Micron in High-Volume Production of HBM4 Designed for NVIDIA Vera Rubin, PCIe Gen6 SSD and SOCAMM2
HBM4 36GB 12-high in high-volume production, shipping since Q1 CY2026: over 11 Gb/s per pin, over 2.8 TB/s per stack, 2.3x bandwidth and over 20% better power efficiency vs Micron HBM3E; HBM4 48GB 16-high at samples. Issuer claims.
How deeply this was read. The whole press release, but not as raw HTML: Micron's investor site blocks plain fetches behind a browser check, so it was read through a page-rendering fetch, and each number below was pulled out as an exact quoted sentence rather than a summary. Nothing here comes from secondary reporting.
Whose claim this is. Micron's, about Micron's own product, published on the day of NVIDIA's GTC conference. Every performance figure is written as a floor ("over", "greater than") rather than a measured value, and nobody outside Micron has checked any of them. Treat them as the maker's stated specification, as of 16 March 2026.
What the release says
The headline item, in Micron's words: "HBM4 36GB 12H in high-volume production, designed for NVIDIA® Vera Rubin — greater than 2.8 TB/s and with 20% better power efficiency."
The body gives the detail: "With HBM4, Micron achieves over 11 Gb/s pin speeds, enabling a bandwidth greater than 2.8 TB/s, representing a 2.3 times bandwidth and greater than 20% power efficiency improvement over its HBM3E."
And on timing: Micron "has begun volume shipment of its HBM4 36GB 12H in the first quarter of calendar year 2026 and is designed for NVIDIA Vera Rubin."
On the next, taller stack: "Micron has demonstrated advanced packaging capability of stacking 16 die of HBM by shipping samples of HBM4 48GB 16H to customers." and "This milestone delivers a 33% increase in capacity per HBM placement compared to the HBM4 36GB 12H offering."
Key figures
| Item | Figure | Status |
|---|---|---|
| HBM4 36GB, 12 dies high | over 11 Gb/s per pin; over 2.8 TB/s per stack | Volume production, shipping since Q1 calendar 2026 |
| vs Micron's own HBM3E | 2.3× the bandwidth; over 20% better power efficiency | Issuer comparison |
| HBM4 48GB, 16 dies high | 33% more capacity per placement than the 36GB part | Samples only, shipped to customers |
| Design target | NVIDIA Vera Rubin | Named in the release |
The release also announces two non-HBM products that are outside this rung's memory-bandwidth question and are recorded only for completeness: the Micron 9650 PCIe Gen6 SSD (in high-volume production; up to 28 GB/s sequential read) and 192GB SOCAMM2 server memory modules (in high-volume production; the release says up to 2 TB of memory and 1.2 TB/s of bandwidth per CPU).
What the release does not say
Asked for directly, the page contains nothing on: the DRAM process node, the base-die process, the interface width (2,048-bit), yields, price, capacity, market share, or whether 2026 output is sold out. So this is a capability-and-timing source, not a supply-or-price one.
How it ties to the rung's numbers
The rung's unit is the cost of memory bandwidth delivered, because in the decode phase — the part of answering where the model produces one token at a time — the chip mostly waits on memory. So the bandwidth a single stack delivers, and the energy it burns delivering it, feed straight into how many tokens a GPU can produce per second and per watt, and so into the token price.
A check we did ourselves (arithmetic, not a claim in the release): NVIDIA's own Rubin material puts 288 GB of HBM4 and 22 TB/s on each GPU (Vera Rubin platform). 288 GB is eight of these 36GB stacks, and eight stacks at "over 2.8 TB/s" is over 22.4 TB/s. The two vendors' numbers agree with each other: this is the part that fills a Rubin GPU.
The second thing it does is date HBM4's arrival in volume. Until now the rung carried HBM4 as a roadmap item — a January 2026 outlook put the taller 16-high 48GB HBM4 in production in Q4 2026 (SK hynix outlook). Micron says the 12-high part was already shipping in volume in Q1 2026, and that the 16-high part was sampling by March. SK hynix's own July results say its HBM4 mass shipments began in Q2 2026 (SK hynix 2Q26). Two of the three suppliers now say HBM4 is a shipping product in the first half of 2026.
Limitations
- Issuer claims, written as floors. "Over 11 Gb/s" and "greater than 2.8 TB/s" are minimums the maker chose to publish; the actual shipping figure could be higher, and nothing independent confirms either.
- The "20% better power efficiency" baseline is Micron's own HBM3E, not the industry's, and the release does not say what efficiency is measured in (energy per bit, per stack, per unit of bandwidth).
- "High-volume production" has no volume attached. No units, wafers or share are given, so it cannot move a supply estimate.
- Dated. This is the state on 16 March 2026. Cite it with that date.
- Read through a rendering fetch, not from the raw page, because of the site's bot check.