COMPUTEX 2026 Preview: AI Infrastructure Enters the 'AI Factory' Era

Overview

This report provides Citi’s forward-looking analysis ahead of COMPUTEX 2026 and GTC Taipei. The core conclusion is that investment logic in AI infrastructure is undergoing a structural shift—from a narrow focus on GPUs, HBM, and CoWoS capacity toward a broader 'AI factory' system design. With the introduction of NVIDIA’s Rubin architecture and Vera CPU, performance bottlenecks are no longer confined to accelerators alone but now extend to networking, optical communications, custom chips, and system-level bandwidth. This transition implies that the value proposition for Taiwan’s tech supply chain will expand comprehensively from single-chip manufacturing to full-rack integration, power management, thermal solutions, and advanced packaging.

Core views

Architectural Paradigm Shift: From Servers to Composable AI Factories The report notes that NVIDIA is leading a transformation in AI infrastructure—from traditional server deployment models to a 'composable AI factory' framework. Under this new paradigm, racks are no longer stacks of independent servers but are redefined as functional modules: GPU racks, Vera CPU racks, LPU racks, CMX context memory racks, Spectrum networking racks, and power/liquid-cooling infrastructure. The Rubin Ultra NVL576 system interconnects eight MGX NVL racks to form a single NVLink domain with 576 GPUs, signifying that the 'rack' has evolved into a standardized module within larger Pod-scale compute domains. ODMs like Hon Hai, Quanta, and Wistron are expected to showcase such modular solutions at the event. Redefining the CPU Role: Vera as the 'Traffic Controller' of AI Factories The Vera CPU is highlighted as a key focus of COMPUTEX 2026. Citi emphasizes that Vera is not designed to compete with general-purpose server CPUs but is purpose-built for the AI factory orchestration layer—handling request scheduling, agent workflow control, memory indexing, and system telemetry. As inference workloads evolve toward Agentic AI, CPUs must deliver high single-thread performance and power efficiency, and the SOCAMM memory architecture is better suited than traditional DIMMs for high-density, low-latency orchestration needs. NVIDIA’s management reaffirmed a ~$200 billion CPU market opportunity, with an estimated 2 million Vera CPUs to be stocked in 2026. Hon Hai and Quanta will demonstrate Vera-based rack systems. Interconnects and Optical Communications: Data Movement Becomes the New Bottleneck As AI systems scale from single racks to Pod-level deployments, data movement efficiency—not raw compute—becomes the primary bottleneck. The report argues that Ethernet, via Spectrum-X, is evolving into a dedicated AI Fabric. For Rubin Ultra, physical scale exceeds the effective range of copper cables, significantly elevating the importance of co-packaged optics (CPO). The adoption curve is projected as follows: 2025–2026 dominated by pluggable optical modules; 2026–2027 sees CPO on switch sides and direct optical links introduced; post-Rubin Ultra enters a phase of broad optical fabric expansion. Strategic importance rises for optical engines, FAUs, silicon photonic PICs, and fiber management components. Power and Thermal Management: Co-Design Challenges in Electrical-Thermal Systems The Rubin architecture pushes rack power consumption to unprecedented levels. Compared to the Hopper era’s 40–60kW, the Rubin NVL72 rack TDP is projected at 180–220kW+, with Rubin Ultra potentially reaching several hundred kilowatts. At 54V, a 216kW rack draws 4,000A—a major challenge for busbars, connectors, and thermal management—forcing the industry toward 800V DC power distribution, 110kW power shelves, and full liquid cooling. Future AI racks are essentially precision 'electro-thermal machines,' requiring deep co-design across power conversion, backup energy, fluid distribution, and telemetry safety systems. Companies like Delta, AVC, and Auras stand to gain incremental opportunities in power and thermal solutions.

Analysis framework

The report employs a 'generational technology evolution + system bottleneck migration' analytical framework. First, by comparing parameter shifts across three NVIDIA architectures—Hopper, Blackwell, and Rubin (e.g., rack power jumping from 40kW to over 200kW)—it identifies that chip process scaling alone can no longer solve system-level challenges. Second, based on NVIDIA’s product roadmap (e.g., NVL576, Vera CPU, Spectrum-X), it deduces that AI infrastructure value is diffusing from the 'compute core' to peripheral systems like interconnects, power, and cooling. Finally, it maps this technological trend onto specific segments of Taiwan’s supply chain to assess which niches (e.g., ABF substrates, optical assembly, power shelves) can capture new system-level value, leading to an overall bullish conclusion on the supply chain.

