HBM Market Analysis (2026) — Pricing, Supply Chain & Forecast Dashboard
As of October 2026, track high bandwidth memory market dynamics. HBM3 ~$200/stack, HBM3E ~$300/stack, HBM4 ~$550/stack (est., 36GB 12-hi). SK Hynix took 50% of HBM revenue in Q2 2026 (Counterpoint Research). Compare HBM configurations across NVIDIA H100, H200, B200, AMD MI300X, MI325X, Google TPU v5p, and other AI accelerators. Track vendor market share, supply chain qualification signals, and revenue forecasts through 2027. Data available via free JSON API at /api/v1/hbm for programmatic access.
Comparing Generations (2022-2027) • Source: Official Company Specifications
Source: Silicon Analysts Projections
"The unit of compute is no longer the individual GPU. The emergence of the NVIDIA GB200 NVL72 and Google TPU Pods signals that system-level bandwidth prioritizes over individual chip peak."
Public shipment-mix estimates × per-SKU HBM content, published as EB ranges (1 EB = 109 GB). Not a training-vs-inference split — dominant SKUs are dual-use, so a workload attribution of HBM would be dishonest; this is the honest SKU-class / customer-type cut.
By customer type
AI HBM Demand Creates Memory Crisis
How HBM supply crisis is draining resources from consumer electronics and reshaping the memory market.
NVIDIA GPU Prices Double
HBM as a key cost driver behind GPU price escalation and supply shortages.
OpenAI & Google's $84B Custom Silicon War
HBM demand surge from custom silicon investment exceeding supply 1.4-1.6x.
NAND Price Explosion: How AI Demand Is Driving SSD Costs Higher
Broader memory market context — AI demand impacts across NAND and DRAM.
High Bandwidth Memory (HBM) has become the critical enabler—and bottleneck—for AI accelerator performance. Every major AI chip from NVIDIA's H100 to Google's TPU v5p relies on HBM stacks for the memory bandwidth needed to feed massive transformer models. Understanding HBM price trends, supply constraints, and technology roadmaps is essential for forecasting AI hardware costs.
The HBM market forecast shows demand growing faster than supply through at least 2026. Only three companies—SK Hynix, Samsung, and Micron—produce HBM, and capacity expansion requires 12–18 months of lead time. SK Hynix holds the dominant market share with early qualification on NVIDIA platforms, while Samsung and Micron compete for the remaining allocation. This supply concentration gives HBM vendors significant pricing power, with high bandwidth memory pricing up 20–40% year-over-year.
Each HBM generation delivers higher bandwidth and capacity at increased cost. HBM3 (used in H100) offers 819 GB/s per stack at 24GB capacity. HBM3E (used in H200, B200, MI300X) pushes to 1.18 TB/s with 36GB stacks. The upcoming HBM4, expected in 2026, will further increase bandwidth with a wider interface but at higher HBM3E cost per bit. The transition between generations creates pricing crossover periods where previous-gen inventory becomes more cost-effective for certain applications.
SK Hynix took 50% of HBM revenue in Q2 2026, ahead of Samsung (33%) and Micron (18%), per Counterpoint Research; its lead was built on early HBM3E qualification with NVIDIA. Samsung is ramping aggressively with improved yield on its HBM3E stacks after early quality issues. Micron entered the HBM3E market later but has secured allocation with several AI chip vendors. This tool tracks all three vendors' revenue, market share, pricing, and qualification status in real time.
Related: HBM Price History · Packaging Cost Model · Cost Bridge Chart · Price/Performance Frontier
The full free tier of the HBM dataset, in plain tables. The sourced feed (qualification status) is current as of October 2026; the leading indicators below are Silicon Analysts editorial estimates, last revised 2026-08-14 — free context with no single external source, labelled as such in the API’s provenance. Vendor share is Counterpoint Research’s Q2 2026 split of HBM revenue. Identical to what GET /api/v1/hbm returns without an API key. For the hand-curated pricing reference, see HBM memory pricing and specifications. There is no public HBM spot market, so the spot-estimate series was retired on 28 July 2026; the API still returns its frozen rows, marked as retired.
