INFLCT ← All theses

HBM MEMORY

Memory, not just logic, is the AI-compute chokepoint

70/100 conviction
▲ +2 strengthening

Last reviewed 17 July 2026 · our own structural assessment, not a price target

The thesis

High-bandwidth memory (HBM) is emerging as an adjacent — and possibly tighter — bottleneck to AI compute alongside logic and power. Each accelerator generation stacks more HBM (Nvidia's H100 carried 80GB; Nvidia specs its Vera Rubin GPU package at 288GB HBM4), and HBM consumes disproportionately more wafer capacity per bit than standard DRAM — TrendForce estimates HBM takes ~18% of DRAM wafer input for only ~8% of bits in 2025 — so HBM is quietly draining commodity DRAM capacity. The exposure sits with the three HBM makers and the wider DRAM value chain.

The case for

Tightness is confirmed by the latest primary results, and the outlook LENGTHENED. Micron's fiscal Q3 2026 (quarter ended 28 May 2026, announced 24 June 2026): revenue $41.5B, GAAP gross margin 84.6%, guiding fiscal Q4 to $50.0B at ~86%. Micron said it has 'already shipped over $1 billion in HBM4 revenue' (cumulative to date) and that its 'HBM4 12-high volume ramp is tracking twice as fast as HBM3E 12-high'. Critically, Micron now expects supply-demand conditions for both DRAM and NAND to remain tight 'beyond calendar 2027' — a one-year extension of the 'beyond calendar 2026' it guided in December 2025 — and says it does 'not have line of sight as to when memory supply will be able to catch up with increasing demand'. It has locked 16 take-or-pay long-term agreements running calendar 2026-2030, roughly $100B of remaining performance obligations and $22B of customer cash deposits. SK Hynix reported a company-reported 72% operating margin in Q1 2026 (23 Apr 2026) and said HBM demand exceeds its planned capacity for three years. The structural mechanism: TrendForce (2 Jun 2026) estimates HBM rising from ~18% of DRAM wafer input in 2025 toward 22% (2026) and 30% (2027) while accounting for only ~8%, 9% and 13% of bits — HBM consumes disproportionate wafer capacity and drains commodity DRAM.

The case against

The rate of price increase is decelerating off a record base — but the cause is not what it first looks like, and we previously got this wrong. Micron's blended DRAM average selling price (which INCLUDES HBM and an explicitly 'favorable mix', so it is not a conventional-DRAM contract price) rose in the 'low-60s percentage range' sequentially in fiscal Q3, versus mid-60s in fiscal Q2, and Micron guides that its fiscal Q4 gross-margin outlook 'reflects a meaningful moderation in the rate of price increases'. Micron does NOT attribute industry-wide moderation to its own contract ceilings, and we should not either. TrendForce (3 Jul 2026) attributes the slowdown to weaker consumer demand and a higher comparison base, forecasting conventional DRAM contract prices +13-18% QoQ in Q3 2026, and notes a portion of server DRAM procurement is governed by long-term agreements; its 9 Jul 2026 release adds that several US cloud providers have entered multi-year LTAs that restrict suppliers from raising prices for those clients — a multi-supplier, customer-driven mechanism, not a Micron-specific one. The sharper genuine risk is affordability: TrendForce says record contract prices mean PC and smartphone customers are 'reaching their affordability limit', with higher component costs weighing on PC and handset volumes. Concentration is also easing: SK Hynix's HBM share moved from ~62% (Q2 2025) to ~57% (Q3 2025) on Counterpoint estimates as Samsung reached HBM4 parity, and Nvidia's buyer power caps supplier pricing power.

What would change our mind

Re-scored 68 -> 70 on 2026-07-14; re-anchored 2026-07-17 on Micron's fiscal Q3 2026 (24 Jun 2026), which supersedes the December-2025 disclosure previously carried here. TWO CORRECTIONS TO OUR OWN REASONING: (1) the prior rationale capped the upgrade because 'the RATE of DRAM price increase is decelerating, which caps upside' — but in the same release Micron EXTENDED its tightness outlook a full year to 'beyond calendar 2027' (from 'beyond calendar 2026' in Dec 2025) while guiding revenue up ~21% sequentially at ~86% gross margin, so the deceleration is a second-derivative move off a record base, not a downturn. (2) We then over-corrected and claimed the moderation was 'self-inflicted' via Micron's own price ceilings — that causation is NOT supported: Micron never attributes industry-wide moderation to its contract structure. TrendForce attributes it to weaker consumer demand and a higher base, with multi-supplier LTAs a contributing mechanism. The honest bear is affordability-driven demand destruction in PCs/smartphones. Conviction held at 70 pending review — on the corrected evidence this score is arguably too low and is flagged for re-scoring.

Where the exposure sits

Exposure sits primarily with the three HBM makers: SK Hynix (KRX 000660, ~57% HBM share on a Counterpoint estimate for Q3 2025 and easing, South Korea), Samsung Electronics (KRX 005930, HBM4 in mass production and shipping since Feb 2026, qualified for Nvidia/AMD), and Micron Technology (NASDAQ:MU, USA). Second-order exposure runs through the broader DRAM value chain and HBM packaging/advanced-packaging supply. Note the durability risk is contestable and Nvidia's buyer power caps supplier pricing power.

Show our work — the evidence

The key evidence behind the score, dated and sourced — supporting and disconfirming. This is how the conviction is built, and what we re-check as it moves.

Tags mark how each fact is sourced: Company disclosure (the company said it) · Official data (a government or agency figure) · Analyst estimate (a third party's number) · Reported (via the press).

Get each thesis as it moves.