Trends›HBM is the AI memory bottleneck
HBM is the AI memory bottleneck
High-Bandwidth Memory stays the binding AI memory-bandwidth constraint through 2030, growing from $2B (2023) to a >=$25B baseline (bull case $150B if it displaces DDR5/6), with SK Hynix holding leadership.
83 ↓1 confidence, computed 23 Oct 2026
- Technical
- High weight
Rests onfacts underneath
What rests on thismarkets it underwrites or threatens
Underwrites
Threatens
Evidencethe schedule the figure draws
The evidence, newest first
- Oct 2026
against · claim The same release says HBM4 shipments began in earnest in Q3, yield is ahead of plan, HBM4 will pass HBM3E in both volume and revenue in Q4, and investment will rise significantly next year as Yongin Phase 1 reaches full… SK hynix 3Q26 financial results release (PR Newswire, dated 27 Oct 2026) - Oct 2026
for · observable Record quarter: revenue KRW104.6tn (+32% q/q, +328% y/y), operating profit KRW83.3tn at an 80% operating margin (76% in 2Q26). SK hynix 3Q26 financial results release (PR Newswire, dated 27 Oct 2026) - Oct 2026
for · claim Micron's roadmap slide moves HBM to an annual cadence (HBM4 2026, HBM4E 2027, HBM5 2028, HBM5E 2029, HBM6 2030) and labels HBM "JEDEC + custom beginning with HBM4E", with vHBM and vEIM as customer-co-designed variants. Micron technology roadmap webinar (21 Oct 2026; TechTimes 22 Oct, tier 5) + Nomura memory note via Seoul Economic Daily (20 Oct 2026, tier 4) - Oct 2026
against · claim Supply response is now visible in the equipment order book: Lam puts calendar-2027 WFE above $200B against mid-$150B in 2026 with growth "led by DRAM", and memory was 46% of its systems revenue in the September quarter. Lam Research Q1 FY27 call (21 Oct 2026, transcript coverage tier 5) - 2027 WFE above $200B led by DRAM; Digitimes 19 Oct 2026 - Musk points to China's 2028 DRAM supply surge - Oct 2026
for · claim The second of the three DRAM suppliers to report this cycle printed a record quarter one week after Micron (KRW 107.4tn operating profit, +782.5% YoY, roughly +20% QoQ; revenue KRW 195tn, below the ~KRW 199tn consensus). Samsung Electronics preliminary Q3 2026 — operating profit KRW 107.4tn on revenue KRW 195tn, first quarter above KRW 100tn (Korea Herald, 8 Oct 2026) - Sep 2026
neutral · claim HBM is currently LOWER gross margin than conventional DRAM: the 2027 HBM price rises are described as "narrowing the gross margin gap with conventional DRAM", and the cloud unit margin (83%, flat) was held back by "higher HBM mix". Micron FQ4 FY26 prepared remarks (30 Sep 2026) - HBM margin vs conventional DRAM - Sep 2026
for · observable Micron "completed agreements for the vast majority of our calendar 2027 HBM bit supply with significant price increases year over year", and expects memory supply-demand "much tighter in calendar 2027 and 2028 than ... Micron FQ4 FY26 8-K press release + prepared remarks (30 Sep 2026) - Sep 2026
for · claim A seven-year-old stealth company disclosed $468M of private capital (GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford, Gates Frontier) plus a separate CHIPS R&D letter of intent for up to $245M, to build ferroelectric memory stacked on logic that it claims… Kepler Computing exits stealth with $468M for ferroelectric-on-logic AI memory built on retrofitted 28nm lines (Hardware Busters / Hoodline, 10-11 Sep 2026) - Aug 2026
against · claim Europe's largest-ever semiconductor round, 3 Aug 2026, for a company whose stated design choice is SRAM INSTEAD OF HBM — DX-1 targets the decode stage without high-bandwidth memory, explicitly to remove the advanced-packaging and HBM-allocation dependency, stitched by a photonic… - Aug 2026
against · claim The other direction on the same source, logged separately rather than averaged. HBF consortium widens — Google and Tenstorrent join; UCIe attach, 512GB/stack, 0.4-3.0 TB/s - Aug 2026
for · observable SK hynix and Sandisk published the first High Bandwidth Flash standard through the OCP on 4 Aug 2026 — but the roadmap in the same cycle's coverage puts the FULL spec in early 2027 and HBF SAMPLES in early 2028… First HBF standard spec published at FMS 2026 — but samples slip to early 2028 (SK hynix newsroom) - Jul 2026
against · claim The counter-current in the same story: Etched's 144GB/chip design uses HBM3E, the generation freeing up as Nvidia Rubin pulls HBM4, and SK Hynix keeps HBM3E mainstream through 2026 (TrendForce). Etched sits on HBM3E as Rubin migrates to HBM4 — trailing-generation relief (TrendForce/JPR) - Jul 2026
for · observable Etched ($10.3B post, ~$1.04B raised) has never disclosed an HBM supply contract; its observable supply mechanism is equity alignment — Rambus controller IP (Jun 2025), TSMC-linked VentureTech Alliance investing (Dec 2025), and SK Hynix joining the Series C cap table… Etched Series C — SK Hynix takes equity in a customer; HBM secured via cap table, not contracts - Jul 2026
for · primary SK Hynix's own CEO told Reuters on 10 Jul 2026 that 2027 will be the industry's WORST-EVER supply shortfall, with demand expected to outstrip supply 'even beyond 2030'. Situational Awareness LP case study (compiled from public reporting) - SK Hynix CEO (Reuters) + Goldman Sachs - Jul 2026