Who has passed whose qualification, by HBM generation and target customer.
| Supplier | Generation | Target | Status | Note |
|---|---|---|---|---|
| SK Hynix | HBM3 | NVIDIA H100 | volume_shipping | Industry-first HBM3 to NVIDIA; mass production began June 2022. |
| Micron | HBM3E 8-Hi | NVIDIA H200 | volume_shipping | Micron HBM3E 24GB 8-Hi for NVIDIA H200; volume production Feb 2024. |
| SK Hynix | HBM3E 8-Hi | NVIDIA H200 | volume_shipping | First in industry to mass-produce HBM3E; supply to customer from late March 2024. |
| Samsung | HBM3 8-Hi | NVIDIA H20 | qualified | Samsung HBM3 passed NVIDIA qualification, limited to China-market H20. |
| SK Hynix | HBM3E 12-Hi | AMD MI325X | volume_shipping | Prep to ship 12-Hi 36GB HBM3E for MI325X (forward-looking). |
| Micron | HBM3E 8-Hi | NVIDIA GB200 | volume_shipping | Micron 8-Hi HBM3E in GB200/B200; shipping from early 2025. |
| Micron | HBM3E 12-Hi | NVIDIA GB300 | volume_shipping | HBM3E 12H designed into GB300; Micron began volume production (reported Mar 2025). |
| Samsung | HBM3E 8-Hi | Broadcom ASIC | qualified | Improved 8-Hi HBM3E passed Broadcom pre-mass-production evaluation. |
| Micron | HBM3E 12-Hi | AMD MI350 | volume_shipping | Micron 12-Hi HBM3E (288GB) in MI350 series; mass production from Q3 2025. |
| Samsung | HBM3E 12-Hi | AMD MI350 | volume_shipping | Samsung HBM3E Shinebolt...stacks 12 layers of D1a nm high-k metal gate 24 Gb DRAM dies...found in AMD MI350X floorplan analysis (published Oct 8, 2025) |
| Micron | HBM4 12-Hi | NVIDIA Vera Rubin | volume_shipping | Micron (2026-03-16): HBM4 36GB 12-high in high-volume production, designed for NVIDIA Vera Rubin; volume shipments began in calendar Q1 2026. |
| SK Hynix | HBM4 12-Hi | NVIDIA Vera Rubin | volume_shipping | NVIDIA's CEO (2026-06-05): all three memory makers passed HBM4 qualification and are in mass production for Vera Rubin. SK hynix's Q2 2026 results report HBM4 mass shipments from the second quarter. |
| Samsung | HBM4 | Google TPU | qualified | Samsung is currently shipping HBM4 samples to key customers such as NVIDIA for qualification testing. Sources expect feedback to come back 'positive' as early as this month. |
| Samsung | HBM4 12-Hi | NVIDIA Vera Rubin | volume_shipping | Samsung (2026-02-12) began HBM4 mass production and shipped commercial 12-high products to customers; Chosun Ilbo the same day names NVIDIA as the first customer. |
| Samsung | HBM4 12-Hi | AMD MI400 | volume_shipping | Korea Herald (2026-07-24): AMD began Helios production with Samsung supplying the HBM4, 12-layer 36GB stacks on each MI455X; rack shipments start late Q3 2026. |
| Samsung | HBM3E 12-Hi | NVIDIA GB300 | volume_shipping | Korea Economic Daily (2025-10-30): Samsung's HBM shipments to NVIDIA began in Q3 2025; Samsung named no customer, and volumes were still being negotiated in October. |
| SK Hynix | HBM3E 12-Hi | NVIDIA GB300 | volume_shipping | TrendForce (2025-05-23, citing Deal Site and ZDNet Korea): SK hynix was already supplying 12-layer HBM3E for NVIDIA's GB300 with a dominant share of orders. |
| SK Hynix | HBM3E 8-Hi | NVIDIA GB200 | volume_shipping | TechInsights teardown of NVIDIA HGX B200 (2025-04-14): the Blackwell GPU carries SK hynix HBM3E in eight 8-high stacks. |
Editorial estimates (last revised 2026-08-14): upstream signals that move ahead of reported HBM revenue, with the methodology behind each.
| Indicator | Value | Signal | Source | Methodology |
|---|---|---|---|---|
| AI Server Demand Proxy | +38% YoY | bullish | Taiwan MOEA Export Orders | Calculation: (Dec 2025 Export Value / Dec 2024 Export Value) - 1. Tracks dollar value of orders received by ODM giants (Foxconn, Quanta) for AI servers. |
| Raw Chip Output Signal | +32% YoY | bullish | Korea MOTIE Chip Exports | Calculation: (Dec 2025 "Memory" Export Value / Dec 2024 Value). A divergence where Value > Volume indicates increasing Average Selling Price (ASP). |
Stack count, capacity, bandwidth and estimated per-accelerator HBM bill, live from the tracked dataset. Actively-procured accelerators (H200, Blackwell, MI325X/MI355X, HBM4-class, Trainium 3) are priced on the Aug-2026 contract basis — matching the cost model on the HBM pricing reference — while legacy accelerators (H100, MI300X, TPUs) keep their earlier tracked market-range vintages.
