報告_JPM_台灣散熱零組件_20260930

PDF 原檔：報告_JPM_台灣散熱零組件_20260930_original.pdf

圖片清單（已驗證 2026-09-30）

ingest 時建立的「眼見為憑」圖片索引，是 lib/ 嵌圖的唯一真相來源；嵌入時只從這裡挑分類為「真資料圖」的，不照 trimmed 引用順序猜。

其餘無 ≥40KB 候選圖片（本份報告僅抽出 1 張圖）。

原始內容

Taiwan Cooling Components

Accelerating cooling upgrades in Nvidia VRU and ASICs to unlock continued TDP growth

We see an accelerating cooling spec upgrade trend in both AI ASICs and GPU syste s, and the content uplift looks larger than investors' expectations, especially for Nvidia's next-gen VR Ultra (VRU), contrary to recent arket concerns on potential de-specing. We continue to believe that cooling is the key bottleneck to unlocking continued TDP growth of all AI chips and igration of AI server syste s. For VRU, we see new designs and spec upgrades across cold plate (double-sided design and diamond copper material adoption), quick disconnect (MQD upgrading to ZQD), and lid/heat spreader (3-piece/ removable lid design). We believe the stronger-than-expected content growth in Nvidia VRU, along with the further increasing ASIC and AMD GPU cooling value will continue to benefit the leading suppliers of each co ponent. Looking forward to the next 1-2 years, we believe the CPO/NPO switches and higher-spec optical transceivers/ odules will beco e the next driver with ainstrea adoption of liquid cooling in future generations (likely fro 1.6T/3.2T gens). We prefer AVC (OW) and Fositek (OW) in the cooling space.

• old plate: double-sided design likely adopted in Nvidia VRU, AWS Trn3 and AMD MI450 to drive +50% content; diamond copper to add value. The double-sided cold plate design looks set to beco e the next step-up in cooling specs. We expect adoption in AWS Trainiu 3 (Trn3) and AMD MI450 co pute trays, with Nvidia VRU likely next. This new double-sided design was required given the VPD (Vertical Power Delivery) architecture, which oves the power discretes, SOCAMM, and capacitors to the backside of chip board, requiring incre ental cold plates for backside cooling. VRU Bianca cold plates ay also upgrade their aterial to 'dia ond copper' for better cooling effectiveness, which adds further value per unit. This co bination supports eaningful cooling content upside for VRU (~50% increase vs. VR200) versus arket expectations for flat-to-down cooling content on de-spec concerns. Outside these platfor s, Google TPU v8 should drive solid cooling content via gold-plated cold plates plus additional units adopted for surrounding ICs. TPU v9's larger chip die and higher TDP also increase the odds of the double-sided cold plate adoption.
• Quick disconnect (QD): No content change with the potential VRU 9switch tray design of 1.5U with 'double-layer architecture'; meaningful upgrades in VRU (ZQD adoption) with ~100% content increase per rack. In the VRU syste , we believe Nvidia will upgrade so e of its QDs fro the previous MQD (Multi Quick Disconnect)/UQD (Universal Quick Disconnect) to the higher-spec ZQD (Zero-drip/flow rate Quick Disconnect) with ~2x ASP, in order to help i prove the cooling efficiency with better liquid flow rate and li ited leakage. Per JPMe, the total QD content per VRU NVL72 rack will increase by ~100% vs VR200/GB300 racks, as the ZQD with 2x ASP will be adopted in 1) the co pute tray Bianca odules (upgrading to ZQD06 spec along with GPU cold plate upgrades), 2) entry QDs for all 9 co pute trays and 9 'double-layered' switch trays (the 2 out-of-tray plug QDs upgrading to ZQD14 spec), and 3) rack anifold socket QDs (upgrading to ZQD14 as well).
• Heat spreader: No delay of 3-piece/removable lid adoption in Rubin Ultra, even pulling forward some Rubin GPU potentially from 4Q26/1Q27. We believe Nvidia's Rubin Ultra GPU will adopt the 're ovable lid + stiffener +

See page 9 for analyst certification and important disclosures, including non-US analyst disclosures.

