AI server supply chain cools and pulls back, but the 2027-2028 demand outlook is still being revised upward

Overview

This report is Bernstein's 2Q26 AI Server Pulse, covering the global semiconductor, hardware, Internet, and software value chains, with updates focused on AI capex, data-center projects, GPU/ASIC/CPU models, AI-server shipments, and the financial and share-price performance of supply-chain companies. The report notes that the AI supply chain has experienced a significant pullback since June, driven by potential delays to Nvidia Kyber racks, the impact of Kimi K3, momentum-stock deleveraging, and concerns about overbuilding in AI; however, component suppliers continue to accelerate capacity expansion to address demand in 2027-2028.

Core views

The core view is that a near-term cool-down will not alter medium-term expansion. Data-center investment and CSP capex continue to be revised upward, and the server market is expected to exceed US$1 trillion by 2028. GPUs remain the core of the XPU market, but ASIC penetration is accelerating, driven by inference demand, performance/cost optimization, and efforts to reduce reliance on external GPUs. PCB backplanes, T-glass, ABF substrates, HBM, CPO, power racks, and test equipment are key bottlenecks and beneficiaries. From a valuation perspective, most AI supply-chain companies in the sample have experienced P/E compression, but consensus 2026/2027 EPS estimates have mostly been raised, creating a combination of improving fundamentals and valuation concerns.

Analysis framework

The report uses a cross-industry supply-chain tracking framework that combines CSP capex, data-center project announcements, GPU/ASIC/CPU shipment models, CoWoS and HBM supply, server ODM/OEM revenue, and changes in share prices, valuations, and EPS expectations for 29 AI server supply-chain stocks.

Methodology notes

• Capex and Project TrackingAI Capex & Datacenter Schedule

  Uses data-center projects involving CSPs, neo-cloud providers, Oracle, and sovereign wealth funds/funds as leading indicators of AI hardware demand.

  The report compares consensus capex estimates from March 2026 with those following 2Q26 results and aggregates planned and under-construction data-center investment, capacity, commissioning schedules, and financing arrangements.

• Accelerator Market ModelGPU vs. ASIC XPU TAM

  Breaks the AI data-center accelerator market into GPUs and ASICs and considers the impact of HBM pricing on market size.

  The report estimates 2027 XPU TAM at approximately US$500B, with GPUs still accounting for the majority, while ASIC share increases due to growth in inference and custom-chip demand; higher HBM prices and whether ASIC/GPU vendors procure HBM on behalf of customers will affect revenue measurement.

• Supply Chain and Shipment ModelAI Server & Supply Chain Model

  Derives supply-chain beneficiaries across racks, servers, components, substrates, PCBs, CPO, and test equipment.

  The report raises its assumptions for 2025-2028E server and GPU AI-server shipment CAGRs and tracks the progress of platforms including Rubin, Rubin Ultra, AMD Helios, Google TPUv8, and Amazon Trainium3.

Asset mapping & comparison

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

• META PLATFORMS INC (META.US)

  A major CSP and AI data-center capex spender, as well as an important driver of ASIC and server demand.

  Strengths

  Investment in the Hyperion project has been raised to US$50B, with planned AI data-center projects in Canada, Indiana, Texas, and elsewhere; the company is connected to both ASIC/GPU and neo-cloud supply chains.

  Weaknesses

  The scale of capex is large, and the market remains concerned about AI ROI and overbuilding.

  Comparison

  Compared with CSPs overall, Meta's project expansion is notable, and the report assigns it an Outperform rating with an US$800 target price.

  Risks

  Data-center construction delays, changes in financing structure, AI-product monetization below expectations, and hardware-architecture changes.

• ADVANCED MICRO DEVICES INC (AMD.US)

  GPU and AI-server platform supplier; Helios racks are expected to begin shipping in 4Q26.

  Strengths

  Helios racks target customers including Meta, Microsoft, Oracle, OpenAI, and Anthropic; the acquisition of ZT Systems strengthens end-to-end AI-server and rack design capabilities.

