AI server supply scarcity remains the core theme, while capital spending, ASICs, and memory fundamentals are being revised upward in tandem
AI summary card
AI server supply scarcity remains the core theme, while capital spending, ASICs, and memory fundamentals are being revised upward in tandem
Bernstein believes that the core opportunities in the 2026 AI server chain are expanding from pure GPUs to data center capital expenditures, ASICs/TPUs, HBM memory, PCB/CPO, power and cooling, and selected hardware suppliers, with supply-constrained segments significantly outperforming in both share price and valuation.
- Capital expenditure expectations for major CSPs, neo-clouds, and Oracle have been revised sharply upward, with combined 2027 capex now expected to reach about US$822B, significantly above the November 2025 forecast.
- Total planned and under-construction data center investment is about US$960B, with projects from Oracle, Amazon, Meta, Microsoft, CoreWeave, and others continuing to move forward.
- Inference workloads reached an inflection point in late 2025, accelerating ASIC demand; the report expects ASIC share in the XPU market to rise from about 10% last year to about 20% by 2027.
- GPUs still account for the majority of AI accelerators, and NVIDIA maintains a leading advantage in training and full-system integration; however, ASIC solutions such as TPUs and Trainium are growing rapidly in inference scenarios.
- High memory prices have not weakened AI server demand, and supply-demand conditions for both HBM and conventional memory remain tight, continuing to benefit memory suppliers.
- Year to date, scarcity-driven segments such as PCB, CPO, memory, and test equipment have risen about 100%-160%, significantly outperforming the 0%-16% gains of CSPs and ODMs/OEMs.
Report interpretation
Overview
This report is Bernstein's 1Q26 AI Server Pulse, continuing the 4Q25 AI tracker framework and focusing on AI capital spending and data center projects, GPU and ASIC shipment progress, updates to the AI server market model, and the financial performance of AI supply chain companies. The core conclusion is that global AI infrastructure buildout has not slowed; instead, scarcity continues to show up in capital spending, project commitments, and supply chain pricing. Investors are rotating from large cloud vendors and ODMs/OEMs toward supply-constrained segments such as PCB, CPO, memory, testing, power, and cooling.
Core views
The report believes the AI server supply chain remains in an upward cycle. Consensus 2026 capital expenditure forecasts for the four major U.S. hyperscalers are about 30% higher than in November 2025; including neo-clouds and Oracle, combined 2027 capex is expected to reach about US$822B. Planned and under-construction data center investment is around US$960B, with Oracle, Amazon, Meta, Microsoft, and CoreWeave continuing to advance large-scale projects. Inference demand reached an inflection point in late 2025, accelerating demand for ASICs and TPUs, but GPUs remain the core of the AI accelerator market, with NVIDIA retaining advantages in training performance, HBM4 adoption, and full-system integration. High memory prices have not disrupted AI server demand and instead reinforce the earnings leverage of memory suppliers.
Analysis framework
The report uses a cross-industry supply chain tracking approach, combining CSP/neo-cloud capital spending, data center project progress, GPU and ASIC shipment models, monthly sales of Taiwanese ODMs/OEMs, memory prices and HBM shipments, and the stock performance of the AI server supply chain to assess the impact of AI infrastructure investment on semiconductor, hardware, internet, and software companies.
Methodology notes
Uses capital expenditure expectations and order commitments from companies such as Amazon, Google, Microsoft, Meta, Oracle, CoreWeave, and Nebius to judge the intensity of AI data center construction.
The report compares market consensus between November 2025 and after 4Q25 earnings, noting that 2026 capex forecasts were revised up by about 30%, with combined 2027 capex expected to reach about US$822B.
Estimates the size of the global server market using high-end GPU servers, rack shipments, number of chips deployed, and server ASPs.
The report expects the global server market and high-end GPU AI server shipments to achieve CAGRs of about 7% and 35%, respectively, from 2025 to 2027, and forecasts NVIDIA server rack shipments of about 61K/88K in 2026/2027.
Compares the evolving roles of GPUs and ASICs in training, inference, cost, energy efficiency, system integration, and supply chain dependence.
