Taiwan cooling components: Cooling upgrades in Nvidia VR Ultra and AI ASICs could extend Taiwan suppliers’ content growth

Overview

JPMorgan’s sector report contends that rising AI-chip thermal requirements are driving broader liquid-cooling upgrades across Nvidia GPU and cloud ASIC systems. It sees cold plates, quick disconnects and heat spreaders as key beneficiaries, with AVC and Fositek its preferred Taiwan cooling names.

Core views

JPMorgan argues that cooling, rather than being de-specified, is becoming the bottleneck that enables continued growth in AI-chip thermal design power and AI-server migration. It expects Nvidia’s Vera Rubin Ultra (VRU) to adopt upgrades across cold plates, quick disconnects (QDs), and lids/heat spreaders. The firm expects these changes to deliver larger component-content gains than investors anticipate over the next one to two years, while later liquid cooling in CPO/NPO switches and higher-spec optical modules could add another growth driver from the 1.6T/3.2T generation. At the VRU compute-tray level, JPMorgan estimates cooling content—including cold plates, QDs and inner manifolds—at US$3,000-3,500, 30-50% above the current Vera Rubin design. The proposed double-sided cold-plate architecture is linked to vertical power delivery, which shifts power components, SOCAMM and capacitors to the rear of the chip board and requires backside cooling. VRU could use six cold plates for two Bianca modules, versus two in the prior upper-half tray arrangement, while diamond-copper material around HBM cooling areas adds value. The report also expects eight additional Mini QDs and upgrades of Bianca-module QDs from MQD to ZQD, which carries roughly twice the ASP. JPMorgan models VRU compute-tray QD content at about US$750, versus US$400-500 for Vera Rubin. The tray would retain 22 principal QD positions but upgrade two out-of-tray plugs to ZQD14 and eight Bianca-module QDs to ZQD06, while adding eight Mini QDs. At the 1.5U VRU switch-tray level, two stacked switch boards should require two cold-plate modules and lift total cooling content to at least US$3,000. QD content per switch tray is modeled at about US$260 versus about US$150, reflecting eight internal MQDs rather than four and an upgrade of the two external plug QDs to ZQD14. Across an NVL72 rack, the firm estimates QD content at about US$20,500, compared with about US$11,000 for Vera Rubin and US$9,400 for GB300. The report also expects Rubin Ultra’s removable-lid-plus-stiffener-and-screws design to proceed without delay. It estimates heat-spreader/lid content at four to five times that of the current one-piece Rubin lid, with limited Rubin trial production potentially beginning in 4Q26 and mass production in Rubin Ultra potentially in 1Q27. JPMorgan sees the schedule as somewhat ahead of market expectations and potentially supportive of near-term revenue and earnings for the three-piece-lid supplier. It also identifies AMD MI450’s two-piece lid, with roughly a tenfold ASP increase versus MI350, plus future CPO/NPO adoption and ASIC upgrades as structural support for heat-spreader demand. Cloud ASIC platforms reinforce the argument that cooling content can rise despite lower chip TDP. AWS Trainium 3’s Teton 3 Max system is estimated to generate US$3,000-3,500 of cooling content per compute tray, roughly 30% above Nvidia Vera Rubin, because of eight double-sided cold plates for four ASICs plus a tray-sized baseboard cold plate. With 18 compute trays, 10 switch trays and a rack manifold, total content is modeled at about US$86,000 per rack excluding the CDU, versus about US$66,250 for Vera Rubin NVL72. Google TPU v8 is estimated at US$2,000-2,500 per compute tray and about US$38,000 per rack: its lower TDP is offset by four large gold-plated cold plates for TPUs and eight smaller plates for surrounding ICs. TPU v9, potentially reaching 2,000W, could adopt a removable lid and stiffener; JPMorgan says this remains unconfirmed but views the broader upgrade trend as certain. For stock implications, JPMorgan prefers AVC because its cold-plate and manifold position spans GPU and ASIC projects, which it believes supports better-than-peer margins and earnings growth. It favors Fositek for expected near-term gross-margin upside, increasing ZQD content, share gains and expansion into higher-spec rack-manifold QDs. The firm says both trade at valuations it considers undemanding relative to history and could deliver 3Q26 margin upside and longer-term earnings upside from more substantial specification upgrades.

