AI Infrastructure Enters a Full-Stack Systems Competition, with Investment Opportunities Expanding from Accelerators to the Entire Supply Chain

Overview

UBS attended the 2026 OCP APAC Summit in Taipei. The central conclusion was that AI infrastructure is no longer simply about procuring GPUs, but about full-stack systems optimization across compute, memory, networking, power delivery, cooling, security, advanced packaging, storage, and rack/data center architecture. As AI cluster scale and per-rack power continue to rise, capital expenditure will flow more broadly to infrastructure components that support high accelerator utilization.

Core views

The report believes that power delivery, advanced packaging, and network optics are the three key constraints in the next phase of AI buildout. In power delivery, 400V/800VDC and side-mounted power architectures could support AI racks of several hundred kilowatts or even megawatt-scale power; in packaging, CoWoS, SoIC, FoCoS, 3.5D, and hybrid bonding make packaging a boundary for system performance; in networking, CPO and optical interconnects can reduce power consumption, improve reliability, and overcome copper interconnect distance limitations. Agentic AI will also bring more workloads for CPUs, memory, orchestration, and system management, expanding beneficiaries beyond GPUs and accelerators.

Analysis framework

Based on keynote speeches, technical sessions, and industry participant views at OCP APAC Summit 2026, the report reviews technology roadmaps, bottlenecks, and potentially benefiting supply chains across power delivery, advanced packaging, network optics, the compute stack, and rack-level systems integration.

Methodology notes

• Value Chain AnalysisFull-Stack AI Infrastructure

  Co-optimization from chips to data centers

  Analyzes compute, memory, networking, power delivery, cooling, packaging, storage, and rack architecture as an interdependent system rather than as isolated components.

• Technology Roadmap AnalysisScale-Out Expansion

  Scale-up, scale-out, and in-package scaling

  Expands AI system scale and efficiency through intra-rack accelerator interconnects, inter-rack/data-center interconnects, and the integration of more compute, memory, and I/O into a single package.

• System ArchitectureOpen Standards

  Multi-vendor interoperability

  OCP and related open specifications aim to reduce lock-in to proprietary architectures and support modular, composable AI cluster deployments.

Asset mapping & comparison

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

• TSMC

  Beneficiary of advanced packaging and 3D integration

  Strengths

  CoWoS, 3DFabric, and advanced silicon integration capabilities support system-level scaling of AI chips.

  Weaknesses

  Advanced packaging capacity expansion, yields, and supply-chain coordination require high execution capability.

  Comparison

  Compared with models relying only on logic process technology, system-level packaging capabilities are becoming a more important competitive dimension.

  Risks

  Customer demand volatility and risks to advanced packaging capacity and yields.

• ASE

  Beneficiary of advanced packaging and CPO

  Strengths

  VIPack, FoCoS, 2.5D/3D, CPO, and vertical power delivery solutions cover key aspects of AI packaging.

  Weaknesses

  The technology roadmap is diverse, creating uncertainty around mass-production validation and customer adoption timing.

  Comparison

  Compared with traditional assembly and test, AI packaging places greater emphasis on heterogeneous integration, optical-electronic coordination, and power-path optimization.

  Risks

  Packaging yields, thermal management, optical alignment, and large-scale mass-production risks.

• ASpeed

  Beneficiary of server management and security infrastructure

  Strengths

  BMC and server management capabilities align with AI cluster manageability requirements.

  Weaknesses

  Affected by server platform design cycles and customer adoption.

  Comparison

  Management, telemetry, and security become more important in high-density AI racks.

  Risks

  Server demand cycles and intensifying competition.

• MediaTek

  Beneficiary of custom chips and systems integration

  Strengths

  Expanding demand for custom chips and system-level integration creates opportunities.

  Weaknesses

  The contribution from specific AI infrastructure projects still depends on customer design wins.

  Comparison

  Open ecosystems are driving more specialized SoC and heterogeneous computing designs.

  Risks

  Design-win, product execution, and customer concentration risks.

• Hon Hai、Quanta、Wistron、Wiwynn

  Beneficiaries of server ODM and rack integration

  Strengths

  Can benefit from liquid-cooling-native racks, modular platforms, and high-power AI rack construction.