The report believes growth in inference workloads is driving ASIC demand, but GPUs still maintain leadership in training and high-performance system integration; ASIC market share is expected to rise from about 10% to about 20% by 2027.
Explains the rise/fall and valuation changes of AI server supply chain stocks through supply scarcity, technology penetration, and earnings revisions.
The report notes that sectors such as PCB, CPO, memory, and test equipment are up about 100%-160% year to date, while CSPs and ODMs/OEMs are up about 0%-16%, reflecting investor preference for scarce segments.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- NVIDIA CORP (US.NVDA)Core beneficiary of AI GPUs and server racks
- Strengths
- Leading training workload performance, Rubin uses HBM4, strong NVLink and full-system integration capabilities, and the data center opportunity remains large and still in an early stage.
- Weaknesses
- Rubin rack shipments were lowered to about 3K in 4Q26, and some inference demand may shift to ASICs.
- Comparison
- Compared with ASICs, GPUs still account for the majority of the AI accelerator market, especially for training and large-scale general-purpose computing.
- Risks
- Supply chain delays, geopolitical conflict, customer capex pacing, and ASIC substitution risk.
- ASIC/TPU ecosystemHigh-growth alternative compute path at the inference demand inflection point
- Strengths
- CSPs are adopting ASICs to improve performance-per-dollar, reduce TCO, and decrease dependence on external GPU supply chains; TPU is expected to become the largest component of the ASIC market.
- Weaknesses
- ASIC use cases are more specialized, with weaker ecosystem and software generality than GPUs.
- Comparison
- GPUs remain better suited for training and general-purpose high-performance computing, while ASICs are growing faster in inference and customized scenarios.
- Risks
- Project execution, advanced packaging and HBM supply, and changes in customer demand.
- MediaTek (2454.TW)Direct beneficiary of TPU projects
- Strengths
- The report says the company is overcoming execution issues and is expected to push its first TPU project into mass production within the year.
- Weaknesses
- Sensitive to a single large project and execution timing.
- Comparison
- Compared with its traditional mobile chip business, AI ASIC projects provide higher growth elasticity.
- Risks
- Mass production delays, customer project adjustments, and intensifying competition.
- Memory and HBM suppliersBeneficiaries of high-bandwidth memory and conventional memory demand from AI servers
- Strengths
- AI demand remains strong, high memory prices have not weakened server demand, and HBM shipments and pricing support earnings.
- Weaknesses
- Cyclicality remains high and the segment is sensitive to supply-demand turning points.
- Comparison
- Compared with ODMs/OEMs, memory suppliers benefit more directly from tight supply and rising prices.
- Risks
- Delays in HBM4/Rubin timing, overly rapid supply expansion, and slowing downstream capital spending.
- Oracle (US.ORCL)Beneficiary of AI data center and cloud infrastructure expansion
- Strengths
- RPO increased to US$553B, with large AI contracts driving growth; the report believes Oracle could become the third-largest hyperscaler.
- Weaknesses
- Data center construction cycles are long, and capital requirements and customer contract execution are complex.
- Comparison
- Compared with traditional CSPs, Oracle is more dependent on leasing, customer prepayments, and expansion of dedicated AI training/inference data centers.
- Risks
- Financing, project delivery, customer concentration, and risk around capex return realization.
- CoreWeave (US.CRWV)Neo-cloud compute supply provider
- Strengths
- Revenue backlog continued rising from 4Q25, with new contracts from Meta, Anthropic, Jane Street, and others, using a build-to-order model.
- Weaknesses
- The report rates it Underperform with a target price of $67; capital expenditures are high and it depends on contracted customers.
- Comparison
- Compared with large CSPs, CoreWeave has higher upside elasticity but also higher financial leverage and customer concentration risk.
- Risks
- Financing costs, contract fulfillment, GPU supply, customer concentration, and utilization risk.
- PCB, CPO, power, and cooling supply chainScarce segments in AI servers
- Strengths
- Benefiting from higher AI server dollar content and supply constraints, with strong year-to-date valuation and share price performance.
- Weaknesses
- Some gains are already large, and order and capacity delivery still need to be verified.
- Comparison
- Compared with CSPs and ODMs/OEMs, scarce component segments have posted much higher gains year to date.