Methodology notes

• Industry/Industrial Analysis FrameworkUpstream-Midstream-Downstream Value Chain Transmission

  Restructuring of the AI Infrastructure Value Chain

  Rather than viewing chip performance in isolation, the report analyzes how value flows downstream and upstream once GPU compute density hits physical limits—toward midstream (ODM full-rack design) and upstream components (optical modules, power, cooling). This approach helps investors understand why power and thermal solution providers can also benefit from the AI boom beyond GPUs.

• Industry/Industrial Analysis FrameworkPenetration S-curve

  CPO Technology Adoption Pathway Prediction

  The report stages the rollout of co-packaged optics (CPO): 2025–2026 dominated by pluggables → 2026–2027 switch-side CPO and direct optical links → post-Rubin Ultra full optical fabric. This uses the S-curve method for new technology adoption—inferring inflection points based on physical bottlenecks (copper limits) and product cycles rather than linear extrapolation.

• Company Fundamentals and Financial Framework

  System-Level Physical Constraint Analysis

  By analyzing the physical relationship among rack power (kW), current (A), and voltage (V), the report deduces the necessity of 800V DC and liquid cooling. This engineering-based fundamental analysis shows that certain supply chain opportunities arise not from commercial choice but from immutable physical laws, thus offering higher certainty.

Asset mapping & comparison

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

• Hon Hai

  Core ODM partner; expected to showcase Vera Rack and SPX networking rack systems

  Strengths

  Full-stack system integration capability across GPU, CPU, and networking racks; deeply aligned with NVIDIA’s next-gen architecture

  Comparison

  Alongside Quanta, a primary demonstrator of NVIDIA’s AI factory architecture with the broadest product coverage

• Quanta

  Core ODM partner; expected to showcase Vera Rack and SPX networking rack systems

  Strengths

  Leader in AI server rack design and manufacturing; actively developing both Rubin and Vera platforms

  Comparison

  Jointly with Hon Hai, forms the first tier of system-level solution providers

• Delta

  Core power solutions provider; expected to showcase Rubin-optimized power systems

  Strengths

  Technology leader in high-power-density shelves, 800V DC distribution, and BBU; directly benefits from doubled rack power demands

  Comparison

  Holds superior market share and technical depth in AI server power versus peers

• TSMC

  Foundation of AI chip manufacturing and advanced packaging

  Strengths

  CoWoS capacity remains essential for AI compute scaling and is irreplaceable in Rubin/Vera chip production

  Comparison

  As an upstream cornerstone, offers higher certainty than mid/downstream assemblers, though marginal upside increasingly comes from system-level support

• Wistron/Wiwynn

  AI server ODMs; expected to showcase VR NVL72 and other rack solutions

  Strengths

  Actively involved in NVIDIA’s next-gen rack architecture development; strong competitiveness in the AI server niche

  Comparison

  Second-tier AI rack integrators, closely following Hon Hai and Quanta

Key data

• Vera CPU Stock Expectation for 2026Approximately 2 million unitsReflects NVIDIA’s commitment to the AI orchestration layer and ODM shipment potential
• Rubin NVL72 Rack TDP180–220kW+3–4x increase from Hopper era (40–60kW), directly driving demand for liquid cooling and high-voltage power
• Rubin Ultra NVL576 GPU Count576 GPUsComposed of 8 interconnected MGX NVL racks, marking the rack as a modular unit in Pod-scale compute domains
• NVIDIA’s Projected CPU Market Opportunity$200 billionCEO Jensen Huang reiterated that Vera is part of AI infrastructure, not merely a replacement for traditional server CPUs
• Rubin Ultra Rack Power Potential300–600kWMakes 800V DC power distribution essential to overcome the 54V/4000A current transmission bottleneck

Impact & implications

The report argues that the most significant signal from COMPUTEX 2026 is that AI infrastructure has entered a new era of system-level design. For Taiwan’s tech supply chain, this means a shift in evaluation criteria: success will no longer depend solely on exposure to individual components (e.g., GPUs or CoWoS) but on the ability to deliver complete AI systems. Beyond core players like TSMC, structural growth opportunities will emerge in ABF substrates, test sockets, thermal modules, power shelves, busbars, optical assembly, BMC controllers, and rack integration. Suppliers capable of co-optimizing with NVIDIA’s next-gen architectures will secure stronger positions in the 'AI factory' era.

What to watch

• Physical configurations of Vera CPU racks and Rubin NVL72 systems demonstrated by major ODMs at COMPUTEX 2026
• NVIDIA’s latest disclosures on Spectrum-X and CPO technology roadmaps
• Validation progress of 800V DC and full liquid cooling solutions from power and thermal vendors
• Actual 2026 ramp-up trajectory and end-customer feedback for Vera CPU