Taiwan cooling components for AI server liquid cooling: Cooling upgrades in Nvidia VR Ultra and ASIC platforms could drive larger-than-expected component-content growth

Overview

This report examines how rising AI-chip power requirements and changing server architectures are increasing liquid-cooling content across Nvidia GPU systems and cloud-provider ASIC platforms. JPMorgan argues that the scale of specification upgrades is underappreciated and identifies AVC and Fositek as preferred beneficiaries.

Core views

JPMorgan's central thesis is that cooling is becoming the key constraint on continued thermal-design-power growth in AI chips and the migration of AI servers to liquid cooling. It expects specification upgrades in both Nvidia GPUs and cloud-service-provider ASIC systems to create more cooling-component content than investors expect, particularly in Nvidia's next-generation Vera Rubin Ultra (VRU). The report explicitly challenges concerns that VRU could be de-specified, arguing instead that upgrades in cold plates, quick disconnects and heat-spreader/lid designs should lift value for leading component suppliers. At the VRU compute-tray level, JPMorgan estimates cooling content, including cold plates, QDs and the inner manifold, at US$3,000-3,500, 30-50% above the current Vera Rubin design. The increase is tied to a double-sided cold-plate design required by vertical power delivery, which relocates power discretes, SOCAMM and capacitors to the back of the chip board and requires incremental backside cooling. The proposed design would use six cold plates for two Bianca modules, compared with two in the prior configuration, while diamond-copper material around HBM areas and additional MiniQDs add value. VRU compute-tray QD content is modeled at about US$750, compared with US$400-500 for Vera Rubin. At the rack level, the report expects a major QD upgrade. Some MQD/UQD connectors are expected to shift to higher-spec ZQD connectors, which have roughly twice the ASP and are intended to improve liquid flow while limiting leakage. ZQD06 is expected for Bianca-module cold plates, while ZQD14 is expected for out-of-tray plug QDs and rack-manifold sockets. Together with added connector counts, JPMorgan estimates total QD content per VRU NVL72 rack at about US$20,500, versus about US$11,000 for Vera Rubin and US$9,400 for GB300—an approximately 100% increase versus VR200/GB300 racks. The 1.5U double-layer switch-tray architecture is also important: two switch boards per tray could lift switch-tray cooling content to more than US$3,000 and QD content to about US$260, versus about US$150 for Vera Rubin. Heat-spreader content is another upgrade path. JPMorgan expects Rubin Ultra to adopt a removable-lid, stiffener and multiple-screw design rather than the current one-piece lid. This three-piece design could raise heat-spreader/lid content by 4-5x versus Rubin's one-piece lid. Based on industry checks, the firm believes limited Rubin GPU trial production could begin in 4Q26, with Rubin Ultra mass production potentially in 1Q27; this timing is described as somewhat ahead of market expectations. The report also cites AMD MI450's two-piece-lid adoption, with about a 10x ASP increase versus MI350, as evidence of a broader structural opportunity for heat spreaders. The report extends the cooling-content argument to CSP ASICs. AWS Trainium 3 is expected to use a double-sided cold-plate design with eight cold plates for four ASICs plus a baseboard cold plate. JPMorgan estimates US$3,000-3,500 of cooling content per Trainium 3 compute tray, roughly 30% above Nvidia Vera Rubin, and about US$86,000 of cooling content per Teton 3 Max rack, compared with less than US$70,000 for VR NVL72. Google TPUv8 is estimated to have US$2,000-2,500 of compute-tray cooling content despite lower chip TDP than Nvidia Vera Rubin, because it uses four large gold-plated cold plates for TPUs and eight smaller plates for surrounding ICs. TPUv8 rack cooling content is estimated at about US$38,000. For TPUv9, the report sees a potential removable-lid-plus-stiffener solution if higher TDP and larger die sizes materialize, though this remains unconfirmed. JPMorgan expects the sector to outperform and favors AVC and Fositek. AVC is preferred for its leading position and market share in GPU and ASIC cold-plate projects, which the report believes can support better-than-peer margin and earnings growth. Fositek is favored for potential near-term gross-margin upside, higher-spec QD content growth, continued share gains and TAM expansion. The report also sees CPO/NPO switches and higher-spec optical transceivers or modules as a next growth driver over the coming one to two years as liquid cooling becomes mainstream in future 1.6T/3.2T generations.

