On this page · 10 sections
- HBM4E status on 7 November 2026: targets versus shipments
- How to read sample, qualification and mass production
- Which GPUs will use HBM4E
- Why HBM4E tightens supply rather than easing it
- What HBM4E means for GPU and cloud pricing in 2027
- Worked example: an 8-high switch, and booking before January
- India: what changes for Indian GPU buyers
- What to do now: a buyer's decision table
- How eCorpIT can help
- References
Summary. HBM4E is sampling, not shipping in volume. SK hynix told analysts on 29 July 2026 that it had finished HBM4E sample deliveries and aims for full-scale mass production in 2027. Samsung says it shipped the industry's first 12-layer HBM4E samples on 29 May 2026 (48 GB per stack, up to 16 Gbps per pin, up to 3.6 TB/s per stack), and Korean press reports of its 29 October call say mass production starts in the first half of 2027. Micron said in March that it expects to ramp HBM4E volume in calendar 2027, and in September that it had agreed most of its 2027 HBM supply at significantly higher prices. TrendForce forecasts blended HBM prices to rise 121% in 2027, with HBM4E ramping gradually from the second half. The first GPU shown with HBM4E is NVIDIA's Rubin Ultra, demonstrated with 1 TB per package and due in 2027. Our verdict: HBM4E will not ease GPU supply in 2027, and it will raise the cost of the parts that use it. Plan 2027 on HBM4-based Vera Rubin and Blackwell, book cloud capacity before AWS reprices Capacity Blocks in January 2027, and treat Rubin Ultra dates as targets.
About this article. For CTOs, infrastructure leads and finance teams planning 2027 GPU capacity. On 7 November 2026 we checked each supply, timing and price claim against SK hynix, Samsung and Micron results releases, prepared remarks and call transcripts, NVIDIA and AWS pages, TrendForce press releases and DGFT Policy Circular 08/2025-26 of 17 December 2025. Call-only statements are cited to the transcript or the Korean press that reported them, and TrendForce figures are labelled as forecasts. This is general information, not investment, procurement or legal advice; prices change, we have not tested the hardware, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
HBM4E status on 7 November 2026: targets versus shipments
Since SK hynix's July statement, Samsung has put a half-year on its target and Micron has named a custom design with NVIDIA. The table separates what has happened from what companies intend, using their own materials and TrendForce.
| Company | What has actually shipped | What is a target | Latest primary source |
|---|---|---|---|
| SK hynix | HBM4E samples to customers, completed in the second quarter of 2026 | Full-scale mass production in 2027 (Q2 call, 29 Jul 2026) | Q3 results and call, 28 Oct 2026 |
| Samsung | First 12-layer HBM4E samples to major customers, 29 May 2026 | Mass production in 1H 2027, volume ramp in 2H 2027 (Q3 call, as reported) | Q3 results release, 29 Oct 2026 |
| Micron | HBM4 ramp under way; no HBM4E shipment announced | HBM4E volume ramp in calendar 2027; custom HBM4E (NVHBM) with NVIDIA | FQ4 2026 prepared remarks, 30 Sep 2026 |
| NVIDIA | Vera Rubin, which uses HBM4, in full production | Rubin Ultra with 1 TB of HBM4E, due 2027 | Vera Rubin platform page; GTC 2026 demo |
| TrendForce (forecast) | Not applicable | HBM4E ramps gradually in 2H 2027 | Press release, 29 Sep 2026 |
Three sourcing notes. SK hynix's 2027 target comes from its July call as reported by Edaily; its written release says only that HBM4E samples shipped in the first half. On 28 October it said development and qualification are on customers' schedules. Samsung's half-year timing comes from its call, as reported by Money Today and Seoul Economic Daily (which called it the first specific mass-production date for this generation); the written results release says only that 2027 AI demand will far exceed industry supply. Micron gave a year, not a quarter, in March and has not updated it. None of the three has announced volume revenue shipments of HBM4E.