  Weaknesses

  The company still needs to demonstrate that its ecosystem, software, and large-scale deployment capabilities can compete with NVIDIA.

  Comparison

  GPUs still account for the majority of the XPU market, but AMD has growth opportunities in alternative GPU supply and rack solutions; the report assigns it an Outperform rating with a US$600 target price.

  Risks

  NVIDIA's platform advantage, progress of the ROCm ecosystem, customer qualification timing, and HBM costs and supply.

• NVIDIA (NVDA)

  A core supplier in the GPU and rack ecosystem; the progress of Kyber/Rubin racks affects sentiment across the entire supply chain.

  Strengths

  Performance advantages in training workloads allow GPUs to continue dominating the XPU market, while Rubin racks are expected to ship in 4Q26.

  Weaknesses

  PCB backplane design and yields for Kyber racks are near-term bottlenecks.

  Comparison

  Although ASIC share is rising, GPUs remain dominant, and NVIDIA retains a relative lead in high-end training and rack ecosystems.

  Risks

  Kyber/Rubin delays, ASIC substitution, HBM pricing, and accelerated development of CSP in-house chips.

• Broadcom (AVGO)

  A key supplier of ASIC services and custom accelerators.

  Strengths

  The company has secured a long-term TPU partnership with Google through 2031 and expanded its partnership with Meta, benefiting from structural ASIC penetration.

  Weaknesses

  ASIC revenue scale may be affected by HBM procurement by the supplier and differences in pricing conventions.

  Comparison

  Compared with GPU suppliers, Broadcom benefits more directly from rising demand for CSP custom chips.

  Risks

  Customer-direct HBM procurement, ASIC project timing, advanced-packaging capacity, and competition from TPU and other ASICs.

• MediaTek

  A beneficiary of Google TPU projects; the report notes that it won two TPU projects.

  Strengths

  Expected to grow rapidly over the next two years, with 2026 ASIC revenue guidance raised to above US$2B.

  Weaknesses

  Growth is highly dependent on TPU project ramp timing and customer concentration.

  Comparison

  Rated Outperform within the Asian semiconductor coverage universe, benefiting as ASICs move from pilot programs to scaled deployment.

  Risks

  HBM procurement models, TSMC capacity allocation, project delays, and slower-than-expected expansion of external TPU customers.

• Unimicron / Delta

  Key Asian hardware supply-chain names, representing opportunities in IC substrates/PCBs and power and hardware components, respectively.

  Strengths

  The report names Unimicron as its top near-term pick and Delta as its top long-term pick; component suppliers receive advance payments, capex sharing, and LTA support.

  Weaknesses

  Share prices have risen significantly, and valuations are highly sensitive to sentiment toward the AI chain.

  Comparison

  Relative to CSP stocks, PCB, substrate, memory, and test-equipment names have remained the leading YTD performers.

  Risks

  Kyber design changes, yields, overly rapid capacity expansion, customer order adjustments, and technology-route changes.

