HBM-segment revenue share in Q2 2026 stood at 50% for SK Hynix, 33% for Samsung, and 18% for Micron, with CXMT and other suppliers holding negligible positions, per Counterpoint Research's quarterly HBM tracker [S1].
The DRAM parent market in the same quarter grew 57% quarter-on-quarter and 385% year-on-year, with Samsung leading overall DRAM share at 38%, SK Hynix at 25%, Micron at 24%, and CXMT at 10% [S1]. Counterpoint's shipment versus revenue split shows that HBM revenue share is more concentrated than HBM bit share: SK Hynix took 50% of HBM revenue in Q2 2026 against 64% a year earlier, while Samsung's revenue share jumped from 15% to 33% in four quarters [S1].
Q2 2026 HBM revenue share by supplier
In the HBM revenue table for Q2 2026, SK Hynix at 50% remains the largest single supplier, but the lead over Samsung (33%) has compressed to 17 percentage points from 49 points in Q2 2025 [S1]. Micron held 18% in Q2 2026, down 3 points year-on-year from 21% in Q2 2025, while Samsung's climb from 15% to 33% in four quarters is the largest single-vendor shift on the chart [S1]. Counterpoint's HBM shipment table (units, not dollars) shows SK Hynix at 50%, Samsung at 33%, and Micron at 18% for the same quarter, a configuration broadly consistent with the revenue cut once Average Selling Price (ASP) mix per stack is accounted for [S1].
For comparison, the DRAM-only revenue table in Q2 2026 puts Samsung at 38%, SK Hynix at 25%, Micron at 24%, and CXMT at 10%, with SK Hynix down from 39% a year earlier as HBM allocation was prioritised within its wafer mix [S1]. CXMT rose from 4% in Q2 2025 to 10% in Q2 2026 in conventional DRAM, the fastest mover outside the top three, but the supplier has not yet appeared in Counterpoint's HBM revenue table [S1].
Why the share moved: HBM3E mix, HBM4 ramp, ASP per stack
Most 2026 HBM revenue is still HBM3E, with HBM4 expected to become visible in shipments in the second half of the year, per Counterpoint's Q2 2026 commentary [S1]. Bank of America estimates the 2026 HBM market to reach US$54.6 billion, a 58% year-on-year increase, while Yole Group's Chiplet Summit 2026 outlook puts HBM revenue closer to US$60 billion in 2026, up from about US$35 billion in 2025 [S2][S5]. The gap between these two outlooks is small relative to the segment's growth, and both anchor on the same AI-server demand curve.
Stack-level ASPs illustrate why revenue share diverges from bit share. As of October 2026, Silicon Analysts tracks HBM3 at about US$200 per stack, HBM3E at about US$300 per stack, and HBM4 at about US$550 per stack (estimated, 36GB 12-high) [S3]. SK Hynix's HBM lead is built on HBM3E qualifications with NVIDIA's Blackwell Ultra accelerators and on being the first HBM3E supplier for Google's TPU v7p and v7e, as cited in SK hynix's January 2026 market outlook [S2]. Samsung's 33% HBM revenue share reflects a stronger HBM3E shipment mix and the first HBM4 volumes flowing in H2 2026 [S1].
Capacity allocation, HBM bit demand, and who is squeezed
HBM and conventional DRAM share the same fabs and wafers, and manufacturers are reallocating wafer starts toward HBM, server DRAM, and enterprise SSDs, per IG's July 2026 coverage of the memory supercycle [S4]. TrendForce expects 2Q26 conventional DRAM contract prices to rise 58–63% quarter-on-quarter and NAND contract prices 70–75% quarter-on-quarter, with 1Q26 conventional DRAM already revised to +90–95% quarter-on-quarter and PC DRAM expected above +100% quarter-on-quarter [S5]. Data centres are projected to consume more than 70% of high-end memory output in 2026, which is the structural reason a tight HBM line is also a tight conventional DRAM line [S4].
