Daily Brief

Daily perspectives and analysis from institutional research
← Back to daily briefs
Daily Brief

2026-09-03 Daily Quick Read | Hilo Research

Summary

Global markets are navigating the deepening of AI infrastructure and macroeconomic divergence. AI computing demand is shifting from training to inference, accelerating the expansion of memory architecture, custom ASIC, and optical interconnect supply chains, but crowding in Japanese semiconductor positions has reached extreme levels. At the macro level, uncertainty over Federal Reserve forward guidance has pushed up rate hike expectations, while weak domestic demand in China has prompted a policy shift toward existing home sales and fiscal expansion; U.S. industrial activity remains robust, supported by AI investment, but tariff-driven front-loading poses risks of an inventory correction. On the corporate side, Meta's litigation settlement could unlock its product pipeline and catalyze a valuation rerating; digital advertising fundamentals are strong, but platforms are underperforming e-commerce due to capital expenditure pressure, while Bitcoin is attracting structural fund inflows as a hedge against currency debasement.

2026-09-0363 reports9 institutions
Published: Content updated:
01

Evolution of AI Memory Architecture and Emerging Storage Technologies

4 Related reports

Key views

Bernstein establishes an architectural hierarchy framework from on-chip SRAM to shared storage, noting that AI workloads are divided by compute, bandwidth, capacity, or latency constraints: inference prefill is compute-bound, while decoding is memory-bound because the KV cache grows with token count and concurrent users. Expanding context windows and RAG datasets will increase demand for DRAM, SSDs, HDDs, and new intermediate tiers.

Bernstein evaluates several emerging storage paths: HBF, led by SanDisk and SK hynix, aims to supplement HBM with NAND but faces difficulty bridging the performance gap; Samsung's zHBM faces thermal and hybrid bonding yield risks; NVIDIA's NVHBM could reshape supply chain value distribution, weakening traditional memory suppliers' bargaining power on base dies; PIM offers theoretical efficiency advantages, but ecosystem restructuring costs limit near-term penetration.

Goldman Sachs tracking shows South Korea's 8-month memory chip exports rose 290% year-on-year, with DRAM exports up 412%, the fastest pace since 2008. DDR4 spot prices carry a 43% premium over contract prices, and DDR5 carries a 16% premium, indicating solid near-term support for memory pricing. Meanwhile, Goldman Sachs is more positive on HBM, raising its forecast for SK Hynix's 2027 HBM pricing growth from 50% to 100%, implying 2027 HBM operating profit of USD 500 billion.

Goldman Sachs notes TrendForce raised its 3Q26 PC DRAM price growth forecast to +18–23% sequentially, with server DRAM maintained at +13–18%. However, affected by elevated customer inventories, mobile DRAM is expected to grow only 8–13% sequentially in 3Q26 and slow further in 4Q26, showing clear structural divergence.

Goldman Sachs found divergence through HBM material import proxy indicators: plastic film imports related to Samsung Electronics rose 134% year-on-year in 7, while epoxy resin imports related to SK Hynix fell 18% year-on-year, interpreted as a clear recovery in Samsung's HBM competitiveness.

Current market environment

KV cache growth in AI inference workloads has become the core memory constraint, driving the industry from sole reliance on HBM toward multi-tier storage architectures. The current DRAM and NAND markets show tight supply-demand dynamics, with South Korean export data hitting records and spot premiums elevated, but weak consumer electronics demand is slowing mobile DRAM growth.

Future market changes

The price spread between HBM and conventional DRAM continues to widen, and intermediate-tier technologies such as HBF are gradually moving toward commercialization.

Medium to long term

Triggers

  • Continuous expansion of AI large model context windows
  • Exponential growth in concurrent inference requests
  • Constrained HBM capacity expansion

Transmission channels

  • KV cache demand exceeding existing HBM capacity limits
  • Driving low-cost, high-capacity solutions such as NAND/HBF into the memory hierarchy
  • Altering the value distribution landscape among memory suppliers

Indicators to watch

  • Announcement of HBF product sampling and mass production timelines
  • Changes in the HBM ASP gap between SK Hynix and Samsung
  • Progress in the transfer of NVHBM base die design rights

Invalidation conditions

  • AI model compression technologies significantly reducing memory requirements
  • HBM overcapacity leading to a price collapse

Institutional disagreements

The extent of Samsung's competitiveness recovery in the HBM market

Different views

  • Based on material import data, Goldman Sachs believes Samsung's HBM competitiveness is clearly recovering, narrowing the gap with SK Hynix.
  • Market consensus still holds that SK Hynix maintains absolute dominance in HBM, and Samsung's increased material imports may merely represent R&D or small-batch trial production rather than a large-scale share reversal.

Opportunities and risks

AI-Dedicated Memory and Multi-Tier Storage Architectures

Consensus opportunity

During the inference decoding phase, the KV cache grows exponentially, HBM is in short supply with soaring prices, and there is a huge performance and cost gap between conventional DRAM and HBM, urgently requiring new architectures to fill it.

Potential beneficiaries

  • SK Hynix
  • Samsung Electronics
  • SanDisk
  • NVIDIA

Risks

  • zHBM thermal and yield performance falling short of standards
  • Difficulty overcoming the cross-tier performance gap for HBF
  • NVHBM standardization weakening memory vendor differentiation

Indicators to watch

  • HBM ASPs continuing to rise beyond expectations
  • Successful tape-out of emerging storage technologies
  • Changes in CSP procurement order structure
Related reports(4)

This content is compiled based on institutional research report views, is for research reference only, and does not constitute investment advice.

Zhejiang ICP No. 2022035445-5
Disclaimer: Market data, charts, indicators, research views, and other information provided on this website are intended solely for information display, research communication, and educational reference. They should not be regarded as personalized investment advice, securities recommendations, trading instructions, solicitations, or guarantees of return. While we strive to improve the reliability of our data and content, such information may still be subject to delays, errors, incompleteness, or untimely updates due to source differences, methodological limitations, system processing, or market volatility. Users should exercise independent judgment based on their own circumstances and bear all risks and responsibilities arising from the use of this website.

Settings

Sign in to view recent logins