Computex 2026 highlights opportunities in AI server power, BMC, and CPO switches

Overview

This report summarizes JPMorgan's second set of observations on the Asian technology hardware supply chain during Computex 2026, focusing on areas including the NVIDIA Vera Rubin/VR200 platform, the AMD Helio system, AI server power architecture, BMC, SiC/GaN power semiconductors, capacitors/BBU, CPO switches, and AI PCs. The overall conclusion is that demand for AI datacenter hardware should continue expanding, but there is divergence across segments in mass-production timing, adoption rates, and degree of benefit.

Core views

The core views include: first, NVIDIA's Kyber GPU rack may be delayed due to backplane signal integrity issues, and Rubin Ultra NVL144 still needs to choose between dual NVL72 interconnect and NVLink CPO switching; second, demand for VR200 and HVDC power racks has upside potential, with product displays and management commentary from Delta and others pointing to higher adoption rates; third, ASPEED BMC benefits from rising AI server system complexity and longer order visibility, and 2027 revenue forecasts may be revised upward; fourth, early Vera CPU demand is positive for BMC, liquid cooling, and DRAM, but has mixed implications for the CPU socket supply chain; fifth, due to lower space efficiency in the ORW form factor and a longer design cycle, AMD Helio cabinet-level mass production for CSPs may be delayed until 1H27; sixth, CPO switches, SiC, capacitors, and BBU, among other power and interconnect-related segments, are likely to become incremental beneficiaries of expanding AI datacenter capital expenditure.

Analysis framework

The report mainly cross-validates observations from Computex exhibits, JPMorgan booth visits, supply-chain research, discussions with company management, and the team's existing forecasts, with a focus on assessing investment implications from mass-production timing, per-rack power/component configuration, adoption rates, supply-chain content per system, and customer demand visibility.

Methodology notes

• Supply chain researchCross-validation of Computex exhibits and company interviews

  Validate new platform designs and adoption rates through trade show displays, management commentary, and supply-chain checks.

  The report combines products shown by NVIDIA, AMD, and ODM/power/switch vendors at Computex with JPMorgan's research on customer demand, production ramps, and component configurations to determine which segments face upside to revenue TAM or timing risks.

• Platform teardownAI server rack-level BOM and content analysis

  Break down supply-chain beneficiary segments across layers such as racks, trays, power shelves, BMC, power semiconductors, BBU, and switches.

  The report emphasizes differences in the system form factors of VR200, Vera CPU, Helio, and CPO switches, and uses these to infer changes in BMC attach rate, power semiconductor content, capacitor count, BBU capacity, and ODM assembly efficiency.

• Production ramp assessmentMP timing and adoption-rate scenario analysis

  Use mass-production timing, adoption rates, and design maturity to judge the pace of revenue recognition.

  For example, Kyber remains about 1.5 years away from mass production and faces signal integrity challenges, AMD Helio rack-level production has been pushed to 1H27, while scale-out CPO switches are expected to enter mass production earlier than scale-up CPO.

Asset mapping & comparison

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

• ASPEED Technology Inc. (5274.TWO)

  Beneficiary target in AI server BMC

  Strengths

  Order visibility extends into 2027, AST2700 is ramping, and NVIDIA modular systems increase BMC attach rate.

  Weaknesses

  Near-term revenue momentum may be constrained by supply limitations.

  Comparison

  Compared with about 3 months of visibility in traditional servers, agentic AI-related server buildouts bring unusually long demand visibility.

  Risks

  If AI server buildouts slow, supply improvements disappoint, or customers adopt custom designs, upside to BMC revenue may be lower than expected.

• Delta Electronics, Inc. (2308.TW)

  Beneficiary target in HVDC power supply, BBU, and gray-space power infrastructure

  Strengths

  Comprehensive HVDC product portfolio, with management expecting around 20% HVDC adoption in the VR200 generation, while also expanding into SOFC, SST, PCS, energy storage, and microgrid controllers.

  Weaknesses

  The company explicitly does not enter liquid-cooling gray-space products such as chillers and water facilities.

  Comparison

  Compared with the previous assumption of about 15% power-rack adoption, a 20% HVDC adoption rate implies upside to power revenue TAM.

  Risks

  Changes in VR Ultra configuration, rising costs of power components such as SiC, lower-than-expected HVDC penetration, or delays in gray-space product ramp-up.