Silicon Analysts' September 2026 HBM bit-demand cut estimates 0.35–0.69 EB for NVIDIA Blackwell-class SKUs in 2026, 0.03–0.44 EB for Google TPU-class, 0.05–0.23 EB for AMD Instinct-class, 0.06–0.09 EB for AWS Trainium, and 0.07–0.09 EB for NVIDIA Rubin [S3]. Aggregated by customer type, merchant GPU SKUs (NVIDIA, AMD, Intel) account for 0.51–1.09 EB and hyperscaler ASIC SKUs (TPU, Trainium, Maia, MTIA) for 0.13–0.57 EB in 2026 [S3]. Goldman Sachs forecasts HBM demand for custom ASIC-based AI chips to grow 82% and reach one-third of the market, the second leg of demand after the NVIDIA Blackwell ramp [S2].
Three-supplier comparison: who fits which buyer
SK Hynix (Q2 2026 HBM revenue share 50%) is the incumbent choice for buyers tied to NVIDIA H200/B200/Blackwell Ultra roadmaps and for Google TPU v7p/v7e, with HBM3E qualified and HBM4 in mass production since September 2025 [S1][S2]. Samsung (33% HBM revenue share) is the broadest second source, with HBM3E in volume and HBM4 shipping in H2 2026, and it leads parent DRAM share at 38% across the full memory portfolio [S1]. Micron (18% HBM revenue share) is the only US-headquartered HBM supplier, qualifies HBM3E on NVIDIA H200 and AMD MI300X/MI325X, and is the natural pick where US-sourced memory is a procurement requirement [S3].
For a 2026 HBM buyer, the decision is rarely price-only: allocation is the binding constraint. TrendForce's 2Q26 contract-price guidance (DRAM +58–63% QoQ, NAND +70–75% QoQ) means that a single missing allocation can idle an AI-server build line, which is why the choice of HBM supplier typically tracks the AI accelerator SKU rather than the lowest stack price [S5]. BofA's view, repeated in SK hynix's January 2026 outlook, is that SK hynix will sustain a total HBM market share above 50% through 2026 on HBM3E dominance, with Samsung and Micron splitting the residual volume and HBM4 share shifting in H2 2026 [S2].
Limits, risks, and what the share table does not show
Counterpoint's HBM revenue table covers only the three named suppliers (SK Hynix, Samsung, Micron) and rounds residual share into 0% for others, so any future CXMT or foundry-driven entry would have to clear both JEDEC HBM stack qualification and NVIDIA/AMD/Google vendor qualification before showing up here [S1]. Stack-level ASPs cited by Silicon Analysts (HBM3 US$200, HBM3E US$300, HBM4 US$550 estimated) are public-source estimates, not contract prices, and can diverge from actual contract ASPs by stack height, density, and binning yield [S3]. HBM4's 16-high hybrid-bonding roadmap implies 576 GB per-package peak capacity and 13.0 TB/s peak bandwidth, but commercial parts on those limits are not yet shipping in volume as of October 2026 [S3].
Yield is the structural choke point. HBM stacking requires wafers thinned to a fraction of normal thickness and thousands of through-silicon-via (TSV) micro-bonds per stack, with a single defect scrapping the unit, which is why HBM yields lag conventional DRAM yields and capacity cannot be ramped as fast as wafer starts can be reallocated [S4]. If HBM4 yields miss plan at any of the three suppliers, the 33% / 18% revenue shares for Samsung and Micron become the most exposed numbers in the table, and SK Hynix's 50% share would re-extend on relative outperformance [S1][S4].
The 2026 HBM revenue share map is therefore a three-supplier oligopoly with wafer allocation, not price, as the binding lever. Trackable signals into Q4 2026 include Counterpoint's Q3 2026 HBM revenue share update for confirmation of Samsung's HBM4 ramp, TrendForce's 4Q26 contract DRAM and NAND price guidance for any easing of the allocation squeeze, and BofA's revised 2026 HBM market-size figure against the US$54.6–60 billion range.