How to read sample, qualification and mass production
Memory makers announce the first step of a ramp loudly and later steps quietly. Only the last two rows change what a GPU buyer can rent or buy.
| Stage | What it means for a GPU buyer | HBM4E on 7 Nov 2026 |
|---|---|---|
| Samples shipped | GPU and ASIC makers can start validation; no volume | Done at SK hynix and Samsung |
| Qualification | The customer validates speed, thermals and yield on its own package | Under way; SK hynix says it is on customers' schedules |
| Mass production start | Volume wafer starts; first revenue shipments follow | Target 2027 (Samsung says 1H 2027) |
| Volume ramp | Enough bits for a GPU launch at scale | Target 2H 2027 (Samsung, as reported; TrendForce forecast) |
| Systems you can buy or rent | OEM servers and cloud instances ship | Rubin Ultra due 2027; Kyber rack timing disputed |
A slip at any stage moves every later one, so 2027 mass production and a 2027 GPU you can rent are different claims.
Which GPUs will use HBM4E
NVIDIA Rubin Ultra. At GTC 2026 on 16 March 2026, NVIDIA showed a Rubin Ultra compute tray. Tom's Hardware reported four compute chiplets and 1 TB of HBM4E per package, due in 2027 in the 144-package Kyber rack, and said it could not confirm tape-out. The German site heise reported from the keynote that Rubin Ultra starts in the second half of 2027. On 5 July 2026 the analyst firm SemiAnalysis reported that the Kyber NVL144 rack had slipped by more than 12 months to 2028; NVIDIA told Tom's Hardware only that its roadmap was intact. Our GTC 2027 buyer briefing covers that dispute, rack power and cooling.
Vera Rubin uses HBM4, not HBM4E. NVIDIA says Vera Rubin is in full production. Its NVL72 specification lists 20.7 TB of HBM4 across 72 GPUs, with 288 GB per Rubin GPU. Rubin Ultra's 1 TB per package is roughly 3.5 times that (1,000 ÷ 288 = 3.47), so HBM4E pricing matters more for it than HBM4 pricing did for Rubin.
Custom accelerators through NVLink Fusion. Micron says NVHBM, which it calls the first custom HBM4E implementation, will be adopted on NVIDIA's next-generation GPUs and NVLink Fusion platforms. NVIDIA describes NVHBM as a custom HBM base die carrying its own memory controller, offered through several memory partners. It claims up to 30% more bandwidth, 15% lower HBM power and up to 25% more compute die area than standard HBM4E, and named Amazon's Annapurna Labs as the first collaborator. SK hynix also said custom HBM, tuned to each customer's system, is growing in importance. HBM4E demand will not come from NVIDIA's own GPUs alone.
Everyone else. TrendForce says GPU and ASIC makers are weighing 8-high stacks for 2027 to stretch limited supply. Beyond NVIDIA and Annapurna Labs, the materials we reviewed name no accelerator tied to a specific supplier's HBM4E.
Why HBM4E tightens supply rather than easing it
HBM4E adds speed and capacity per stack, but four constraints mean it will not add enough bits to loosen 2027 supply.
Bigger dies and more process steps. On Samsung's 29 October call, memory executive Kim Jae-joon said HBM4E and HBM5 dies are larger and need more metal layers, and that moving bit output from conventional DRAM to HBM is a physical limit on supply, according to Digital Today. The same report says Samsung expects the shortfall to last through 2028.
HBM and server DRAM share the same wafers. TrendForce says AI server demand keeps HBM and conventional DRAM competing for limited advanced-process and wafer capacity. Micron expects industry DRAM bit shipments to grow only in the low-20s percentage range in both 2027 and 2028, with supply constrained in both years, and HBM bits to grow faster than conventional DRAM through 2028.
Demand is roughly double supply. SK hynix said on 28 October, in interpreted remarks, that customers want more than twice the HBM volume it can supply today.
Every stack needs a base die. Moving from 12-high to 8-high stacks means more stacks, base dies and back-end capacity for the same bits. SK hynix said it has built this into its plans. Samsung makes its HBM4E base die on its own 4nm foundry process, and its foundry business credited HBM base-die sales growth for its third-quarter improvement. Micron has said TSMC makes the base die for its HBM4E, so base-die supply also depends on foundry capacity.
