AI Infrastructure Shifts from GPU Procurement to Full-Stack System Optimization

Overview

UBS attended the OCP APAC Summit held in Taipei on August 10-11, 2026. The core conclusion from the conference is that AI infrastructure is no longer simply a GPU procurement cycle, but a full-stack systems competition encompassing compute, memory, networking, power delivery, cooling, security, advanced packaging, storage, and rack and data center architecture. Open standards and multi-vendor collaboration are driving large-scale AI clusters toward higher power density, greater bandwidth, and improved energy efficiency.

Core views

First, power architecture will become critical to scaling high-power AI racks, with 400V/800V DC solutions likely to be gradually adopted. Second, performance bottlenecks are shifting from transistor scaling to packaging yield, memory, thermal design, power delivery, and data movement, raising the importance of coordinated optimization across advanced packaging and system technologies. Third, co-packaged optics can alleviate limitations in copper interconnect distance, power consumption, and reliability. Fourth, agentic AI broadens demand for CPUs, memory, networking, and systems management, expanding AI investment from accelerators to the complete data center value chain.

Analysis framework

Based on keynote speeches, thematic forums, and presentations by exhibitors at the OCP APAC Summit, the report synthesizes the shared views of hyperscale cloud providers, chip design companies, packaging houses, equipment manufacturers, server ODMs, and power-infrastructure vendors on technology roadmaps and supply-chain opportunities.

Methodology notes

• Value Chain AnalysisFull-Stack AI Infrastructure

  System-level coordination from chips to data center operations

  Compute, storage, networking, packaging, power delivery, cooling, and rack integration are viewed as an interdependent whole rather than isolated hardware procurement categories.

• Technology Roadmap AnalysisScale-Up, Scale-Out, and Scale-In

  In-rack interconnects, cross-rack cluster interconnects, and in-package integration

  AI system performance, bandwidth, and energy efficiency are improved by expanding tightly coupled accelerator domains, extending cluster interconnects, and integrating more compute and I/O within packages.

• Supply-Demand AnalysisCapital Expenditure Diffusion

  AI investment extends from accelerators to the infrastructure value chain

  Operating expensive accelerators at high utilization requires concurrent investment in power delivery, liquid cooling, interconnects, packaging, and rack-level management.

Asset mapping & comparison

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

• TSMC

  Beneficiary of advanced packaging and 3D integration

  Strengths

  CoWoS, SoIC, and 3DFabric capabilities support demand for AI chips and heterogeneous integration.

  Weaknesses

  Advanced packaging capacity expansion, yields, and customer product timing may affect the pace of realization.

  Comparison

  Compared with pure-play logic foundry services, system-level packaging capabilities are more important to the AI value chain.

  Risks

  Demand volatility, capacity constraints, technology iteration, and geopolitical risks.

• ASE

  Beneficiary of advanced packaging and co-packaged optics

  Strengths

  Covers technology paths including fan-out packaging, 2.5D/3D, FoCoS, and co-packaged optics.

  Weaknesses

  Validation cycles for high-end packaging are lengthy, while scaled costs and yields still require optimization.

  Comparison

  Benefits as packaging evolves from a back-end manufacturing process into a system-integration boundary.

  Risks

  Changes in technology choices, delayed customer qualification, and intensifying competition.

• ASpeed

  Beneficiary of server management and security infrastructure

  Strengths

  Data center manageability, baseboard management controllers, and security demand rise with greater AI rack complexity.

  Weaknesses

  Demand is closely tied to server shipments and platform designs.

  Comparison

  Compared with compute chips, it benefits more directly from increasing rack management and operations complexity.

  Risks

  Server demand volatility, customer concentration, and competition risks.

• MediaTek

  Beneficiary of custom chips and system-level integration

  Strengths

  The open ecosystem and custom silicon trend create expansion opportunities.

  Weaknesses

  The scale of its data center business and timing of customer adoption remain uncertain.

  Comparison

  Compared with general-purpose accelerator suppliers, it is more focused on customization and ecosystem collaboration opportunities.

  Risks

  Uncertainty in design wins, R&D investment, and supply-chain execution risks.