- Risks
- Easing supply-demand conditions after capacity expansion, changes in technology roadmaps, and customer order cuts.
Key data
- Planned and under-construction data center investmentabout US$960BFurther increased from the roughly US$840B figure in November 2025.
- Combined 2027 capital expenditures of CSPs, neo-clouds, and Oracleabout US$822BEquivalent to roughly 38% CAGR from 2025 to 2027.
- Adjustment to 2026 capital expenditure expectations for the four major U.S. hyperscalersup about 30% versus November 2025Market consensus was significantly revised upward after 4Q25 earnings.
- AI data center accelerator marketabout US$190B to about US$500B by 2027The XPU definition includes GPUs and ASICs and includes HBM but excludes XPU attach.
- ASIC market sharerising from about 10% to about 20% by 2027The inference workload inflection point is accelerating ASIC adoption.
- TPU market opportunityabout US$30B in 2026, about US$70B in 2027, and about US$90B in 2028The report expects TPU to approach about 70% of the ASIC market in 2027 and still account for about 60% in 2028.
- High-end GPU AI server shipment growthabout 48% growth in 2026, followed by about 23% growth in 2027Mainly driven by a smooth ramp of GB300 racks and subsequent demand.
- NVIDIA server rack shipmentsabout 61K in 2026 and about 88K in 2027Around 6.6M NVIDIA chips are expected to be deployed into downstream servers in 2026.
- Global server market sizeabout US$520B in 2026, approaching US$900B in 2028Driven by high-end AI servers, Vera Rubin/Rubin Ultra racks, and ASIC servers.
- AI financingUS$202B in 2025, with 1Q26 exceeding the full-year 2025 amountOpenAI raised about US$122B in 1Q26, with its valuation increasing from US$500B to US$730B.
Impact & implications
The investment implication is that the bottleneck in AI infrastructure buildout is spreading from cloud budgets to supply chain capacity, and the market is increasingly willing to pay higher valuations for scarce capacity, critical materials, advanced packaging, memory, optical interconnects, power, and cooling. NVIDIA remains a core beneficiary in GPU training and full-system solutions, but inference demand will expand opportunities across ASIC/TPU, the MediaTek and Broadcom ecosystems, and the in-house chip ecosystems of Google and Amazon. Memory, HBM, and server component suppliers benefit from strong demand and elevated pricing; although ODMs/OEMs are growing revenue, weaker profitability and valuation compression remain differentiating factors.
Risks
- Geopolitical events such as U.S.-Iran conflict may affect individual data center facilities or market risk appetite.
- Large AI data center projects require ongoing financing; if financing arrangements or customer prepayments fall short of expectations, construction pace may slow.
- If key chips and server racks such as Rubin, HBM4, TPU, and Trainium are delayed, supply chain revenue recognition may be affected.
- ASIC growth may divert some incremental GPU demand and change profit allocation across the supply chain.
- Although ODMs/OEMs are growing revenue, profitability, inventory, and valuation compression may still weigh on stock performance.
- Scarce segments such as PCB, CPO, memory, and test equipment have already risen significantly; if supply eases or demand expectations are revised down, valuation pullback risk will be high.
- Capital expenditures are not always a precise proxy for hardware spending, as some companies include land, buildings, and improvement expenditures in capex.
What to watch
- Financing arrangements and compute procurement contracts of AI supply chain companies such as OpenAI, NVIDIA, and Oracle.
- CSP capex guidance, financing plans, and budget revisions for 2026-2027.
- Execution progress of large projects such as Stargate, Oracle data center construction, Amazon government AI/HPC infrastructure, and Meta Hyperion.
- Timing of Rubin chip and rack launches, GB300 volume ramp, and Vera Rubin mass production progress in 4Q26.
- Progress of Google TPUv8, Amazon Trainium3, and ASIC procurement projects by Meta and Anthropic.
- Whether supply of T-glass, ABF substrate, HBM, CoWoS, CPO, power racks, and cooling remains tight.
- Monthly sales, inventory days, and margin changes of Taiwanese ODMs/OEMs.
- Whether memory prices, HBM shipments, and HBM4 timing continue to support memory supplier profitability.