New capacity is coming, but most of it lands during 2027 and ramps slowly. The timings below are the companies' own statements.
| Facility | Company | Output timing (company statement) |
|---|---|---|
| M15X | SK hynix | Ramping to full capacity by end-2026 for HBM4 |
| Yongin cluster, phase 1 | SK hynix | Full-scale operation in 2027; cleanrooms open in phases as demand allows |
| HBM advanced packaging, Singapore | Micron | Initial output early calendar 2027 |
| ID1 fab, Idaho | Micron | Wafer output from mid-calendar 2027 |
| Tongluo, Taiwan | Micron | Meaningful product shipments mid-calendar 2027 |
Micron's caveat applies to all of them: new-fab output becomes meaningful only a few quarters after first wafers.
What HBM4E means for GPU and cloud pricing in 2027
GPU list prices are not public, so watch memory contracts upstream and cloud rates downstream. The table draws on company statements, TrendForce forecasts and AWS's pricing page, current and archived.
| Signal | Figure | Source and date | Type |
|---|---|---|---|
| Micron 2027 HBM supply | Vast majority of bit supply agreed, with significant price increases year on year | Micron prepared remarks, 30 Sep 2026 | Contracted |
| SK hynix 2027 HBM supply | Volume and price for next year finalised with major customers | SK hynix Q3 call, 28 Oct 2026 | Contracted |
| Samsung 2027 HBM supply | Next year's HBM supply sold out at target prices | Samsung Q3 call, 29 Oct 2026, as reported | Contracted (reported) |
| HBM blended ASP | Up 121% year on year in 2027 | TrendForce, 29 Sep 2026 | Forecast |
| 8-high vs 12-high pricing | 8-high about 10 to 20% more per Gb | TrendForce, 29 Sep 2026 | Forecast |
| AI server revenue | Nearly $920 billion in 2027, up more than 90% | TrendForce, 14 Oct 2026 | Forecast |
| AWS P6-B300 Capacity Block | $16.146 per accelerator-hour, up 15% in October; next update January 2027 | AWS pricing page, checked 7 Nov 2026 | List price |
Read the rows together. All three memory makers have priced most of their 2027 HBM before HBM4E reaches volume, and each says the new prices are higher. SK hynix expects the profitability gap between HBM and conventional DRAM to narrow meaningfully, and Micron describes the same narrowing. SK hynix has also signed long-term agreements covering more than half of its total 2027 to 2028 capacity. TrendForce links memory shortages directly to higher prices for next-generation AI server racks, and says 8-high adjustments may ease some GPU cost pressure without reversing the HBM price rise.
Our reading, which is analysis rather than company guidance: memory is becoming a larger share of accelerator cost, and HBM4E parts carry the most memory. A package with 1 TB of HBM4E is far more exposed to a 121% rise in blended HBM prices than a 288 GB HBM4 GPU. Expect Rubin Ultra-class systems to carry a premium, and do not expect HBM4E to push down prices of current GPUs. AWS says its Capacity Block prices are updated regularly based on supply and demand; our analysis of AWS's 2026 Capacity Block price rises covers the FinOps response.
Worked example: an 8-high switch, and booking before January
Example A: memory per GPU with 8-high stacks. Samsung sells HBM4E as 48 GB 12-layer stacks and plans 32 GB 8-layer stacks, so each DRAM die holds 4 GB in both. Take a package with eight HBM sites (an assumption for illustration):
- Eight 12-layer stacks: 8 × 48 GB = 384 GB.
- Eight 8-layer stacks: 8 × 32 GB = 256 GB, one-third less memory.
- With TrendForce's forecast 10 to 20% per-Gb premium for 8-high, the memory cost index falls from 384 to between 256 × 1.10 = 281.6 and 256 × 1.20 = 307.2.
