AI server supply chain cools in the short term, while capex and compute demand still support long-term expansion

Overview

This report extends Bernstein's 1Q26 AI tracking framework and reviews 2Q AI capex, data center projects, GPU and ASIC competition, server shipments and financial performance of supply-chain companies. The report argues that the AI supply chain has entered a valuation and share-price correction phase since June, but industry demand has not weakened in tandem: cloud providers continue to raise investment, data center project scale is expanding, server and accelerator forecasts are being revised upward, and component makers are also expanding capacity ahead of 2027–2028 demand.

Core views

First, the short-term cooling in market sentiment mainly stems from the risk of delays to Nvidia Kyber racks, the impact of Kimi K3, deleveraging in technology momentum stocks and concerns over returns on AI investment, rather than a broad decline in end-market compute demand. Second, major cloud providers and new cloud providers are still expanding capex, and global planned and under-construction data center investment has approached US$1.2 trillion. Third, GPUs will still account for the majority of the AI accelerator market, but increasing inference workloads, cost-efficiency requirements and supply-chain autonomy will drive faster ASIC growth. Fourth, agentic AI will transform inference from one-off calls into continuous multi-step workflows, thereby simultaneously driving demand for CPUs, general-purpose servers, memory and storage. Fifth, investors should remain selective within the supply chain, focusing on segments with upward earnings revisions, constrained capacity or customer prepayment support, while avoiding companies with financing pressure and insufficient valuation realization.

Analysis framework

The report uses a combined top-down and bottom-up approach: it first tracks cloud provider capex, data center projects and AI financing; then builds shipment models for GPUs, ASICs, CPUs, servers and racks; subsequently compares share prices, valuation multiples and earnings estimate changes for 29 AI server supply-chain companies; and on that basis forms company ratings, target prices and key recommendations.

Methodology notes

• Demand forecastingCapex and data center project tracking

  Measure medium-term AI infrastructure demand using cloud provider capex, announced project investment and power capacity.

  The report aggregates data center projects from major cloud providers, new cloud providers and sovereign capital, and combines changes in consensus expectations to judge construction cycles and hardware demand direction; it also notes that capex is affected by leasing structures and accounting policies and should not be directly equated with hardware spending.

• Market sizingGPU and ASIC accelerator model

  Estimate the XPU market by shipments, market share, advanced packaging and HBM value.

  The model distinguishes between GPUs and custom ASICs, and separately evaluates the impact of TPUs, CoWoS packaging and HBM; because some suppliers include HBM in chip revenue, HBM prices and procurement methods can significantly affect market size and share estimates.

• Stock screeningCross-comparison of supply-chain earnings and valuations

  Compare share-price performance, changes in valuation multiples, earnings estimate revisions and supply-chain positions at the same time.

  The report covers 29 supply-chain companies and identifies different situations such as share-price gains despite valuation compression, upward earnings revisions and beneficiaries of capacity bottlenecks, avoiding judgments on fundamentals based solely on short-term share-price momentum.

Asset mapping & comparison

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

• META

  Major cloud provider and AI data center capex entity, rated Outperform with a US$800 target price.

  Strengths

  Large-scale projects such as Hyperion continue to expand, and external capital is being introduced through joint-venture structures; AI infrastructure investment and commercialization capabilities are strong.

  Weaknesses

  Capex is massive, and returns on hardware investment need continued validation.

  Comparison

  Compared with some pure leasing cloud providers, Meta has both self-build capability and partners such as Blue Owl and BlackRock to share funding pressure.

  Risks

  Excessive AI capacity buildout, project delivery delays, advertising business volatility and capital returns below expectations.

• AMD

  GPU, CPU and AI server rack supplier, rated Outperform with a US$600 target price.

  Strengths

  Helios racks are expected to enter mass production at the end of 2026, ZT Systems strengthens end-to-end rack design capability, and agentic AI is also favorable for CPU demand.

  Weaknesses

  Its AI GPU ecosystem and training performance still face NVIDIA's leading advantage.

  Comparison

  Compared with NVIDIA, AMD has lower market share, but rack integration capability and multi-cloud customer adoption provide room to catch up.

  Risks

  Product ramp-up below expectations, software ecosystem gap, customer validation delays and competitive price cuts.

• NVDA

  Core supplier in the AI GPU and rack market, rated Outperform with a US$315 target price.

  Strengths

  Leading performance in training workloads and software ecosystem, with GPUs expected to still account for the main share of the XPU market.

  Weaknesses

  Kyber rack PCB backplane design and yield issues could delay deliveries, while custom ASIC penetration is rising.

  Comparison

  Compared with ASICs, NVIDIA has advantages in generality and training performance, but faces competition in inference cost and customer autonomy.

  Risks

  Rack delays, changes in architecture specifications, ASIC substitution, rising HBM costs and concerns over returns on AI investment.

• MediaTek

  Beneficiary of Google's TPU custom ASICs, rated Outperform with a target price of NT$4,380.

  Strengths

  Has secured two TPU projects, with the first project scheduled for production in 2026, and has raised 2026 ASIC revenue expectations to more than US$2 billion.

  Weaknesses

  Revenue is highly dependent on the progress of major customer projects and advanced packaging resources.

  Comparison

  Compared with the traditional consumer electronics business, data center ASICs provide higher medium-term growth flexibility.

  Risks

  Project delays, customer concentration, changes in HBM procurement models and insufficient advanced packaging capacity.