Technology

Megan Hsueh AC

(886-2) 2725-9249 megan.hsueh@jpmorgan.com .P. Morgan Securities (Taiwan) Limited

okul Hariharan

(852) 2800-8564 gokul.hariharan@jpmorgan.com .P. Morgan Securities (Asia Pacific) Limited/ .P. Morgan Broking (Hong Kong) Limited

Albert Hung

(886-2) 2725-9875 albert.hung@jpmchase.com

.P. Morgan Securities (Taiwan) Limited/ .P.

Morgan Securities (Asia Pacific) Limited/ .P.

Morgan Broking (Hong Kong) Limited

ultiple screws' design (or the so-called 3-piece lid design) with no delay, which should drive the heat spreader/lid content increase by 4-5x vs the current 1-piece lid design for Rubin GPU. Based on our industry check, Nvidia ight even adopt this new design in s all volu e of Rubin GPU potentially fro 4Q26 for trial production and enter ass production in the next gen Rubin Ultra GPU potentially in 1Q27. We believe the adoption schedule is slightly ahead of arket expectations, which ight provide near-ter rev/ earnings upside for the 3-piece lid supplier. In addition, AMD MI450 series adoption of 2-piece lid (~10x ASP increase vs MI350 series), future CPO/NPO ra p with higher stiffener content, and the potential ASIC spec upgrades in the longer ter are structural trends for the heat spreader arket.

• Stock implications. We prefer AVC (OW) in the cooling space given the stronger cold plate spec upgrade trend in both GPU and ASIC syste s with its leading position and arket share in each project, enabling the co pany with better-than-peer argin and earnings growth with unde anding valuation (and underperfor ance vs peers). We also like Fositek (OW) given its near-ter GM upside and longer-ter QD upgrade/content growth story with continued share gains and TAM expansion.

Investment thesis

We believe the cooling co ponent sector will continue to outperfor the arket, and in this space, we prefer AVC and Fositek. Co pared to other cooling na es, these two na es 1) still trade at an unde anding valuation (vs historical) with near-ter share price perfor ance underperfor ing other cooling na es (or even Taiex for AVC), 2) can continue to deliver both near-ter and longer-ter earnings upside (e.g. 3Q argin beat, 2028/28E EPS upside fro ore eaningful spec upgrades), and 3) serve as the leading players in the specific co ponents, enabling the ajor beneficiaries of spec upgrades and ongoing argin expansion (with s aller-than-expected ASP pressure).

Table 1: We prefer AV and Fositek in the liquid cooling space

Source: Company data, Bloomberg Finance L.P., J.P. Morgan estimates

Nvidia GPU server: VR Ultra likely to see broad-based cooling spec upgrades

Table 2: Nvidia VR Ultra (VRU) system to see meaningful cooling spec upgrades

Note: For QD units, the (N+2) means N units of QD inside tray + 2 units of QD (plug QD) outside of tray, so totalling (N+2) units of QD per tray. Source: J.P. Morgan estimates, Company data.

ompute tray level: cooling content to increase another 30-50% in VRU driven by double-sided cold plate adding diamond copper materials and ZQD adoption

• old plate: We believe the cooling content (including cold plate, QD and inner anifold) of VRU co pute tray can reach US$3,000-3,500 level, which suggests 30-50% content increase vs current VR design, and should be better than arket expectation of li ited changes or even de-spec. The stronger content growth in VRU co pute tray is ainly driven by 1) the double-sided cold plate design for Bianca modules given VPD (vertical power delivery) adoption (placed at the backside of chip): aking it 6 cold plates (2 existing + 4 at the backside for rack GPU) on the upper half of tray for the two Bianca odules (vs 2 CP units in VR upper-half tray) + the existing 3 cold plates at the botto half for the CX-9 chips and power distribution board (PDB), 2) diamond copper material adoption: specifically added for the cold plate areas attaching on HBMs, 3) additional QDs required (additional 4 sets/8 units of Mini QD, in our view) for the additional cold plates (the ones for the VPD placed at the backside of chip), and 4) QDs for the Bianca module cold plates upgrading to ZQD fro MQD previously with ~2x ASP increase.
• QD: In VRU syste , we believe Nvidia will upgrade so e of its QDs fro the previous MQD (Multi Quick Disconnect)/UQD (Universal Quick Disconnect) to the higher-spec ZQD (Zero-drip/flow rate Quick Disconnect) to help i prove the cooling efficiency with better liquid flow rate and li ited leakage. Hence, we

