報告_JPM_HBM市場_20260922

PDF 原檔：報告_JPM_HBM市場_20260922_original.pdf

圖片清單（已驗證 2026-09-24）

其餘 001-009（<40KB）為內文小型示意圖表（HBM S-D ratio、bit demand mix 等柱狀圖），依 trimmed 引用順序對應內文 Figure 1-9，未逐張放大驗證；內容與 trimmed 文字描述一致，供參考但不列入「已驗證真資料圖」清單。

原始內容

Memory Market Update

HBM S-D update; pricing tailwind and 63% 2026E-28E demand CAGR even after the de- pec

• HBM model revision: Three years rolling forecast largely unchanged, tighter 2026E-27E vs less tight 2028E Following the Asia Tech research team's CoWoS model u date last week, we revise our HBM industry model (higher 2026E-27E demand mainly on greater system unit sales demand vs. lower 2028E demand, reflecting content adjustment; 8Hi solution mix increase and 16Hi ush-out to 2029E at the earliest). On the other hand, three years' rolling accumulated HBM bit demand is largely unchanged, at 163bn Gb. While the HBM S-D shortage trend continues through our forecast eriod, we sense a slight HBM S-D trend im rovement from -20% in 2026E to -16% in 2028E. Reflecting the worsening accumulated S-D (over four months of shortage vs. rocurement demand), we revise u our ricing forecast and overall HBM TAM (US$160bn-282bn for CY27E-28E, re resenting 18-24% of the three global memory makers' overall DRAM revenue sales). On the HBM su ly side, we make a mild u ward revision in CY27E-28E, reflecting a higher HBM wafer allocation (es ecially SEC).

Figure 1: HBM S/D ratio (CoWoS adjusted)

Source: Company data, J.P. Mor an estimates.

Table 1: HBM estimate revisions

Source: J.P. Mor an estimates.

This document is being provided for the exclusive use of [讀者資訊已遮蔽].

Technology - Semiconductors

ay Kwon AC

(82-2) 758-5725

jay.h.kwon@jpmorgan.com

Sangsik Lee

(82-2) 758 5146

ang ik.lee@jpmorgan.com

J.P. Morgan Securitie (Far Ea t) Limited, Seoul Branch

Figure 2: Accumulated HBM bit sur lus and glut ratio (weeks) Weeks, 1GB equiv. Mn - RHS

Source: Company data, J.P. Mor an estimates.

www.jpmorganmarkets.com

• HBM bit demand continues to be above conventional DRAM, S-D constraint factor thesis intact Among the overall DRAM market, we ex ect HBM to show the secondhighest growth in the next two years (+63% CAGR), after AI-server grade DRAM/ SOCAMM (+92%). This is followed by non-HBM server DRAM (+37%) and PC/ Mobile (-3%). This trend can be consistently witnessed from the HBM bit demand mix continuously rising over the time (from 4% in 2025 to 12% in 2028E), resulting in greater wafer ca acity allocation. We forecast a 930k DRAM WSPM increase from 2025-end to 2028-end, and 58% of this incremental ca acity will be used to su ort HBM-grade frontend wafer fabrication (installed base antici ated to grow from 390k at 2025-end to 930k at 2028-end, or from 19% to 31% during the same eriod). As 16Hi ado tion is delayed, a meaningful wafer bit enalty scenario is unlikely; however, we ex ect 3x or above die enalty conce t to be largely sustained and continue to view growing HBM wafer allocation as a key differentiation factor to DRAM S-D in the AI-led H/W ca ex cycle.

Figure 3: Bit demand growth CAGR by a lication (2026E-28E)

Source: J.P. Mor an estimates.

7%

Figure 4: HBM % ca acity mix in DRAM

Kwfpm, % - RHS

3,000

93%

2,500

2,000

1,500

1,000

500

-

87%

13%

2023

2024

81%

19%

2025

77%

23%

100%

72%

28%

2026E

69%

31%

2027E

2028E

Total DRAM capacity HBM capacity

% allocation to HBM

% allocation to non-HBM

Source: Company data, J.P. Mor an estimates.