Analysis framework

JPMorgan compares cooling architectures, component counts, ASPs and per-tray and per-rack content across Nvidia GPU systems, AWS Trainium 3 and Google TPUv8. It then links design changes—such as double-sided cold plates, higher-spec QDs and removable lids—to supplier exposure, earnings potential and valuation observations.

Methodology notes

• Industry AnalysisVolume-price decomposition

  Component-content analysis using unit counts, specifications and ASPs

  The report estimates cooling value by identifying changes in cold-plate, QD and lid quantities and specifications, then translating those changes into per-tray and per-rack dollar content.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI-chip thermal requirements transmitted to cooling-component suppliers

  The report links higher GPU and ASIC power requirements and server-design changes to demand, content and margin implications for cold-plate, QD and heat-spreader suppliers.

• Competition & strategyEconomic Moat and Competitive Advantage

  Supplier positioning and project share

  AVC and Fositek are assessed through their stated leadership, market share, project exposure, share-gain potential and ability to capture higher-specification content.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Asia Vital Components (3017.TW)

  Preferred cooling supplier with exposure to cold-plate and manifold upgrades in Nvidia and CSP ASIC projects.

  Strengths

  Leading position and market share across projects; expected better-than-peer margins and earnings growth.

  Comparison

  Trades at about 20x 12-month forward P/E versus a historical 18-22x range; the report notes underperformance versus peers and Taiex.

• Fositek (6805.TW)

  Preferred QD supplier exposed to ZQD adoption, incremental MiniQD demand and rack-manifold QD expansion.

  Strengths

  Potential near-term gross-margin upside, continued share gains and longer-term TAM expansion.

  Comparison

  Trades at about 20-25x 12-month forward P/E versus a historical 15-30x range.

• Jentech Precision Industrial Co. (3653.TW)

  Covered heat-spreader/lid supplier exposed to Rubin and Rubin Ultra lid-design changes.

  Strengths

  Potential benefit from removable-lid adoption and higher heat-spreader content.

  Comparison

  Trades at about 50x 12-month forward P/E versus a historical 30-50x range.

  Risks

  Partial Rubin GPU adoption in trial volumes and only 4-5x content growth versus a one-piece lid at mass production.

Key data

• VRU cooling content per compute trayUS$3,000-3,500Estimated 30-50% above the current Vera Rubin design.
• VRU total cooling content per NVL72 rack~US$98,500Compared with ~US$66,250 for Vera Rubin NVL72; excludes CDU.
• VRU rack QD content~US$20,500Compared with ~US$11,000 for Vera Rubin and ~US$9,400 for GB300.
• VRU switch-tray cooling content>US$3,000Driven by a 1.5U tray with two switch boards.
• Rubin Ultra heat-spreader/lid content growth4-5xVersus Rubin's one-piece lid design.
• AWS Teton 3 Max cooling content per rack~US$86,000Compared with ~US$66,250 for Nvidia Vera Rubin NVL72; excludes CDU.
• Google TPUv8 cooling content per compute trayUS$2,000-2,500Similar to Nvidia Vera Rubin despite lower chip TDP.

Impact & implications

The report says escalating cooling requirements should expand the addressable value of cold plates, QDs and heat spreaders across Nvidia and ASIC platforms. It views AVC as positioned to benefit from cold-plate upgrades and Fositek from QD upgrades, margin potential, share gains and TAM expansion.

Risks

• For Jentech, the report identifies risks that Rubin may use the removable lid only in partial trial volumes and that content growth may be limited to 4-5x versus a one-piece lid at mass production.
• Google TPUv9 removable-lid adoption remains unconfirmed despite the report's expectation of continued specification-upgrade trends.