• Unimicron and Delta

  Near-term and long-term preferred Asian hardware names, respectively, both rated Outperform.

  Strengths

  Unimicron benefits from PCB and IC substrate bottlenecks and customer support for capacity expansion; Delta benefits from long-term upgrades in data center power and infrastructure.

  Weaknesses

  After earlier share-price gains, valuations and expectations are relatively high, making them sensitive to capacity ramp-up and order delivery.

  Comparison

  Unimicron is more of a near-term capacity-tightness trade, while Delta is more tied to the long-term logic of rising power density and infrastructure upgrades.

  Risks

  Rack specification changes, customer project delays, overly rapid capacity expansion and continued industry valuation compression.

• Memory, storage, PCB, IC substrates and test equipment

  High-boom bottleneck segments in the AI server supply chain.

  Strengths

  Some sample stocks have risen about 100% to 200% in 2026, while customers are supporting capacity expansion through prepayments, capex sharing and long-term agreements.

  Weaknesses

  Share-price volatility and cyclicality are high, and valuation compression in the memory and PCB sectors has been notable recently.

  Comparison

  Compared with cloud provider stocks, these segments have shown stronger relative performance year to date, but are more sensitive to changes in supply and demand and the capex cycle.

  Risks

  Excess new capacity, price declines, specification changes, substandard yields and customer concentration.

• CoreWeave

  GPU cloud service provider, rated Underperform with a US$67 target price.

  Strengths

  Backlog is about US$100 billion, and the build-to-demand model provides strong revenue visibility.

  Weaknesses

  2026 capex guidance is as high as US$31 billion to US$35 billion, creating significant financing needs and balance-sheet pressure.

  Comparison

  Compared with large cloud providers that combine self-build and leasing, CoreWeave is more dependent on external financing, customer contracts and leased data center capacity.

  Risks

  Rising financing costs, customer concentration, capex overruns, contract execution and oversupply of AI compute.

Key data

• Planned and under-construction data center investmentAbout US$1.18 trillionFurther increased from about US$960 billion in April 2026.
• Forecast adjustment for 2026 capex by major cloud providersRaised by about 15% versus March 2026Reflects the market's continued increase in AI infrastructure investment expectations after second-quarter results.
• 2027 capex by cloud providers, new cloud providers and OracleAbout US$1.11 trillionExpected CAGR of about 61% during 2025–2027.
• 2027 XPU market sizeAbout US$500 billionIncludes GPUs and ASICs, with GPUs still accounting for the majority.
• ASIC share of the XPU market in 2027About 20%Previously about 10%, with inference demand and cost efficiency driving higher penetration.
• 2026 TPU shipment growthOver 90% year over yearExpected to account for about 42% of CoWoS-packaged ASIC shipments.
• 2028 AI data center ASIC marketAbout US$140 billion to US$160 billionIncludes HBM; excluding HBM, the size is roughly halved.
• Server shipment forecastTotal server CAGR of 15% and high-end GPU AI server CAGR of 22% during 2025–2028The global server market is expected to exceed US$1 trillion in 2028.
• Nvidia rack shipment forecast61,000 racks in 2026 and 88,000 racks in 2027Rubin racks and AMD Helios racks are expected to begin shipping in 4Q26.
• 2Q26 AI primary-market financingUS$147 billionUp 213% year over year, accounting for about 70% of all venture investment in the quarter.

Impact & implications

At the industry level, AI infrastructure investment remains in an expansion phase, but the beneficiaries are spreading from GPUs to ASICs, CPUs, memory, storage, PCBs, IC substrates, test equipment, power and optical interconnects. At the stock level, short-term valuation compression offers potential opportunities in quality names with continued fundamental upgrades, but high capex, financing dependence and project delays will widen stock-level dispersion. Investors should place greater emphasis on order visibility, customer funding support, capacity scarcity and earnings delivery rather than simply chasing AI theme exposure.

Risks

• Nvidia Kyber racks and PCB backplane design, yield and delivery progress fall short of expectations.
• Insufficient returns on AI investment by cloud providers and frontier model companies trigger capex cuts.
• Large AI projects depend on complex financing arrangements, and deterioration in interest rates, debt or equity financing conditions could delay construction.
• The pace of ASIC substitution for GPUs, HBM prices and changes in procurement methods lead to deviations in market size forecasts.
• Memory, PCB, substrate and test equipment manufacturers face overcapacity after expanding capacity ahead of demand.
• AI supply-chain valuations continue to compress, and earnings upgrades are insufficient to offset declining risk appetite.
• Frequent changes in rack architecture, cooling, optical interconnects and accelerator specifications cause inventory or R&D losses.

What to watch

• Capex guidance and financing plans from major cloud providers over the next 12 months.
• Financing and capacity cooperation arrangements among companies such as OpenAI, Nvidia, Oracle and CoreWeave.
• Design finalization, yield and mass production progress for Nvidia Rubin and Rubin Ultra racks.
• Adoption of AMD Helios racks at Meta, Microsoft, Oracle and other customers.
• Mass production ramp-up of Google TPUv8, Amazon Trainium3 and MediaTek TPU projects.
• Actual demand pull from agentic AI for CPUs and general-purpose servers.
• Supply bottlenecks in T-glass, ABF substrates, HBM, PCB backplanes and advanced packaging.
• Commercial shipment pace of power racks, co-packaged optics and near-packaged optics.
• Whether supply-chain companies' 2026–2027 earnings expectations continue to be revised upward.