  Weaknesses

  System complexity, supply-chain coordination, and delivery requirements are increasing.

  Comparison

  Value content is extending from traditional server manufacturing to full-rack and data-center-level integration.

  Risks

  Customer capital expenditure, component supply, and project execution risks.

• Delta、King Slide、BizLink

  Beneficiaries of power delivery, mechanical, and interconnect infrastructure

  Strengths

  800VDC, liquid cooling, cable interconnects, mechanical components, and high-density rack construction provide incremental content value.

  Weaknesses

  Demand realization depends on platform iteration and the pace of standards implementation.

  Comparison

  Compared with traditional data centers, AI racks have materially higher requirements for power density, cooling, and interconnect performance.

  Risks

  Safety certification, technology roadmap changes, and industry competition risks.

Key data

• OCP Membership ScaleMore than 400 members in 2025OCP was initiated by Facebook (now Meta) in 2011 to promote open data center hardware designs.
• 800VDC Penetration AssessmentApproximately 10%Feedback in the report indicates penetration of around 10% at the initial stage of the Rubin platform, with subsequent platforms expected to increase further.
• Traditional vs. Agentic Inference CPU/GPU RatioShifting from 1:4 to 1:1The report cites AMD's view that agentic AI increases CPU demand for orchestration, applications, databases, and storage.
• CPO Reliability and Power ConsumptionReliability improves by approximately 10x, while optical interconnect power consumption declines by more than 70%Broadcom disclosed CPO metrics related to Tomahawk 6 Davisson.
• ASE Optical-Electronic Integration PotentialUp to 32x higher bandwidth and approximately 6x lower energy lossBrings optical components closer to the compute engine to support future 200Gbps-per-lane interconnects.
• Vertical Power Delivery EfficiencyPower loss reduced from approximately 10% to approximately 2%ASE noted that vertical power delivery can also save more than 30% of X/Y plane area.
• AI Chip Packaging Scale9x package size, 600x silicon area, 400x die count, and 1,000x transistor densityApplied Materials' description of the evolution in advanced packaging complexity.
• Arm Data Center Core ShipmentsMore than 1.5 billion coresArm stated this is cumulative shipment volume since Neoverse was deployed by cloud service providers in 2022.

Impact & implications

The beneficiaries of AI capital expenditure are expanding across the full infrastructure value chain. On the semiconductor side, advanced packaging, optical packaging, BMC/security chips, custom chips, and system-level integration capabilities are becoming more important; on the hardware side, server ODMs, power systems, liquid cooling, cable interconnects, mechanical components, and rack integration will benefit from higher rack density and AI cluster construction. The report explicitly favors semiconductor beneficiaries including TSMC, ASE, ASpeed, and MediaTek, and is positive on hardware and infrastructure-related companies including Hon Hai, Quanta, Wistron, Wiwynn, Delta, King Slide, and BizLink.

Risks

• AI data center capital expenditure may fall below expectations or projects may be delayed.
• Standardization, certification, and mass-production progress for technologies such as 800VDC, CPO, advanced packaging, and liquid cooling may fall short of expectations.
• Safety, cooling, reliability, and maintainability challenges for high-power racks may raise deployment costs.
• Advanced packaging, HBM, optical components, and power electronics supply chains may face capacity or yield constraints.
• Competition between open standards and proprietary ecosystems may change value allocation across segments.
• Many of the technical metrics cited in the report are derived from conference presentations and company views; actual commercialization outcomes still require validation.

What to watch

• Actual adoption rates and safety certification progress for 400V/800VDC in next-generation AI racks.
• Customer adoption and mass-production timelines for open optical interconnect specifications such as CPO, optical I/O, and OCI.
• Capacity, yields, and customer demand for CoWoS, SoIC, FoCoS, hybrid bonding, and 3.5D packaging.
• Incremental demand from agentic AI workloads for CPU, memory, storage, networking, and system management resources.
• AI capital expenditure by hyperscale cloud providers, custom-chip deployments, and the pace of liquid-cooled rack construction.
• Orders, product content value, and gross margin changes for Taiwan server ODM, packaging, power delivery, and interconnect suppliers.