• Hon Hai Precision (2317.TW)、Quanta Computer Inc. (2382.TW)、Wistron Corporation (3231.TW)、Wiwynn Corp (6669.TW)

  AI server ODM and rack assembly supply chain

  Strengths

  The modular and cableless VR200 design shortens assembly time, helping improve yield, accelerate ramp-up, and reduce working capital pressure; Wiwynn and Wistron displayed AMD/NVIDIA-related systems.

  Weaknesses

  Platform designs are still evolving, and the AMD Helio rack form factor has lower space efficiency with delayed production timing.

  Comparison

  The VR200 NVL72 compute tray production cycle of about 5 minutes is significantly shorter than the roughly 2 hours for a GB300 tray.

  Risks

  Customer customization, platform delays, yield ramp, rack design changes, and CSP demand volatility.

• Infineon Technologies、STMicroelectronics、Wolfspeed、ON Semiconductor

  SiC/GaN power semiconductor supply chain

  Strengths

  In 800V HVDC racks, SiC could become a key device in AC/DC PSUs, e-Fuse, and 800V-to-54V DC/DC conversion, with SiC content per PSU potentially several times that of traditional silicon-based power semiconductors.

  Weaknesses

  GaN is still in the qualification stage, and high-voltage applications are constrained by voltage limits.

  Comparison

  Datacenters currently contribute only a small share to SiC TAM, but this may rise to a mid- to high-single-digit share over the next 2-3 years.

  Risks

  Qualification progress, cost, supply assurance, customer design choices, and EV-driven volatility in the SiC supply-demand cycle.

• Nichicon、Nippon Chemi-Con、TDK、Musashi

  Traditional capacitor and supercapacitor supply chain

  Strengths

  AI workloads raise requirements for power stability and transient response, and the number of aluminum electrolytic capacitors in NVIDIA's 3RU 110kW power shelf could increase by more than 10x.

  Weaknesses

  Some hyperscalers may choose standalone supercapacitor shelves, so the degree of benefit to traditional capacitors depends on customer architecture.

  Comparison

  Traditional capacitors benefit from higher PSU content, while supercapacitors benefit from solutions pairing a 72kW power shelf with a standalone 1RU supercapacitor shelf.

  Risks

  Platform design changes, customer choices on space utilization, material costs, and supply stability.

• Dynapack (3211 TT)、AES (6781 TT)

  Potential beneficiary targets in BBU and battery packs

  Strengths

  If customers extend backup power duration, BBU capacity and per-rack value would rise mechanically.

  Weaknesses

  Current power-rack BBUs usually support only about 1-2 minutes of ride-through, and whether demand upgrades remains a customer decision.

  Comparison

  Compared with about 5 minutes of backup duration for UPS, existing BBU duration is shorter, leaving room for expansion toward standalone BBU racks.

  Risks

  Customers not increasing backup power requirements, battery cost, safety certification, and system integration complexity.

• NVIDIA/MediaTek RTX Spark AI PC ecosystem

  AI PC theme and edge AI demand observations

  Strengths

  NVIDIA and MediaTek jointly launched the RTX Spark AI PC chip, expected to be broadly available in 3Q26.

  Weaknesses

  The report has not yet seen a significant change in the fundamental ecosystem, powerful agentic AI models can run in the cloud, and there are still concerns about Windows on Arm compatibility.

  Comparison

  The report compares this with the QCOM AI PC rally at end-2024 and believes stock gains driven by AI PCs after Computex may fade.

  Risks

  Lack of a killer app that must run locally, pricing that is too high, compatibility issues, and insufficient consumer replacement demand.

• Wistron、Accton/Edgecore、Ayar Labs、GUC、Hon Hai

  CPO/NPO switch and AI network interconnect supply chain

  Strengths

  The 102.4T Spectrum-6 photonic switch and 3.2T scale-up CPO projects indicate that CPO switches may ramp rapidly over the next 1-2 years, with high margins on switch projects.

  Weaknesses

  Scale-up CPO switch mass production lags scale-out CPO, and some projects are still in the design or co-development stage.

  Comparison

  Scale-out CPO is expected to reach mass production later this year or early next year, while scale-up CPO is expected to reach mass production from end-2027 to early 2028.

  Risks

  CPO maturity, customer adoption pace, yield, optoelectronic integration complexity, and platform timing changes.