• Who will lead the 63% HBM consumption CAGR in the next two years? ASIC will become larger than NVDA in overall HBM consumption next year Des ite a decline in shares, NVDA continues to be the major layer in the HBM market (58% of total CY26E res ective HBM-grade AI chi 's combined HBM bit demand) until this year, with industry-leading HBM content (most leading-edge technology using higher layer and 8 cubes er system in ackage). From 2027, we ex ect ASIC (48% of total HBM bit demand) to take over NVDA's osition (43%) with a substantial system unit sales increase (ASIC: +102% y-y vs. NVDA: +15% next year). NVDA-grade HBM content ( er system in ackage) is likely to be continuously higher vs. ASIC; however, the content remium should shrink over the time (from 1.44x in CY26E to 1.33x/1.25x in CY27E/28E) with the otential s read widening from CY29E (the earliest timeline for 16Hi ado tion under consideration).

This document is being provided for the exclusive use of [讀者資訊已遮蔽].

80%

60%

40%

20%

0%

Figure 5: HBM bit demand mix by customer

%

Source: Company data, J.P. Mor an estimates.

Figure 6: NVDA vs. ASIC content and remium (x)

GB, x - RHS

Source: Company data, J.P. Mor an estimates.

• Tighter S-D and wafer economics balancing initiatives support a better-thanexpected HBM ASP outlook As we highlighted in our recent SEC (link) and SKH (link) 3Q26 review and u date re orts, we icked u stronger HBM ASP evidence vs. our initial ex ectation. We encil in a 54%/25% blended HBM ASP increase for our to - down HBM industry model and ex ect the HBM industry TAM to rise to US$282bn by CY28. Stronger ricing vs. conventional DRAM in the next two years should hel ut HBM back to one-quarter of the total DRAM TAM by CY28 (from 24% in 2025 to 10%/16%/22% in 2026E/27E/28E). In terms of rofit margin, we estimate it to be near a mid-60% to low-70% OPM across different su liers, below that for LTA-secured server DRAM (above 80%), but a shar im rovement from the revious year. As su liers re rice every year, HBM margins/ ricing of the non-LTA ortion could continue to rise until CY28E, in our view.
• Multi-tiering SKUs result in higher-than-expected 8Hi product shelf life, led mainly by NVDA We believe NVDA/ASIC customers established the universal s ecs for standard HBM and qualified su liers with minimum s ec requirements in the ast and laced roducible ca acity and cost structure as the key evaluation metrics. Moving forward, we ex ect to see multi le SKUs within standard HBM (not a customized version, but different segment tiering targeting slightly different workloads for various end customers). The 12Hi vs. 8Hi s lit within NVDA Vera Rubin Ultra is a ty ical exam le, and we have witnessed a similar a roach by other customers (MI450/455x, as well as TPU v8 Sunfish/Sunfish Ultra). 12Hi is likely to be the mainstream version for the erformance segment vs. the 8Hi, an economic version for the mass-tier segment. Due

Figure 7: HBM ASP and Y/Y% trend

Source: Company data, J.P. Mor an estimates.

Figure 8: HBM vs. DRAM OPM

Source: Company data, J.P. Mor an estimates. Note: Blended mar in from SEC/SKH

This document is being provided for the exclusive use of [讀者資訊已遮蔽].

to multi le technology validation issues and cost reasons, 16Hi ado tion has been ushed back to the 2029E line-u at the earliest, er our su ly chain checks. Therefore, we ex ect TCB to be the mainstream back-end ackaging technology solution for HBM over HCB in the foreseeable future.

Table 2: Diversifying HBM s ec across customers and SKU

Source: Company data, J.P. Mor an estimates for Rubin Ultra/Feynman/MI455 and beyond series/TPUv9.

• Intensifying competition among suppliers for value share gains amid undersupplied market conditions We have witnessed HBM value share s lit status move closer to that of the overall DRAM market value share mix, with followers (SEC and MU, which is covered by Harlan Sur) catching u against the leader (SKH) in the ast 12-24 months. We estimate that SEC and MU's combined share could rise close to 54% this year, followed by 59% next year (SEC commanding a greater share vs. MU u on larger ca acity, successful NVDA enetration and ASIC roject scale-u initiatives). While su liers are likely to take a slightly different a roach in addressing the HBM bit demand - SKH (strategic customer being NVDA and balanced bit roduction across different tiering of HBM layer roducts with a otential focus on 8Hi) vs. SEC ( riority to grab the largest HBM share in the market and scale u its wallet share within NVDA after successful HBM4 enetration) vs. MU (targeting to book equivalent share to that of overall DRAM market share) - HBM ASP negotiation is not largely im acted by the com etition, given a shortage market environment. We ex ect custom or bes oke HBM to be the next com etitive field, with roducts coming out as early as late 2028.