believe the VRU co pute tray QD unit will be largely the sa e as VR with 20 (inside tray)+2 (outside of tray) units, but the 2 units of plug QD outside of tray will be upgraded to ZQD14 in VRU fro UQD08 in previous gens, and 8 of the 20 inside-tray QDs for the two Bianca odules (upper half of tray) will also be upgraded to ZQD06 fro MQD06, while the re aining 12 units (for the botto half of tray) will stay at the MQD06 spec. Aside fro the 22 units of MQD+ZQD, we believe there ight be an additional 8 units of MiniQD for VRU co pute tray, placed at the botto half of tray. In ter s of the QD content per co pute tray, we odel it to increase to ~US$750 for VRU vs US$400-500 for VR.

Switch tray level: 2x QD unit adoption per VRU switch tray (1.5U) with the double-layered (2 boards per tray) architecture

• old plate: Given the new design of '2 layers of boards per 1.5U switch tray (2 switch boards each at 0.75U)', we odel the cooling content (including cold plate, QD and inner anifold) of the 1.5U switch tray to increase to at least US$3k, based on 2x cold plate odule adopted given 2 switch boards.
• QD: The QD units inside the '1.5U' VRU switch tray will likely double vs that in VR, i plying an increase to (8+2) units fro (4+2) units, given 2 layers/boards per tray. Although the 8 units of QD inside the VRU switch tray should all re ain with the MQD06 spec, the 2x unit increase still drives overall QD content growth. In addition, the 2 QDs outside of VRU switch tray will be upgraded to ZQD14 fro UQD08 in VR, which also adds incre ental content fro ASP increase for these 2 plug QDs. In total, we odel the QD content per switch tray to increase to ~US$260 for VRU vs ~US$150 for VR, based on our esti ates.

Table 3: Meaningful QD spec upgrades in VRU with ~100% content increase per NVL72 rack

Source: J.P. Morgan estimates, Company data.

hip level: 3-piece/removable lid design in Rubin Ultra to drive 4-5x content growth

• Heat spreader/Lid: Co pared to the current 1-piece lid design in Rubin GPU, we believe Rubin Ultra GPU will adopt the 're ovable lid + stiffener + ultiple screws' design (or so called 3-piece lid design), which should drive the heat

spreader/lid content increase by 4-5x. The Rubin Ultra design will require so e screws to attach the lid with the stiffener during chip packaging, and after the packaging and testing process, the screws and lid will then be re oved by the ODM vendors to place cold plates on top of the chip with a stiffener surrounding and a TIM1.5 in between.

Table 4: Nvidia GPU heat spreader lid spec/design and content growth

Source: J.P. Morgan estimates, Company data.

CSP ASIC: more aggressive cooling designs and upgrades vs GPU, despite lower TDP

AWS Teton 3 Max (Trainium 3 liquid-cooled system)

Table 5: AWS Trainium 3 (Trn 3) liquid-cooled rack creates higher cooling content vs Nvidia GPU systems

Note: Cooling content includes cold plate, QD, inner manifold and rac manifold (without CDU or sidecar). Source: J.P. Morgan estimates, Company data.

• old plate: We believe the cooling content (including cold plate, QD and inner anifold) of AWS Trn3 co pute tray can reach US$3,000-3,500 level (~30% higher than Nvidia's VR co pute tray), thanks ainly to the double-sided cold plate design (8 Ps for 4 ASI s) plus 1 incremental tray-sized cold plate for the baseboard with decent a ount of QD adoption. Also, the Trn3 switch tray ay also create US$1,500-2,000 content value, with slightly ore trays per rack (10 switch trays in Teton 3 Max vs 9 units in VR NVL72 rack). Therefore, the total cooling content (including cold plate, QD and inner manifold) per Teton 3 Max rack could reach ~US$86k (vs <US$70k content for VR NVL72 rack).

