandwidth." AMD also says its MI455X uses a CoWoS-L package.
The amount of HBM per package keeps climbing. At its North America Technology Symposium on April 22, 2026, TSMC said it will produce CoWoS packages in 2028 that hold about 10 compute dies and 20 HBM stacks.
Where exactly the bottleneck sits is a matter of debate. On TSMC's Q3 earnings call on October 15, management said "we have a little bit shortage on the back end." The back end is the stage where chips are assembled into packages. In TSMC's own answers on that call, the word HBM never came up. It appeared once, in a question from a Citigroup analyst.
In January 2026, Taiwan's Economic Daily News (經濟日報) listed three supply-side weak spots together: advanced packaging, glass substrates and HBM. Memory makers have their own take. In its Q3 earnings release, SK hynix said it "expects demand to keep exceeding supply through 2030" (translated from the company's Korean-language release).
Who makes HBM: SK hynix, Samsung Electronics, Micron
Three companies make HBM: South Korea's SK hynix and Samsung Electronics, and US-based Micron. According to a June 2026 report by Korean tech daily Electronic Times (Etnews), Nvidia CEO Jensen Huang said all three had passed HBM4 qualification testing. Qualification is the process a customer uses to validate a product before taking deliveries.
Both Korean makers claim a "first," each on its own terms. On September 12, 2025, SK hynix announced it had completed HBM4 development and "prepared world-first mass production of HBM4." In February 2026 (per reports), Samsung Electronics said it had begun HBM4 mass production and "shipped commercial products to customers." Samsung called this a "first in the industry." SK hynix counts from mass-production readiness; Samsung counts from commercial shipments.
Table 3. HBM4 product figures from the three makers (company announcements)
| Item | SK hynix | Samsung Electronics | Micron |
|---|---|---|---|
| Announced | September 2025 | February 2026 (per reports) | June 2025 (samples) |
| Speed / bandwidth | Over 10Gbps per pin | Consistent 11.7Gbps per pin, up to 13Gbps (up to 3.3TB/s per stack at that speed) | Over 2.0TB/s per stack |
| Stack height / capacity | 16-high 48GB on display (April 2026) | 12-high 24–36GB; 16-high up to 48GB | 12-high 36GB in volume production; 16-high 48GB samples (10-K) |
The JEDEC HBM4 standard tops out at 8Gb/s per pin and 2TB/s per stack. Samsung Electronics and SK hynix both say they've beaten it on per-pin speed. Samsung says its 11.7Gbps is "exceeding the industry standard of 8Gbps by approximately 46%." Micron's figure of more than 2.0TB/s per stack is on par with the standard's maximum. Announcement timing and measurement conditions differ, so the three numbers can't be used to rank the companies.
Base dies split two ways: some makers outsource them to a foundry (a contract chipmaker), others make them in-house. In April 2026, SK hynix displayed an HBM4 at a TSMC event, labeled as using a base die built on TSMC's advanced logic. Samsung Electronics says it uses a 4nm logic base die, and its Q3 foundry commentary cites "HBM B-Die" sales. Micron states in its 10-K that its base logic die and DRAM core dies are "designed and manufactured in-house." Say "from HBM4 on, foundries make the base die," and you leave Micron out.
Market share isn't something the companies disclose. A Counterpoint Research forecast, reported by the Korea Economic Daily in February 2026, puts 2026 HBM4 market share at 54% for SK hynix and 28% for Samsung Electronics. The report says only that the two will take more than 80% between them; it gives no separate figure for Micron. These are research-firm estimates, and methods and timing vary from firm to firm.
How to read HBM in earnings reports
A single accelerator carries up to 288–432GB of HBM (Table 2), yet none of the three companies breaks out HBM revenue in its filings. The filings alone don't show how much that revenue is or what share of sales it makes up. How far the HBM4 transition has come can be gauged from company statements, such as when HBM4 will overtake HBM3E and when capacity expansions are scheduled.
SK hynix: generation shift, capacity, contract terms
SK hynix reported third-quarter results on October 28. In the revenue-by-application legend on page 16 of its earnings presentation, HBM falls under "Graphics (including HBM)." There's no standalone HBM revenue or share. Instead, the company offers the following statements (PRISM translations from its Korean-language release and presentation):
- Shipment stage: "With HBM4 shipments ramping in earnest, a sharp rise in server DDR5 sales drove DRAM growth."
- Generation shift: "The company expects HBM4 to overtake HBM3E in both shipments and revenue in the fourth quarter, becoming its main product."
- Capacity: It plans to "bring M15X to full capacity by year-end to increase HBM4 production." M15X is an SK hynix production site.
- Contract terms: Page 11 of the presentation reads: "2027 HBM contract negotiations concluded; profitability gap between HBM and conventional DRAM expected to narrow meaningfully."
The contract-terms line doesn't say how much the gap will shrink, or which side's movement will close it. Revenue and operating profit figures are in our SK hynix Q3 story.
