HBM (High-Bandwidth Memory)

Cross-cuts: ManufacturingCompute
last updated 2026-06-23 · +2 sources in last 30d
DRAMCoWoS (TSMC)HBM / CoWoS BottleneckAdvanced PackagingHybrid Bonding (Cu-Cu)TSV / Through-Silicon ViaHBM (High-B…

the memory denominator for Hbm Bottleneck. Research base: 2026 06 01 Hbm State Of Play (mid-2026), 2026 05 07 Hbm Supercycle Thesis Update.

What it is (physics / mechanism)

HBM stacks DRAM dies vertically over a base die, connects them with through-silicon vias (TSVs), and co-packages the stack next to the accelerator on a silicon interposer (2.5D). The short TSV interconnects replace long PCB traces, cutting energy per bit to ~3-4 pJ/bit (vs ~15 pJ/bit GDDR6).

The load-bearing point: bandwidth scales by bus width, not clock. HBM3E runs a 1,024-bit-per-stack interface; HBM4 doubles I/O to 2,048-bit, which is why throughput roughly doubles at HBM4 even at modest per-pin speeds. You widen the bus rather than chase frequency.

HBM only delivers its bandwidth when co-integrated on the interposer next to the GPU, and that interposer is TSMC CoWoS. So HBM is welded to advanced packaging: see HBM / CoWoS Bottleneck. Multiple analysts argue CoWoS, not the memory die, is the actual gating constraint (Nvidia booked ~55% of TSMC’s 2026 CoWoS capacity; lead times 50-104 weeks).

Generation roadmap

GenPin speedBandwidth / stackCapacity / stackStatus (mid-2026)
HBM2E~3.6 Gb/s~460 GB/sup to 16GBlegacy (A100 era)
HBM3~6.4 Gb/s~819 GB/sup to 24GBH100
HBM3E~8-9.6 Gb/s~1.2 TB/s24GB (8-hi) / 36GB (12-hi)mainstream 2024-2026
HBM46.4 Gb/s JEDEC base, vendors push 11-13~2.0-3.3 TB/s (pin-speed dependent)36GB (12-hi) / 48-64GB (16-hi)ramping 2026
HBM4E / C-HBM4E~12.8 Gb/s~3+ TB/slarger2027; ~40% of 2027 HBM demand (TrendForce)

The HBM4 per-stack bandwidth quoted as 2.0 / 2.8 / 3.3 TB/s is a pin-speed-quoting artefact (different vendors quote different speeds), not a contradiction. Samsung demoed 13 Gb/s → 3.3 TB/s at ISSCC Feb 2026; Micron ships >11 Gb/s → >2.8 TB/s.

At the accelerator level the stacks aggregate: an H100 (HBM3) reads memory at 3.35 TB/s, an H200 (HBM3E) 4.8 TB/s, a Blackwell B200 (HBM3E) ~8 TB/s (NVIDIA datasheets). A DDR5 channel, by contrast, moves only tens of GB/s, which is why AI accelerators pay the HBM premium at all, and why the The Memory Wall binds despite it (compute has outgrown even HBM, +60,000x vs +100x bandwidth over 20yr, 2024 Gholami Ai And Memory Wall).

The structural HBM4 shift: the passive base die becomes a LOGIC base die. This opens customer-specific customisation (controllers, near-memory compute) and turns HBM from a commodity into a design-win business. The competitive fault line is the foundry split: SK Hynix and Micron use TSMC (N12-class) base dies; Samsung uses its own SF4 (4nm) foundry base die. That split is why yields diverge (below).

The market and the supercycle

The suppliers (who is best placed, and why)

Market share (HBM revenue, Counterpoint Q3 2025): SK Hynix 57% · Samsung 22% · Micron 21%. 2026 bit-output projection: SK Hynix ~50%, Samsung ~28%, Micron the balance. CXMT is the rising fourth on wafers but a rounding error on competitive bits (below).

SK Hynix — best-placed overall (~75% confidence). Share leader, Nvidia’s primary HBM supplier (Nvidia ~15% of its revenue), sold out through 2026 into 2027. The durable edge is execution: ~80-90% HBM3E yields, MR-MUF packaging (<2°C thermal delta across a 12-hi stack), and the mature, low-cost, high-yield TSMC N12 base-die path carried into HBM4. Q1 2026 record sales; crossed ~$1T market cap. Weakness: lost the HBM4 first-to-ship headline to Samsung this cycle, and share is eroding off a 60%+ peak.

Samsung — best-in-class HBM4 performance, riskiest path. Won the HBM4 first-mover sprint on paper (passed Nvidia’s 11 Gb/s bar, demoed 13 Gb/s / 3.3 TB/s at ISSCC, began full HBM4 shipments Feb 2026, set to lead Vera-Rubin-specific HBM4 supply). The vertical-integration story (DRAM + own foundry + packaging) matters more as HBM4 moves to a logic base die. But SemiAnalysis is explicit: Samsung still lags SK Hynix on reliability/stability, and its in-house SF4 base die + 1c DRAM runs only ~50-60% yield (below ~60% breakeven), a margin/execution risk. Net: Samsung wins the headline, SK Hynix likely wins the volume.

Micron — solid #3, the US/Western hedge, and a 2026 surprise. At GTC (Mar 2026) Micron announced high-volume HBM4 36GB 12-high for Nvidia Vera Rubin (>11 Gb/s, >2.8 TB/s) and is sampling 48GB 16-high, using a low-cost internal CMOS base die. The only US-based supplier, so it carries the CHIPS-aligned structural angle (ties to Sovereignty Semi Bifurcation). 2026 supply fully booked; FY26 capex ~$20B; >50% gross margin. (Divergence flag: an earlier TrendForce note framed Micron as “limited to mid-tier Rubin CPX”; Micron’s own IR contradicts it. Trust the primary.)

CXMT, the China entrant: real on wafers, not yet on bits (~70% confidence the gap holds to 2028). SemiAnalysis models CXMT’s HBM wafer capacity at 5 → 30 → 55 → 100 kwspm (2025→2028), which “increase[s] the company’s share of global HBM wafer supply from 1% in 2025 to 12% in 2028” (2026 06 23 Semianalysis Cxmt Set To Challenge Dram Incumbents). The headline reads like a fourth force arriving; the yield reads otherwise. On technology readiness SemiAnalysis is explicit: “we believe CXMT is still struggling to stabilize supply for HBM3 8-hi, with even greater challenges in 12-hi.” It models 8-hi front-end/back-end yields at ~35% / ~70%, an overall yield of only ~25%, gated by back-end stacking (thermal stress, die cracking, warpage, bonding defects) and a low front-end wafer-sort yield, with HBM3 attempted without EUV. So 12% of HBM wafers by 2028 ≠ 12% of competitive HBM bits: most of CXMT’s stacks are lost or low-grade, and it remains ~3–4 years behind on the gen roadmap (HBM2E mass production targeted H1 2026, HBM3 in 2026, HBM3E in 2027). The competitive moat the leaders hold is back-end stacking yield, and CXMT’s struggle is the clearest evidence that it holds longer than the wafer-share chart implies. The live risk is sovereignty, not displacement: even sub-scale, low-yield domestic HBM relieves Chinese-AI memory dependence (Sovereignty Semi Bifurcation, Cxmt). Full maker volumes/revenue + the wafer-capacity chart sit on DRAM.

Routes (how to express it)

The bear case (what breaks the supercycle)

Companies using

Connected ideas

Sources

Related concepts

Frontier questions