• Hon Hai、Quanta、Wistron、Wiwynn

  Beneficiaries of AI servers and rack integration

  Strengths

  Possess server manufacturing, liquid-cooled rack, and system-integration capabilities.

  Weaknesses

  Margins are affected by customer bargaining power, component supply, and product complexity.

  Comparison

  Benefits as AI infrastructure shifts from standalone servers toward modular rack and cluster deliveries.

  Risks

  Adjustments in customer capital expenditure, supply-chain bottlenecks, and delivery and certification risks.

• Delta、King Slide、BizLink

  Beneficiaries of power, mechanical, and interconnect infrastructure

  Strengths

  Benefit from growing demand for high-power racks, 800V DC, liquid cooling, cabling, and mechanical components.

  Weaknesses

  Penetration and mass-production timing of new power standards remain uncertain.

  Comparison

  Compared with conventional data center components, higher AI rack power density drives greater value per system.

  Risks

  Safety certification, divergent technology standards, price competition, and deferred demand.

Key data

• Number of OCP MembersMore than 400 in 2025OCP is an industry consortium promoting open data center hardware designs.
• Assessment of High-Voltage DC Power PenetrationApproximately 10% for the Rubin platformThe report states that high-voltage DC adoption will advance as computing power consumption rises; Google and Meta may progress more rapidly with ASIC systems above 400V.
• CPU/GPU Ratio for Traditional versus Agentic AI InferenceChanges from 1:4 toward 1:1The report cites AMD's view that agentic workloads increase requirements for CPU orchestration, databases, storage, and tool execution.
• Advanced Packaging Interconnect Density ImprovementFoCoS approximately 53x; FoCoS Bridge approximately 212xCompared with conventional flip-chip packaging, ASE's heterogeneous integration solutions can increase interconnect density.
• Potential Benefits of Co-Packaged OpticsBandwidth can increase by up to 32x, while energy loss can decline to approximately one-sixthThese are technical metrics presented by ASE at the conference and remain dependent on engineering implementation and scaled adoption.
• Co-Packaged Optics Reliability and Power ConsumptionReliability improves by approximately 10x; optical interconnect power consumption declines by more than 70%These are Tomahawk 6 Davisson-related metrics disclosed by Broadcom.
• Share of Data Center Thermal Management CostsApproximately 40%The report cites Applied Materials' view that thermal management is a key lever for improving energy efficiency.

Impact & implications

For investors, the beneficiaries of AI capital expenditure are broadening. Advanced packaging, optics and co-packaged optics, networking chips, server CPUs, rack integration, 800V DC power delivery, backup power and energy storage, liquid cooling, cabling, and mechanical components may all see incremental demand. The report favors semiconductor beneficiaries TSMC, ASE, ASpeed, and MediaTek; on the hardware side, it favors Hon Hai, Quanta, Wistron, and Wiwynn. Component and infrastructure suppliers such as Delta, King Slide, and BizLink may also benefit.

Risks

• Safety certification, standards harmonization, and mass-production deployment of 400V/800V DC power delivery may be slower than expected.
• Co-packaged optics, 3D packaging, and vertical power delivery still face challenges in thermal management, optical alignment, yields, and serviceability.
• A slowdown in hyperscale cloud providers' AI capital expenditure would affect demand across the value chain.
• Open standards may accelerate competition and compress differentiation and margins for some suppliers.
• Advanced packaging, HBM, optical components, and power equipment may face capacity or delivery bottlenecks.
• Energy consumption, facility upgrades, and reliability requirements associated with high-power racks may increase deployment costs.

What to watch

• The pace of high-voltage DC power adoption by Rubin, Rubin Ultra, and hyperscale cloud providers' custom ASIC platforms.
• Reliability validation, serviceability, and mass-production progress of co-packaged optics in switches and AI clusters.
• Capacity, yields, and customer adoption of CoWoS, SoIC, fan-out packaging, and 3.5D integration.
• Whether agentic AI workloads continue to drive CPU and GPU configurations toward a more balanced mix.
• Membership expansion, specification maturity, and actual deployment of open standards such as OCP, ESUN, OCI, and UALink.
• AI rack orders, delivery capabilities, and margin changes at server ODMs, power vendors, and liquid-cooling suppliers.