- That is a 20% to 27% cut in memory cost (307.2 ÷ 384 = 0.80; 281.6 ÷ 384 = 0.73) for a 33% cut in capacity.
- The same bits now fill 384 ÷ 256 = 1.5 times as many GPUs.
That is the trade TrendForce describes: more GPUs per wafer of HBM, less memory per GPU and a higher price per gigabyte. SK hynix notes that moving from 12-high HBM4 to 8-high HBM4E loses less capacity, because HBM4E dies hold more. Memory-bound models may need more 8-high GPUs for the same job, which erodes the saving.
Example B: booking a GPU block before AWS's January update. AWS lists an 8-GPU p6-b300.48xlarge Capacity Block at $129.168 per hour and charges the rate in force when you buy, even if the block starts after a price change. Figures below are the reservation fee only, before any operating system fee.
- Two-week training block: 14 × 24 = 336 hours; 336 × $129.168 = $43,400.45.
- If January's update were +15% (an illustration only, the size of October's rise from $112.32 to $129.168 an hour), the same block would cost $43,400.45 × 1.15 = $49,910.52.
- Booking before the update would save $49,910.52 minus $43,400.45 = $6,510.07 on one block.
AWS has not said which way January will go. The purchase date, not the start date, sets the price.
India: what changes for Indian GPU buyers
Cloud regions. AWS lists p6-b300 Capacity Blocks in Asia Pacific (Hyderabad), p6-b200 blocks in Mumbai and Hyderabad, and p5 blocks in Mumbai, at the same rates as its US regions. Indian teams can book guaranteed Blackwell capacity in-country at a known dollar rate. Our India GPU cloud pricing guide compares hyperscaler rates with domestic GPU clouds that bill in rupees.
Imported servers. Servers under HSN 8471 are restricted imports that need a DGFT Import Management System authorisation. Under the 2026 procedure, applications close on 15 December 2026 and every 2026 authorisation expires on 31 December 2026. By 7 November 2026 DGFT had not published the 2027 procedure; the previous two were issued in December. If GPU servers declared under HSN 8471 are due in early 2027, confirm your OEM or distributor will hold a 2027 authorisation.
Subsidised compute. For teams that qualify, the IndiaAI Mission's subsidised GPU capacity undercuts commercial clouds; our IndiaAI GPU empanelment guide explains access.
What to do now: a buyer's decision table
| If you are... | Do this | Why |
|---|---|---|
| Renting GPUs for 2027 training | Book Capacity Blocks or reserved terms before AWS's January 2027 update | AWS fixes the rate at purchase; HBM costs are rising |
| Waiting for Rubin Ultra | Plan 2027 on Vera Rubin and Blackwell; treat Rubin Ultra as an option | HBM4E volume is a 2H 2027 target; Kyber timing is disputed |
| Buying on-premises GPU servers | Get HBM stack height and GB per GPU in writing, with a price-validity date | TrendForce expects more 8-high parts in 2027 |
| Building on custom accelerators | Ask the vendor how its 2027 HBM allocation is contracted | Memory makers have sold most 2027 HBM already |
| Running inference at scale | Cut memory footprint with quantisation, KV-cache offload to SSD and batching | Every GB of HBM is getting more expensive |
| Setting a 2027 AI budget | Assume flat or higher accelerator costs | Contracts and forecasts both point up |
Dates to watch: NVIDIA's third-quarter fiscal 2027 results on 17 November 2026, GTC Washington, D.C. from 30 November to 3 December 2026 and GTC 2027 from 15 to 18 March 2027, DGFT's 15 and 31 December deadlines, AWS's January update, and the memory makers' next results, where any HBM4E slip would show first. TrendForce also forecasts conventional DRAM contract prices up 10 to 15% and NAND up 15 to 20% quarter on quarter in the fourth quarter of 2026; our 2027 server memory squeeze guide covers DDR5 and SSD buying.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team sizes 2027 GPU-hours, compares reserved, block and on-demand terms, and reviews server quotes for memory configuration and price validity. Talk to us at /contact-us/.
Last updated: 7 November 2026.