Figure 10: HBM revenue mix to DRAM by layer

This document is being provided for the exclusive use of [讀者資訊已遮蔽].

• Several different custom HBM designs in the pipeline, NVHBM likely to be based on 3nm logic die We are witnessing an increase in bes oke HBM rojects being discussed between memory makers and GPU/ASIC customers. The first official custom HBM roject is likely to be NVDA-led NVHBM (link), due to release in late 2028, followed by mass ram in 2029E. De ending on the customer, the degree of logic design freedom and res onsibility differs. While the NVHBM a ears to be based on 3nm logic die from TSMC (different from the diversified HBM4 logic die su ly chain), memory su liers are in charge of the final qualification incl. com atibility test with its core DRAM die. Gross margin er Gb may shrink due to an increase in cost (logic die gross margin lower vs. HBM-grade core DRAM die), but it remains clear that memory makers are contributing additional value to the overall system design, in our view. Com ared to NVDA, memory makers view ASIC rojects as ones that could command greater valueadd in designing the logic dies.
• Frontier AI model lab in-house chip demand not priced in, source for demand upside risk in the long term Leading AI model lab com anies (e.g., Anthro ic and O en AI) rely rimarily on datacenter ca acity lease business model and use a mixture of GPU/ASIC chi s for the silicon workload. However, there is increasing discussion in the su ly chain that model labs are ex loring the use of their own custom silicon chi s for s ecific workloads (link). While no details have been disclosed, we believe this could be an additional source of HBM bit demand on a long-term horizon. Given the rise in leading AI model lab com anies' direct memory rocurement su ly chain management, we believe it is a worth noting the develo ment to assess the otential demand u side to the memory industry. According to a local media release, HBM4/4E are the discussed HBM s ecs for leading AI model labs, and we sus ect the HBM density could be in the 144-288GB range, de ending on the layers (8hi or 12hi) and the number of cubes er system in ackage (6-, 8- or 12-cube configuration).
• Bullish on memory; SKH > SEC on imminent shareholder return upside After a shar correction during late June to late July (-36% vs. SOX: -19%), major memory stocks (SEC/SKH/MU) have rebounded 39% as of the 21 Se tember close (vs. SOX: +19%), attributable to risk-on a etite, with memory fundamentals seeing limited signs of weakness into 2027E. We believe the HBM de-s ec concern is somewhat riced in and reiterate our bullish view on memory, given: (1) tight HBM S-D, even after the de-s ec; (2) HBM customization leading to higher value-share gain at memory; and (3) resilient HBM ricing into 2027E-28E. In the near term, we retain our SKH > SEC view on constructive shareholder returns, but do see a favorable setu for SEC under the condition of constructive shareholder returns (i.e., raising the shareholder return ool, additional dividends/share buybacks, and strong commitment to shareholder returns).

Figure 11: Global memory makers' share rice erformance, including SOX (Philadel hia Semiconductor index)

Source: Bloomber Finance L.P.

This document is being provided for the exclusive use of [讀者資訊已遮蔽].

Companies Discussed in This Report (all rices in this re ort as of market close on 22 Se tember 2026, unless otherwise indicated)

SK hynix(000660.KS/W1,842,000/OW), SK hynix ADR(SKHY/$188.86[21 Se tember 2026]/OW), Samsung Electronics(005930.KS/W278,000/OW)

Analyst Certification: The Research Analyst(s) denoted by an 'AC' on the cover of this re ort certifies (or, where multi le Research Analysts are rimarily res onsible for this re ort, the Research Analyst denoted by an 'AC' on the cover or within the document individually certifies, with res ect to each security or issuer that the Research Analyst covers in this research) that: (1) all of the views ex ressed in this re ort accurately reflect the Research Analyst's ersonal views about any and all of the subject securities or issuers; and (2) no art of any of the Research Analyst's com ensation was, is, or will be directly or indirectly related to the s ecific recommendations or views ex ressed by the Research Analyst(s) in this re ort. For all Korea-based Research Analysts listed on the front cover, if a licable, they also certify, as er KOFIA requirements, that the Research Analyst's analysis was made in good faith and that the views reflect the Research Analyst's own o inion, without undue influence or intervention.

All authors named within this re ort are Research Analysts who roduce inde endent research unless otherwise s ecified. In Euro e, Sector S ecialists (Sales and Trading) may be shown on this re ort as contacts but are not authors of the re ort or art of the Research De artment.