Analysis framework

JPMorgan compares cooling architectures across Nvidia GPU and cloud-ASIC systems, breaking content into cold plates, QDs, inner manifolds and rack manifolds. It estimates component counts and ASP changes by tray and rack, then links the resulting content changes to supplier positioning, margins and earnings potential. It also compares forward P/E ranges and its EPS estimates with Bloomberg consensus for the discussed Taiwan suppliers.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Cooling-component content analysis across AI chips, compute trays, switch trays and racks

  The report traces how chip packaging, power-delivery and board-design changes require more or higher-spec cooling components, then identifies suppliers positioned to benefit.

• Industry AnalysisVolume-price decomposition

  Component-unit and ASP-based content modeling

  JPMorgan estimates cooling value by counting cold plates and QDs and applying higher ASPs for designs such as ZQD, diamond copper and removable lids.

• Valuation methodsP/E and PEG Valuation

  Forward P/E compared with historical trading ranges

  The report describes AVC and Fositek as trading at valuations it considers undemanding based on 12-month forward P/E versus historical ranges.

Asset mapping & comparison

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

• Asia Vital Components (3017.TW)

  Preferred cold-plate and manifold supplier positioned to benefit from GPU and ASIC cooling upgrades.

  Strengths

  Leading position and market share across projects; exposure to Nvidia VR, major CSP ASICs and AMD MI-series; potential margin and earnings growth.

  Comparison

  JPMorgan cites ~20x 12-month forward P/E versus a historical 18-22x range and recent underperformance of the Taiex by about 2%.

• Fositek (6805.TW)

  Preferred QD supplier positioned to benefit from ZQD upgrades and increasing QD content in VRU racks.

  Strengths

  Potential near-term gross-margin upside, share gains, new customer/project wins and expansion into higher-spec rack-manifold QDs.

  Comparison

  JPMorgan cites 20-25x 12-month forward P/E versus a historical 15-30x range.

• Jentech Precision Industrial Co. (3653.TW)

  Comparable heat-spreader/lid supplier exposed to removable-lid adoption.

  Strengths

  Exposure to Rubin and Rubin Ultra removable-lid adoption.

  Comparison

  JPMorgan cites ~50x 12-month forward P/E versus a historical 30-50x range.

  Risks

  Partial Rubin GPU adoption for trial volumes and only 4-5x, rather than higher, content growth when mass production begins.

Key data

• VRU compute-tray cooling contentUS$3,000-3,500Estimated 30-50% above Vera Rubin, including cold plates, QDs and inner manifold.
• VRU rack QD content~US$20,500About 100% above Vera Rubin’s ~US$11,000 per NVL72 rack.
• Rubin Ultra lid content growth4-5xVersus the current one-piece Rubin lid design.
• AWS Teton 3 Max rack cooling content~US$86,000Excluding CDU; versus ~US$66,250 for Nvidia Vera Rubin NVL72.
• Google TPU v8 rack cooling content~US$38,000Driven by complex and gold-plated cold-plate designs.
• AVC 2026/27/28E EPSNT$108/187/236JPMorgan estimates, 3%/16%/16% above Bloomberg consensus.
• Fositek 2026/27/28E EPSNT$64/121/187JPMorgan estimates, 1%/12%/40% above Bloomberg consensus.

Impact & implications

The report says rising cooling content per AI-server platform should favor leading Taiwan suppliers of cold plates, QDs and heat spreaders. AVC is positioned for GPU and ASIC cold-plate upgrades, while Fositek could benefit from ZQD adoption, higher content per rack, gross-margin improvement and share gains.

Risks

• For the heat-spreader supplier, Rubin GPU may see only partial removable-lid adoption during trial volumes, and mass-production content growth may be limited to 4-5x versus the one-piece design.