Computex 2026: AI Server Architecture Upgrades Accelerate, Power/Cooling/Interconnect Become Key Bottlenecks and Opportunities

Overview

This report is Morgan Stanley’s in-depth summary of the 2026 Taipei Computex, focusing on the evolution of AI-infrastructure hardware. The core conclusion is that next-generation AI servers, represented by the NVIDIA Vera Rubin platform, are migrating from single-chip to rack-level co-design, driving simultaneous upgrades in power, cooling, interconnect, and structural subsystems; RTX Spark PC, despite its revolutionary vision, is constrained by price and ecosystem and is unlikely to become a near-term growth engine for the PC market.

Core views

The report argues that AI servers have entered a new ‘Rack-as-a-Platform’ phase, with the NVIDIA Vera Rubin POD epitomizing this trend—it is not a single product but a standardized combination of GPU racks (NVL72), CPU racks (Vera), storage racks (STX), and compute racks (LPX), requiring the supply chain to leap from component-level to rack-level delivery capability. On key subsystems, the report highlights three core advances: (1) 800 V high-voltage DC (HVDC) power is being rapidly deployed, with NVIDIA outlining a clear three-stage roadmap (3Q26 rack → 2Q27 center → 1Q28 module); vendors such as Delta are already showcasing solid-state transformers (SST) and fuel-cell solutions. (2) Liquid cooling has become a hard requirement; the Vera Rubin liquid-cooling solution is expected to reach mass production in 3Q26, while the more advanced micro-channel lid (MCL) faces industrialization challenges around yield, leakage, and long-term reliability. (3) High-speed interconnect technologies are diversifying, including optical wavelength-selective (OWS) switching, CPO architectures, and Micro-LED optical modules, migrating from optical modules toward CPO architectures. Regarding RTX Spark PC, the report is cautious. Although it integrates Grace CPU, Blackwell GPU, and unified memory with a long-term vision of a ‘24/7 home AI agent’, its $3,000-plus price, initial targeting of AI professionals, and substitutability with Mac Mini limit first-year shipments to an estimated 5–10 million units—insufficient to offset overall PC market headwinds.

Analysis framework

The report employs a typical ‘technology-driven supply-chain framework’: it first anchors on the core platform (NVIDIA Vera Rubin POD), decomposes it into four functional racks (GPU, CPU, storage, compute) and quantifies their ratio (8:1:1:5) to identify incremental opportunities; second, around the central conflict of soaring AI-server power density, it systematically maps the evolution roadmap and commercial milestones of the three bottleneck technologies—power (800 V HVDC), cooling (liquid cooling → MCL), and interconnect (OWS/CPO/Micro-LED); finally, combining booth-level details and executive interviews (e.g., Foxconn chairman Brand Cheng on mass-production timelines), it translates macro technology trends into micro-level firm orders and capacity schedules, completing the full chain from ‘what’ to ‘who’ and ‘when’.

Methodology notes

• Industry/sector analysis frameworkSupply-demand framework

  The sector is driven primarily by supply-side technology iteration and capacity-release cadence

  Rather than relying on end-demand forecasts, the report tracks technology standards defined by platform vendors such as NVIDIA (e.g., Vera Rubin POD specs) and mass-production schedules of supply-chain vendors (e.g., 3Q26 liquid-cooling mass production) to determine the inflection point and sustainability of the industry cycle

• Industry/sector analysis frameworkUpstream-midstream-downstream transmission

  Technology standards are defined upstream and transmitted level by level to midstream modules and downstream structural parts

  The report clearly presents the vertical transmission chain: ‘NVIDIA defines POD → Foxconn/Quanta build racks → FIT supplies connectors/power-power → Chenbro/King Slide supply structural parts’, showing that value is shifting from chips to entire racks and supporting components

• Cycle & cyclical frameworkInflection-point analysis

  Key equipment mass-production dates serve as inflection-point markers

  The report treats concrete milestones such as ‘800 V DC rack-ready 3Q26’ and ‘Vera Rubin liquid-cooling mass production 3Q26’ as the core evidence for judging when the industry enters an acceleration phase, rather than vague ‘positive trend’ statements

Asset mapping & comparison

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

• FIT Hon Teng Ltd (6088.HK)

  Core interconnect and power-solution supplier for the Vera Rubin platform, leading new quick-disconnect (QD) and liquid-cooling busbar designs

  Strengths

  Designated by NVIDIA as the lead designer for Rubin Ultra’s new QD, poised to disrupt the existing QD competitive landscape; liquid-cooling busbar ASP of US$3,000–4,000 with high share-gain certainty