| Accelerator | Vendor | HBM Type | Stacks | Capacity | Bandwidth | Market Range | Suppliers |
|---|---|---|---|---|---|---|---|
| H100 | NVIDIA | 8hi | 5 | 80 GB | 3.35 GB/s | $1,200 - $1,400 | SK Hynix (~70%), Samsung (~30%) |
| H200 | NVIDIA | 8hi | 6 | 141 GB | 4.8 GB/s | $2,400 (Est.) | SK Hynix (~70%), Micron (~20%) |
| B200 | NVIDIA | 8hi | 8 | 192 GB | 8 GB/s | $3,250 (Est.) | SK Hynix (>60%), Micron (~25%) |
| B300 | NVIDIA | 12hi | 8 | 288 GB | 8 GB/s | $4,900 (Est.) | SK Hynix, Micron, Samsung |
| Rubin (VR200) | NVIDIA | 12hi | 8 | 288 GB | 22 GB/s | $4,500 (Est.) | SK Hynix (First 16-Hi demo) |
| VR300 | NVIDIA | 16hi | 16 | 576 GB | 16 GB/s | $4,500+ (Est.) | SK Hynix (Lead), Samsung |
| MI300X | AMD | 12hi | 8 | 192 GB | 5.3 GB/s | $1,100 - $1,300 | Samsung (>60%), SK Hynix |
| MI325X | AMD | 12hi | 8 | 256 GB | 6 GB/s | $4,350 (Est.) | Samsung (Primary), Micron |
| MI355X | AMD | 12hi | 8 | 288 GB | 8 GB/s | $4,900 (Est.) | Samsung ($3B deal signed) |
| MI455X | AMD | 12hi | 12 | 432 GB | 23.3 GB/s | $15,300 (Est.) | Samsung, SK Hynix |
| TPU v5p | 8hi | 4 | 95 GB | 2.7 GB/s | ~$750 (Est.) | Samsung, SK Hynix | |
| TPU v6e | 8hi | 2 | 32 GB | 1.64 GB/s | $400 - $500 | Samsung (>60%), SK Hynix | |
| TPU v7 | 12hi | 8 | 192 GB | 7.4 GB/s | $1,700 - $1,800 | SK Hynix (First), Samsung | |
| Trainium 3 | AWS | 12hi | 4 | 144 GB | 4.9 GB/s | $2,450 (Est.) | SK Hynix, Micron (pin speed 9.6 Gbps) |
Bandwidth, capacity and pin-speed envelope per generation, as currently tracked.
| Generation | Bandwidth | Capacity | Pin Speed | Note |
|---|---|---|---|---|
| HBM2 | 256 GB/s | 8 GB | 2.0 Gbps | Legacy (V100) |
| HBM2e | 460 GB/s | 16 GB | 3.6 Gbps | Standard (A100) |
| HBM3 | 819 GB/s | 24 GB | 6.4 Gbps | Mainstream (H100) |
| HBM3E | 1.2 TB/s | 36 GB | 8.0+ Gbps | Cutting Edge (H200) |
| HBM4 | >2 TB/s | 36 GB | High | 2026 volume part (12-Hi). 16-Hi 48 GB follows in 2027. |
Q2 2026 share of HBM revenue (Counterpoint Research, Sep 2026): the latest quarter of the series on the HBM market share dataset. Counterpoint rounds each share, so a quarter can total 101%. Share of bit shipments is a different basis.
| Supplier | Share | Basis |
|---|---|---|
| SK Hynix | 50% | Q2 2026 share of HBM revenue (Counterpoint Research) |
| Samsung | 33% | Q2 2026 share of HBM revenue (Counterpoint Research) |
| Micron | 18% | Q2 2026 share of HBM revenue (Counterpoint Research) |
Public sources only. Figures are estimates unless a source is linked. Forward-looking series (revenue forecast, supplier revenue, monthly bit demand) are not included here — see the API docs for access tiers.