• QD: AWS Teton 3 Max adopts UQD and MQD in their Trn3 co pute and switch tray designs, and the total QD unit adoption per Teton 3 Max rack is 512 units vs 450 units in Nvidia VR NVL72 rack (both excluding rack anifold socket QDs). Each QD ASP is at ~US$20, so the total QD content per AWS Teton 3 Max (Trn3 liquid-cooled rack) is ~US$10.2k per NRL72 rack.

• Heat spreader: AWS ASICs adopt stiffeners for their heat spreader design, including Trainiu 3. The ASP of these general stiffeners ranges between US$1020, in our view, but each new generation will still see double-digit ASP growth driven by size increase and aterial changes.

oogle TPU v8 rack (liquid-cooled system)

Table 6: Google TPU v8 compute tray content creates similar cooling content vs Nvidia GPU compute tray despite lower chip TDP

Note: We only compare compute tray content for TPU rac , as it does not have switch trays (inter-chip interconnect for TPU) . Source: J.P. Morgan estimates, Company data.

• old plate: We believe the cooling content (including cold plate, QD and inner anifold) of TPU v8 co pute tray can reach US$2-2.5k level, si ilar to Nvidia's VR co pute tray despite lower chip TDP, due to the ore co plicated cold plate design, including 1) ore cold plate unit adoption: a total of 12 cold plates with 4 big ones for TPUs and an additional 8 s all ones for surrounding chips, and 2) goldplating design (given etal TIM adoption) to add incre ental content. Therefore, the total cooling content (including cold plate, QD and inner manifold) per TPU v8 rack could reach ~US$38k.
• QD: Google TPUv8 co pute tray adopts NSP1 QD, unlike Nvidia or other CSP ASIC syste s that generally use MQD or UQD, and it only places 2 units of NSP1 QD per TPU co pute tray (vs 22 units in VR and Trn3 co pute tray). In ter s of ASP, we believe NSP1 QD is 2-3x higher than both MQD and UQD. However, based on our calculation, the total QD content of Google TPU is relatively small at ~US$1.8k with only 32 units per rack (vs 400 or 500 unit level for VR and Trn3 rack), despite higher ASP.
• Heat spreader: Google TPUs also adopt stiffeners for their heat spreader design, including both v8i and v8t. However, for TPU v9, we believe the re ovable lid + stiffener could be a potential solution, given uch bigger size and higher TDP (potentially reaching 2,000W), which ay drive ~10x content increase if the new design is actually adopted. This is yet to be confir ed, but we believe the spec ugprade trend is certain.

ompanies Discussed in This Report (all prices in this report as of arket close on 29 Septe ber 2026, unless otherwise indicated) Asia Vital Co ponents(3017.TW/NT$3,400.00/OW), Fositek(6805.TW/NT$2,515.00/OW)

Analyst ertification: The Research Analyst(s) denoted by an 'AC' on the cover of this report certifies (or, where ultiple Research Analysts are pri arily responsible for this report, the Research Analyst denoted by an 'AC' on the cover or within the docu ent individually certifies, with respect to each security or issuer that the Research Analyst covers in this research) that: (1) all of the views expressed in this report accurately reflect the Research Analyst's personal views about any and all of the subject securities or issuers; and (2) no part of any of the Research Analyst's co pensation was, is, or will be directly or indirectly related to the specific reco endations or views expressed by the Research Analyst(s) in this report. For all Korea-based Research Analysts listed on the front cover, if applicable, they also certify, as per KOFIA require ents, that the Research Analyst's analysis was ade in good faith and that the views reflect the Research Analyst's own opinion, without undue influence or intervention.

All authors na ed within this report are Research Analysts who produce independent research unless otherwise specified. In Europe, Sector Specialists (Sales and Trading) ay be shown on this report as contacts but are not authors of the report or part of the Research Depart ent.