  Weaknesses

  New Voronoi connector not yet finalized; pre-mass-production challenges remain around pogo-pin yield and cost optimization

  Comparison

  Together with Danfoss, one of two main players in the QD market, but holds first-mover advantage in next-gen designs

  Risks

  Voronoi connector industrialization falls short of expectations; final 800 V DC standard is unified by OCP consortium, potentially weakening its customized pricing power

• Accton Technology Corporation (2399.TW)

  AI-cluster network-switching solution provider, showcasing OWS optical switching and CPO switches

  Strengths

  Possesses optical engines (subsidiary InLC) and rack-level integration capability (Edgecore Open Fabric), enabling full-stack offerings

  Weaknesses

  Initial rack-level business may dilute gross margin; must offset via high-value-add modules such as self-developed NICs and liquid cooling

  Comparison

  Compared with pure component vendors, its rack-level solution better matches end-customer ‘one-stop procurement’ trends

  Risks

  Commercial deployment of optical switching slower than expected; direct competition with NVIDIA and Broadcom in the CPO space

• Chenbro (6213.TW)

  AI-server chassis and full-rack solution supplier, displaying a full range of MGX-architecture chassis and VR NVL72-compatible racks

  Strengths

  Actively expanding AI-related racks (IT/CDU/power), with a new Malaysia plant starting production 3Q26 to strengthen global responsiveness

  Weaknesses

  Rack business still in ramp-up; project wins and profitability yet to be proven

  Comparison

  Compared with traditional chassis vendors, its dual-platform layout on NVIDIA MGX/AMD Helios offers greater flexibility

  Risks

  AI-server demand volatility leads to lower-than-expected rack orders; new capacity ramps slower than planned

• Bizlink (3660.TW)

  Core supplier of AI-server power and data interconnect, covering power cables, DAC/AOC, and CPO shuffle boxes

  Strengths

  Post-acquisition of XFS, enhanced fiber assembly and shuffle-box capability; CPO shuffle box expected to ramp from 2027

  Weaknesses

  Emerging businesses (LEO satellites, robotics) currently contribute little, unlikely to form a second growth curve in the near term

  Comparison

  Maintains a solid share in AI-server power cables; shuffle box is the key entry point into the CPO ecosystem

  Risks

  Shuffle-box technology qualification cycle long; overall CPO adoption ramps slower than expected

Key data

• Vera Rubin POD rack ratio8 NVL72 GPU racks : 1 Vera CPU rack : 1 STX storage rack : 5 Groq LPX compute racksBased on official NVIDIA architecture, constituting the minimum deliverable unit of an AI server
• 800V DC deployment timeline3Q26 (rack), 2Q27 (1.6 MW center), 1Q28 (4.8 MW module)Technology roadmap unveiled at NVIDIA GTC Taipei, marking the upgrade of AI data-center power systems
• RTX Spark PC first-year shipment forecast5–10 million unitsDerived from an N1X chip run-rate of 200–250 k units per quarter, far below traditional annual PC volumes
• Vera Rubin liquid-cooling value per GPU tray~US$2,500Covers cold-plate modules, inner manifolds, and NVQD core parts, reflecting the high added value of cooling upgrades

Impact & implications

For the supply chain, this round of upgrades means the value center is migrating from traditional PC/server ODM toward high-barrier, highly customized AI-infrastructure segments. Power-solution vendors (e.g., FIT), network-switch vendors (e.g., Accton), precision-structure firms (e.g., Chenbro, King Slide), and high-speed-interconnect suppliers (e.g., Bizlink) will directly benefit from growing rack-level demand and rising technical thresholds. Investors should focus on each segment’s actual share within the Vera Rubin platform, the delivery of mass-production schedules, and the industrialization breakthrough cadence of new technologies (e.g., MCL, Micro-LED).

Risks

• Kyber rack platform fails to reach mass production for Rubin Ultra, delaying the next-gen AI-server upgrade cycle
• 800 V DC power standard fails to converge (e.g., OCP push stalls), creating equipment compatibility risks and redundant investment
• Micro-channel lid (MCL) industrialization progresses slowly; yield and reliability issues continue to constrain high-end AI-chip cooling solutions

What to watch

• Delivery of Vera CPU and Groq LPX rack mass-production schedules in 3Q26/4Q26
• Actual adoption scale of NVIDIA BlueField-4 STX storage racks among AI customers
• FIT’s new Voronoi connector and NVIDIA Rubin Ultra new QD design finalization and mass-production timeline