Key data

• Planned and Under-Construction Data-Center InvestmentApproximately US$1.18-1.2 trillionHigher than the approximately US$960B estimate as of April 2026.
• Major CSP 2026 Capex ExpectationsRaised by approximately 15% from March 2026Consensus estimates continued to rise after 2Q26 results.
• Combined CSP, Neo-Cloud, and Oracle CapexApproximately US$1110B in 2027, with a 2025-2027 CAGR of approximately 61%Higher than the approximately US$822B forecast in March 2026.
• Capex of the Four Major US HyperscalersApproximately 58% CAGR from 2025-2027, reaching approximately US$940B in 2027Includes Amazon, Meta, Microsoft, and Google.
• Global Server Market SizeExpected to exceed US$1 trillion in 2028Driven jointly by high-end GPU AI servers and general-purpose servers.
• Server Shipment CAGR15% for total servers and 22% for high-end GPU AI servers from 2025-2028EThe report raises its model assumptions.
• High-End GPU Server ShipmentsExpected to grow 48% in 2026Supported by the smooth ramp of GB300 racks and the 2027 demand outlook.
• NVIDIA Server Rack Shipments61K in 2026 and 88K in 2027In 2026, Foxconn is expected to account for approximately 32K, while Quanta and Wistron/Wiwynn are each expected to account for approximately 10-15K.
• XPU Market SizeApproximately US$500B in 2027XPU is defined as GPU plus ASIC.
• ASIC ShareRising from 10% last year to 20% in 2027Adoption is driven by inference workloads and demand for custom chips.
• TPU ShipmentsMore than 90% year-over-year growth in 2026Approximately 42% of CoWoS-based ASIC shipments.
• AIDC ASIC Market SizeUS$140-160B in 2028, including HBMExcluding HBM, the market size is approximately half as large.
• TPU Market OpportunityAt least approximately US$90B by 2028Additional TSMC wafer allocation, advanced-packaging capacity expansion, and external customers are potential upside factors.
• CPO Optical EnginesApproximately 10M units in 2027 and more than 20M units in 2028Shipments over the next two years will focus on scale-out.
• AI Primary-Market FinancingUS$147B in 2Q26, up 213% year over yearApproximately 70% of all venture capital; AI financing grew 243% year over year in 1H26.
• AI Supply-Chain Share Prices and ValuationsSome stocks in the sample rose 100%-200% YTD, but 21/29 experienced P/E compressionValuation contraction reflects AI-bubble concerns, declining risk appetite, and concerns about CSP ROI.

Impact & implications

The investment implication is that the AI supply chain is transitioning from a sentiment-driven phase to one focused on earnings delivery and validation of supply bottlenecks. Upstream memory, HBM, PCB/HDI, IC substrates, test equipment, advanced packaging, and optical interconnects continue to benefit from demand pull-ins and advance stocking for 2027-2028; CSPs and neo-cloud providers, meanwhile, need to demonstrate that capex can translate into revenue, RPO, and cash flow. The report favors Unimicron as its top near-term pick and Delta as its top long-term pick among Asian hardware names, while maintaining positive ratings on Meta, AMD, NVDA, AVGO, Microsoft, and Oracle.

Risks

• The AI supply-chain pullback since June may continue, affected by momentum-stock deleveraging and declining risk appetite for technology stocks.
• Potential delays to Nvidia Kyber racks, with key bottlenecks in the design and yield of PCB backplanes connecting compute blades.
• Concerns about overbuilding in AI data centers and the investment returns of CSPs and frontier labs may weigh on valuations.
• Rising HBM prices and procurement/resale models may disrupt assessments of GPU and ASIC market sizes and margins.
• Changes in CSP capex guidance, financing arrangements, or accounting conventions may weaken the effectiveness of capex as a proxy for hardware demand.
• Changes in ASIC, GPU, and rack architecture specifications may create volatility in supply-chain orders, inventories, and capacity planning.
• If capacity expansion for T-glass, ABF substrates, advanced packaging, CPO, and power racks falls short of expectations, shipments may be constrained.
• Neo-cloud providers and GPU cloud-service providers depend on external financing and contracts with large customers; financing shortfalls or failure to realize contracts could cause demand volatility.

What to watch

• Financing arrangements for major AI supply-chain companies over the next 12 months, including OpenAI, Nvidia, Oracle, and CoreWeave.
• CSP capex guidance, financing plans, and RPO changes, particularly for Amazon, Meta, Microsoft, Google, and Oracle.
• Volume production of Rubin racks, progress on Rubin Ultra, and improvements in the PCB backplane design and yields of Nvidia Kyber racks.
• Customer qualification and shipment pace for AMD Helios racks after 4Q26.
• Scaled deployment of Google TPUv8, next-generation TPU projects, and Amazon Trainium3.
• The sustainability of CPU and general-purpose server demand driven by Agentic AI.
• Capacity allocation for T-glass, ABF substrates, HBM, and advanced packaging.
• The pace of power-rack, CPO, and optical-interconnect deployment from scale-out to larger-scale deployments.