報告_GF_散熱冷板與QD價值_20260909

PDF 原檔：報告_GF_散熱冷板與QD價值_20260909_original.pdf

原始內容

Thermal: VRU NVL72 to Boost Cold Plate & QD Value

Jeff Pu, CFA

SFC CE No. BNO719

jeffpu@gfgroup.com.hk

Henry Huang

SFC CE No. BYA749

henryhuang@gfgroup.com.hk

What's New?

Structurally, the rising AI server TDP and power consumption are driving greater adoption of advanced cooling solutions and higher cooling content per system. For 2027, we expect the trend to continue into Rubin Ultra platform, despite HBM de-spec, driven by the likely doublesided of cold plate per GPU and enhanced material of copper diamond. We also expect QD value to increase.

Comments

Rising AI server power consumption drives cooling content: GPU TDP has surged from Ampere/Hopper's 400W-700W to Blackwell's 1,000-1200W, with Rubin expected to reach up to 2,300W; AI server racks have scaled from 10kW (H200) to 36kW (GH200) in earlier GPU deployments to 130-140kW (GB300) in current configurations, and expected to head towards 600kW to 1MW in future generations. This shift makes AI server platforms increasingly depend on advanced cooling solutions, driving higher cooling content for a system.

Cold plate content per rack to be driven by double-sided and copper diamond: Due to higher TDP (we expect 2,600W to be mainstream for Rubin Ultra), VR200 NVL72 includes both large and small form-factor cold plates within the compute tray and upgrades to a microchannel cold plate (MCCP) architecture; hence, we expect VR200 NVL72 to utilize 5 cold plates in each compute tray (2 large plates and 3 small plates), with switch tray remaining the same, leading to the content per rack of ~$34,200, slightly increase from that of GB300 NVL72. For VR300 (Rubin Ultra), we expect the content value to further increase by 58%, vs. that of VR200 NVL72, driven by: 1) double-sized cold plate for GPU, 2) copper diamond powder/material for GPU's cold plate; 3) doubled units of switch ASICs (for one NVL72).

QD content per rack to rise into Rubin platform: Given Vera Rubin shifts to all liquid-cooled architecture, VR200 NVL72 expands the number of liquid-cooled components, driving more QD use per system. We estimate VR200 NVL72 to utilize 22 QDs in each compute tray and 6 QDs in each switch tray, resulting in content per rack of ~$10,760, 77% increase from GB300 NVL72. For VR300, we expect the content value to further increase by 103%, vs. that of VR200 NVL72, thanks to upgrade to ZQD and doubled usage per switch tray (due to 2 boards per tray).

Key beneficiaries

• Cold plate: AVC (3017 TT)

• QD: Fositek (6805 TT) with 50%+ share, LOTES (3533 TT), FIT (06088 HK Buy)

• ODMs/OEMs: Hon Hai (2317 TT Buy), Supermicro (SMCI US Buy)

Risks: 1) AI demand slowdown, 2) infrastructure deployment, 3) supply constraints, 4) competition.

Rising TDP and power consumption for AI server

AI server power content is rising structurally, driven by increasing system/rack level power demands along with expanding per-chip power envelopes. While conventional host server CPU TDPs typically peak at 300-500W, advanced AI accelerators have scaled aggressively across generations. Take Nvidia's GPU superchips for instance, Single-GPU TDP has surged from legacy Ampere/Hopper's 400W-700W to Blackwell's 1,000-1200W, with upcoming Rubin expected to reach up to 2,300W. However, due to HBM de-spec, we believe the Rubin Ultra series' TDP could be in the range of 1,800 to 2,600. Yet, we believe next-gen platforms of Feynman will increase TDP even further, sustaining the demand for higher-density power delivery.

Figure 1: Nvidia GPU TDP (W)

Sources: Company data, GF Securities (Hong Kong) Brokerage

At system level, rising accelerator TDP directly translates into a step-up in rack-level power consumption. AI server racks have scaled from 10kW (H200) to 36kW (GH200) in earlier GPU deployments to 130140kW (GB300) in current configurations, and expected to head towards 600kW to 1MW in future generations. This shift makes AI server platforms increasingly depend on advanced cooling solutions, driving higher cooling content for a system.

Figure 2: Nvidia rack solutions roadmap

Sources: Tech Insignts, GF Securities (Hong Kong) Brokerage

Cooling content per rack to rise into Rubin platform

As said, cooling architecture has transitioned from a routine supporting function to a system-level performance bottleneck, marking a long-term structural upgrade in AI server cooling design and continuous increase of content value, where Blackwell supports both liquid and air -cooled configurations, while Vera Rubin shifts to an all liquid -cooled architecture, including GPU/CPU to components such as SOCAMM and NIC/ConnectX.

For cold plate, VR200 NVL72 includes both large and small form-factor cold plates within the compute tray and upgrades to a microchannel cold plate (MCCP) architecture; hence, we expect VR200 NVL72 to utilize 5 cold plates in each compute tray (2 large plates and 3 small plates), vs. 6 small plates for GB300 NVL72, while blended ASP rise to ~$292 from ~$240, with switch tray remaining the same, leading to the content per rack of ~$34,200, slightly increase from that of GB300 NVL 72. For VR300 (Rubin Ultra), we expect the content value to be further increase by 58% with more use of small plates, vs. that of VR200 NVL72, driven by: 1) double-sized cold plate for GPU, 2) copper diamond powder/material for GPU's cold plate; 3) doubled units of switch ASICs (for one NVL72).

For QD (Quick Disconnect), Given Vera Rubin shifts to all liquid-cooled architecture, VR200 NVL 72 materially expands the number of liquid-cooled components, driving more QD use per system. We estimate VR200 NVL72 to utilize 22 QDs in each compute tray and 6 QDs in each switch tray, resulting in content per rack of ~$10,760, 77% increase from GB300 NVL72. For VR300, we expect the content value to increase by 103% thanks to upgrade to ZQD and doubled usage per switch tray (due to 2 boards per tray).

Figure 3: Cold plate content forecast

Sources: GF Securities (Hong Kong) Brokerage

Figure 4: QD content forecast

Sources: GF Securities (Hong Kong) Brokerage

Figure 5: Cold plate and QD content forecast for VR300 NVL 72

Sources: GF Securities (Hong Kong) Brokerage

Risks

1) AI demand slowdown, 2) infrastructure deployment, 3) supply constraints, 4) competition.

Rating definitions

Benchmark: Hang Seng Index (Hong Kong)

Company ratings Buy

Stock expected to outperform benchmark by more than 10 %

Hold

Expected stock relative performance ranges between -10% and 10%

Underperform

Stock expected to underperform benchmark by more than 10%

Sector ratings

Positive

Sector expected to outperform benchmark by more than 10%

Neutral

Expected sector relative performance ranges between -10% and 10%

Cautious

Sector expected to underperform benchmark by more than 10%

Hong Kong Company

GF Securities (Hong Kong) Brokerage Limited

Address

27/F, GF Tower, 81 Lockhart Road, Wan Chai, Hong Kong

Telephone

(852) 37191111

Email

evanlee @gfgroup.com.hk