Samsung Electronics: in both memory and foundry
Samsung Electronics' third-quarter earnings presentation is an IR document published before the external review was completed. It notes: "The review outcomes may cause some parts of this document to change." Memory revenue for the third quarter (July–September) came to KRW 140.7 trillion (out of KRW 147.1 trillion for the Device Solutions, or DS, chip division), with no HBM share given. HBM appears in these lines:
- Memory, Q3: "Significantly scaled up HBM4 sales backed by industry-leading product competitiveness." The "industry-leading product competitiveness" it cites as the reason is the company's own assessment.
- Memory, Q4 outlook: "Further increase sales of HBM, Server DRAM and SSDs."
- Foundry, Q3: "Substantial revenue and earnings improvement on HBM B-Die sales growth and expanded advanced-node sales." The presentation doesn't define B-Die. PRISM reads this as the HBM base die discussed above.
We covered Q3 chip profit figures in 5 Things You Missed This Week (October, week 5).
Micron: what the 10-K says
In its 10-K, Micron books HBM under its Cloud Memory Business Unit (CMBU). That revenue line also includes DDR, LPDDR and GDDR, with no separate HBM figure. For fiscal 2026 (ended September 3, 2026), it describes HBM in words:
- "HBM3E 12-high represented the majority of our HBM shipments."
- Micron "began volume production of our HBM4 36GB 12-high" and "delivered samples of HBM4 48GB 16-high to multiple key customers."
- It expects to "ramp volume" of next-generation HBM4E "in calendar 2027."
Transition timing and reporting periods
SK hynix expects HBM4 to overtake HBM3E in both shipments and revenue in the fourth quarter of calendar 2026. Micron wrote that HBM3E 12-high made up most of its HBM shipments in fiscal 2026. The two statements cover different periods, so they can't be compared head-to-head. In January 2026, Taiwan's Economic Daily News reported market watchers' expectations that HBM3E would remain the mainstream product in 2026.
Four lines to check next earnings season
- SK hynix: Do the Q4 results confirm that HBM4 actually overtook HBM3E in both shipments and revenue? If they do, the company's Q3 forecast has been borne out; if they don't, the company may have missed that forecast.
- SK hynix: How did the 2027 HBM contract negotiations end, and how does the "profitability gap" language change? If the results report that the gap narrowed, the company's forecast will have held up.
- Samsung Electronics: Does the Q4 memory commentary include a results line saying HBM sales rose? Does the qualifier in front of HBM4 sales ("significantly" and the like) stay, and does the foundry's "HBM B-Die" line continue? If higher sales are reported as a result, the Q4 outlook has been met by the company's own account; if the qualifier or the B-Die line drops out, the company's description has changed.
- Micron: Does it mention HBM4's share of HBM shipments, and does the line on ramping HBM4E volume in 2027 hold? If Micron says HBM4 made up the majority of its HBM shipments, the company's own wording confirms the switch in its main generation; a change to the HBM4E timing would mean the plan has been revised.
What to watch
- December: Micron fiscal Q1 2027 results. Check item 4 above.
- January 2027: SK hynix and Samsung Electronics Q4 results. Check items 1–3.
- Nvidia Rubin shipments. Rubin carries up to 288GB of HBM4 (Table 2). In its fiscal Q2 2027 earnings press release on August 26, 2026, Nvidia said the Vera Rubin platform is "ramping into full production." It added that racks are running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure (OCI) and Nebius. On October 27, Nvidia and Oracle announced the general availability of Vera Rubin NVL72 on OCI. The August release doesn't include shipment numbers. Next up is Nvidia's fiscal Q3 2027 earnings call on November 17 (US time). Compare its wording with August's "ramping into full production."
- Next-generation standards. JEDEC announced the SPHBM4 standard in July 2026. Going by the press release title, it's a standard for delivering HBM4-class bandwidth on organic substrates. We haven't yet found an HBM4E standard announcement in JEDEC's press release list (as of October 2026).
This content is AI-generated based on source articles. While we strive for accuracy, errors may occur. We recommend verifying with the original source.
Related Articles
Korea's AI Basic Act fines a missing AI label only after an ignored order. In 5 Korean and 4 Taiwanese laws PRISM checked, a missing AI label or disclosure draws a fine from the first offense in two places: election deepfakes and Taiwan ad platforms.
Korea's October exports rose 93.7% to $115.28 billion. Apple said its 47.9% gross margin included about 1 percentage point from tariff refunds. Plus China's draft tech export list.
EU AI Act deepfake labeling has applied since Aug. 2. Dec. 2 ends only a narrow marking grace period and starts a conditional ban on non-consensual sexual-deepfake AI.
SK hynix posted its first 100-trillion-won quarter, with an 80% operating margin. Micron's GAAP margin was 1.0pp higher; TSMC and Samsung were in the 50% range.
Thoughts
Share your thoughts on this article
Sign in to join the conversation