for · observable Private source Nathan Youngblood (Pitt) call — optical memory ruled out for AI, 3W optical compute vs 20W HBM - Jul 2026
against · claim A fourth DRAM entrant now holds RMB57.92bn (~$8.6bn) of fresh capital, a 2028+ supply-relief vector on the bottleneck. CXMT IPO war chest: $8.6bn raised, DRAM-not-HBM allocation - Jul 2026
for · observable CXMT listed 27 Jul 2026 on Q1-2026 revenue of RMB50.8bn, up over 700% YoY, and closed +466% at ~RMB3.3tn. CXMT STAR Market debut, +466% to ~$490bn (CNBC/SCMP) - Jul 2026
against · primary SK Hynix fell 15.4% in Seoul on 13 Jul 2026 (biggest one-day drop in ~two decades; ~17% intraday) on a brokerage warning of a Q3 OP miss — HBM4 shipments expected to scale from Q2 'do not appear to have… SK Hynix record one-day fall (-15.4% Seoul) as HBM4 goes three-supplier (Yahoo Finance / CNBC / TrendForce allocation trackers) - Jul 2026
for · primary Samsung Q2-2026 operating profit hit a record KRW 89.4tn (~$58.5-61bn), +1,810% YoY and +56% on Q1's own record, on revenue KRW 171tn (+129.3% YoY); the memory division alone generated ~KRW 90tn OP on DRAM/NAND ASP surge, and Samsung overtook Nvidia… Samsung Q2-2026 preliminary guidance — record memory profit (Korea Herald / KED Global, verified) - Jul 2026
for · primary HBM now consumes 23% of total DRAM wafer output (up from 19% in 2025) to produce ~9% of bits (~3x wafer per bit). TrendForce HBM wafer share + Nothing CMF cancellation + Apple price pass-through (Thursday Four cost-curve issue verification) - Jun 2026
for · observable Micron projects the HBM TAM reaches ~$100B by 2028 (up from ~$35B in 2025), far exceeding the $25B baseline. Micron HBM TAM ~$100B by 2028 (IG) - Jun 2026
for · observable Micron said its HBM supply for calendar 2026 is fully allocated/sold out with orders stretched to year-end. Micron HBM sold out CY2026 (IG) - Jun 2026
for · observable Micron and Anthropic signed a long-term memory/storage supply + joint-architecture pact (22 Jun 2026), with Micron a strategic investor in Anthropic's round - an AI lab integrating down the memory supply chain to secure scarce HBM/DRAM. Micron x Anthropic memory supply + co-design pact (via Semi Doped Daily) - Jun 2026
for · primary Jefferies projects memory pricing +40-50% QoQ in Q3 2026 and +30-40% in Q4 (~40-45% YoY into 2027) on zero wafer capacity growth; CSPs are locking 50% of capacity via 2-yr LTAs with 40% prepayment, squeezing out consumer/long-tail buyers. Jefferies memory pricing outlook (via Semi Doped Daily, 23 Jun 2026) - Jun 2026
for · primary DRAM AI-demand share rises from 12% (2023) to ~70% (2027) with HBM consuming ~3x the wafer per bit of commodity DRAM, and 1GW of deployed compute implies ~354K wafer starts (>60% memory). SemiAnalysis — To Boldly Go: The Case for Space Datacenters (Space DC TCO) - Apr 2026
for · observable SK Hynix co-developing HBM4 with key customers, sampling 7th-gen HBM4E in 2H 2026 with mass production targeted 2027. SK Hynix HBM4/HBM4E roadmap - Apr 2026
for · observable SK Hynix Q1 2026 revenue +144% YoY to KRW52.6T as memory prices rocketed, HBM the principal ASP driver. SK Hynix memory revenues explode as prices rocket (Blocks & Files) - Apr 2026
for · primary SK Hynix said customer HBM requests already exceed its planned production capacity for the next three years and called the upcycle 'structural, not temporary.' SK Hynix Q1 2026 call — demand exceeds 3 years of capacity - Mar 2026
against · primary Google published TurboQuant (25 Mar 2026), a training-free 6x KV-cache compression; within 24h SK Hynix fell 6.23% and Samsung 4.8%, Micron down >20% over six sessions — markets pricing a real software threat to HBM demand. Google TurboQuant pressures memory stocks (CNBC) - Feb 2026
neutral · observable TrendForce: SK Hynix keeps its 2026 HBM lead but Samsung recovers via HBM3e/HBM4 and Micron expands; growth momentum slowing on chip-upgrade delays and converging supply-demand. HBM Industry Analysis 1Q26 (TrendForce)
What would move itthe measures watched
- watchHBM sold-out status / book-coverage horizon. Capacity 'sold out' or booked years forward = bandwidth still binding; order horizon shortening = constraint loosening
- watchHBM ASP / pricing trajectory. Rising HBM ASPs confirm scarcity pricing; ASP roll-over (as HBM4 ramps, Samsung re-enters) signals supply catching up
- watchHBM TAM forecast revisions ($B by 2028/2030). Upward revisions toward/above the $25B baseline support the bull path; downgrades cut it
- watchPhotonic / exotic memory displacement of HBM. The statement's final clause (photonic memory and RRAM/FeRAM stay niche through 2030) had no metric until 29 Jul 2026. Niche-confirming: no photonic memory at HVM, density stuck at microring pitch, field practitioners rating it non-viable for AI weights. Disconfirming: a foundry-qualified photonic or emerging-NVM part shipping into an AI memory hierarchy, or a sub-diffraction addressing demo that closes the areal-density gap.
- watchSoftware demand-bending (KV-cache compression, quantization). Disconfirming axis: 5-6x KV-cache compression could bend per-token HBM demand — watch whether it shows in order books or is absorbed by larger batches
- watchHBM4 / HBM4E ramp + qualification timing. On-time HBM4 deepens the bottleneck and ASP step-up; slips compress the curve and open share to rivals