Key data

• Kyber GPU rack mass-production target2H27The report states that the Kyber GPU rack is targeted for mass production in 2H27, but backplane signal integrity challenges create uncertainty.
• Potential reduction in VR200 chip TDPfrom about 2.3kW to about 1.8kWEven with the TDP adjustment, the report expects the rack-level AC/DC PSU configuration to remain at four 110kW power shelves per rack.
• Delta's expected HVDC adoption rate for the VR200 generationabout 20%This is above JPMorgan's prior regular + high-voltage power rack adoption assumption of about 15%, implying upside to power revenue TAM.
• AST2700 penetration forecast20-30% by end-2027ASPEED BMC revenue is expected to benefit from server TAM expansion, the AST2700 ramp, and improved supply.
• Vera CPU unit forecastabout 600,000 units in 2026; about 3 million units in 2027Including head nodes, with further upside if demand is stronger.
• CPU count per Vera CPU rack32 per air-cooled rack; 256 per liquid-cooled rackWistron displayed both air-cooled and liquid-cooled Vera CPU rack configurations.
• AMD Helio rack powerabout 240kWThe system consists of 18 compute trays, 72 AMD MI455 GPUs, and 6 switch trays, with total power supply capacity of about 432kW.
• VR200 NVL72 compute tray production cycleabout 5 minutesThis is significantly shorter than the roughly 2 hours for a GB300 tray, benefiting from modular, cableless, and midplane design.
• Datacenter SiC device TAM contributionabout US$400-500mn or less by 2028 or laterThe report believes datacenters' contribution to SiC device TAM could rise from low single digits to mid- to high-single digits within 2-3 years.
• Delta SOFC delivery targetabout 100MW over the next two yearsAt US$4-5/W, this corresponds to about US$400mn in revenue, or around 1% of total revenue.
• Hon Hai CPO switch shipment outlookabout 10,000 units this year, more than doubling next yearThe report expects CPO switches to ramp rapidly over the next 1-2 years.

Impact & implications

From an investment perspective, AI server platforms are expanding from single GPU racks into more complex rack clusters and combinations of CPU/storage/switching/power systems, increasing content opportunities in BMC, power semiconductors, capacitors, BBU, liquid cooling, and ODM assembly. The main beneficiaries are more likely to be Delta, ASPEED, Wiwynn, Quanta, Hon Hai, some CPO switch suppliers, and the SiC/capacitor supply chain; by contrast, the AI PC theme lacks support from a clear local killer app, while AMD Helio and NVIDIA Kyber require continued monitoring of design maturity and mass-production timing.

Risks

• Backplane signal integrity issues at Kyber may delay NVIDIA's next-generation GPU rack design decisions and mass production.
• Changes in VR Ultra configuration may affect adoption of HVDC power racks, while rising costs of power components such as SiC could limit 100% penetration.
• ASPEED's near-term BMC revenue may be constrained by supply, and customer custom designs could also reduce the standard BMC attach rate.
• AMD Helio's ORW form factor has lower space efficiency, and silicon and UBB design cycles may be longer than expected, potentially delaying rack-level mass production for CSP projects.
• AI PCs lack a killer app that must run locally, while Windows on Arm compatibility and end-device pricing may suppress demand.
• New technologies such as CPO, GaN, SiC, BBU, and supercapacitors still face risks around qualification, cost, yield, and supply assurance.

What to watch

• Progress of M10 CCL qualification, solutions to Kyber backplane signal integrity, and NVIDIA's final choice for the Rubin Ultra NVL144 architecture.
• HVDC adoption rates for VR200 and VR Ultra, rollout of 660kW/560kW PSU/BBU solutions, and orders for vendors such as Delta, Lite-On, Flextronics, and Vertiv.
• ASPEED AST2700 ramp-up, pace of supply improvement, and the quarter-over-quarter BMC revenue trend from 4Q26 to 1H27.
• Early project progress for Vera CPU among U.S. hyperscalers, potential shipments in 2Q26, and whether 2027 unit demand exceeds about 3 million units.
• The timeline for AMD Helio from small-batch production to rack-level mass production, especially small-volume progress in late 3Q26 to 4Q26 and mass-production progress in 1H27.
• CPO switch shipment volumes, mass-production timing for Spectrum-6 scale-out projects, and progress of Ayar Labs/GUC scale-up CPO projects from end-2027 to early 2028.
• Changes in datacenter content for SiC, GaN, capacitors, e-Fuse, and BBU, and whether requirements for power safety and backup duration are upgraded.