AI and General-Purpose Servers Enter a Multi-Year Upcycle, with Supply Capacity Determining Actual Shipment Fulfillment

Overview

The report examines the demand outlook for the Asian hardware supply chain against the backdrop of U.S. cloud service providers increasing AI capital expenditure, AI shifting from training toward inference, and the rise of agentic AI. Its core conclusion is that demand for both AI and general-purpose servers is strengthening and will spread to CPUs, power, cooling, and networking equipment. Supply constraints will determine near-term shipment fulfillment, while the PC segment faces weakening demand in the second half of 2026, although brand vendors' margins may be better than the market fears.

Core views

First, the report bases its upward demand revisions on continued capital expenditure increases by U.S. cloud service providers. J.P. Morgan's U.S. hardware team now expects data center capital expenditure to grow 80% and 50% year over year in 2026 and 2027, respectively, above its previous forecasts of 63% and 40%. This marks the third upward revision this year; based on the latest outlooks from leading hyperscale cloud service providers, current forecasts still have approximately 10% upside potential. Although negative free cash flow at leading cloud service providers over the coming quarters is viewed as inevitable, the report believes that their resilient core businesses, strong AI demand indicators, and acceptable and continuously improving returns from AI businesses can still support continued investment. Tier-two cloud service providers and other data center companies have also become important sources of incremental growth, with their capital expenditure expected to rise 65% and 39% in 2026 and 2027, respectively, reaching US$90 billion and US$130 billion. The report estimates cumulative AI capital expenditure of US$5.5 trillion from 2026 to 2030, of which US$4.1 trillion will be debt-financed, while the investment-grade corporate bond market may provide more than US$2.1 trillion for data center investment over the next five years. Enterprise demand also supports the expansion of AI investment. J.P. Morgan's 2026 Asia-Pacific AI implementation survey shows that 88% of respondents incurred AI expenditure over the past 12 months, approximately 3% had not yet invested, and approximately 7% were uncertain. Average AI expenditure as a proportion of operating expenses plus capital expenditure is expected to rise from 4.5% over the past 12 months to 5.8% over the next 12 months. Respondents primarily aim to improve productivity and drive growth, while cost reduction remains a secondary use case. Most respondents remain optimistic about the impact on profits, but the report emphasizes that execution capabilities will increasingly become the dividing line between AI winners and laggards. For AI servers, the report expects AI chip shipments to grow at a 45% CAGR from 2025 to 2028, including a 60% CAGR for AI ASIC chip shipments. The latter will be driven by demand for AWS Trainium and Google TPU, with year-over-year growth of 77% and 88% in 2026 and 2027, respectively. As AI inference scales, more end customers may develop proprietary inference ASICs, expanding the overall ASIC server market. GPU server demand is supported by resilient spending from U.S. cloud service providers and backlog demand from emerging cloud providers and enterprises, with demand visibility extending into 2027. Nvidia Vera Rubin systems are expected to ramp as scheduled from late in the third quarter through the fourth quarter of 2026, with yield issues at leading PCB manufacturers having a limited effect on the overall timetable. The report expects shipments of 70,000 to 80,000 NVL72 racks in 2026, including 60,000 to 65,000 GB300 racks and approximately 10,000 VR racks. Rubin chip board-level production is expected to reach approximately 1.5 million units, but production cycles and component shortages will limit deliveries of VR NVL72 racks to approximately 10,000 units. NVL72 rack shipments are expected to increase to 85,000 to 95,000 units in 2027, primarily comprising VR200. Estimated ODM prices for GB300 and VR200 racks are approximately US$4.1 million and US$6.5 million, respectively, about 20% and 90% higher than GB200, mainly due to higher GPU and memory prices and upgraded networking, power, and cooling configurations. Correspondingly, ODM gross margins are expected to decline from 4.2% for GB200 to 3.6% for GB300 and 2.7% for VR200. However, due to more complex architectures and longer testing times, gross profit per rack is expected to rise from approximately US$145,000 for GB200 to approximately US$148,000 for GB300 and approximately US$175,000 for VR200. The competitive landscape for server ODMs may remain concentrated. Although tier-two ODMs such as Pegatron and Compal are entering the market, complex design capabilities, frequent product upgrades, and high working capital requirements constitute barriers. The report identifies Hon Hai and Quanta as preferred names in the NVL72 supply chain and views Celestica and Wiwynn as leading ASIC server manufacturers. Wiwynn also covers GPU, AI ASIC, and general-purpose servers, providing exposure across multiple platforms. General-purpose servers are another core theme of the report. Agentic AI requires CPUs to handle orchestration, tool and skill execution, storage, and security tasks. The Vera CPU is used not only in NVL72 but also as a standalone processor for storage and CPU servers on the Vera Rubin platform. The report divides CPU demand into three categories: AI server head-node CPUs, standard server CPUs, and agentic AI CPUs. It believes that as AI shifts from training toward inference, head-node and agentic AI demand will drive a "super" general-purpose server CPU cycle. The report estimates that general-purpose server demand will grow 30% to 40% year over year in 2026, while server CPU shipments will increase from 26 million units in 2025 to 68 million units in 2028, representing a 38% CAGR. Demand for the three CPU categories is expected to grow at CAGRs of 74%, 5%, and 155%, respectively, from 2025 to 2028. Driven jointly by shipment expansion and an approximately 10% CAGR in server CPU average selling prices, the server CPU revenue TAM is expected to grow at a 53% CAGR over the same period. Data center CPU revenue growth at Intel and AMD has accelerated over the past three quarters. AMD's outlook for the server CPU TAM to reach US$220 billion by 2030, implying a CAGR of approximately 50% or more from 2025 to 2030, is also cited by the report as supporting evidence for a multi-year cycle. Recent server shipments have already improved: U.S. cloud service provider server shipments increased sequentially in the first quarter of 2026 and outperformed seasonality, while overall shipments in the second quarter appear to have achieved double-digit sequential growth. Demand from U.S. cloud service providers is expected to remain the primary driver over the next six months. Order visibility at Lotes and ASPEED has extended into 2027, while memory suppliers have received feedback from U.S. cloud service providers indicating 50% to 80% demand growth in 2027. The report expects server shipment growth to accelerate to 25% in 2027, driven by backlogged orders from 2026 and strong demand. Its forecasts for cloud service provider and enterprise server shipments in 2026/2027 are growth of 40% to 50% and a slight decline to flat performance, respectively. Overall server shipments may continue to achieve double-digit growth in 2028–2029, but fulfillment will depend primarily on supply chain constraints. Server growth is further increasing demand for power and cooling. The report expects the data center power TAM to reach approximately US$50 billion by 2028, representing an 80% CAGR from 2025 to 2028. Delta's AC/DC server power supply revenue is expected to grow at a 75% to 80% CAGR over the same period. Growth stems from three compounding factors: AI accelerator shipments growing at a 40% to 50% CAGR based on the latest CoWoS estimates, accelerator thermal design power growing at a 30% to 50% CAGR as specifications are upgraded, and the migration of power architecture toward standalone power racks. The report expects power rack adoption of approximately 20% during the Vera Rubin cycle, rising to more than 50% during the VR Ultra cycle. Based on the latest updates, the forecast CAGR for AI accelerator shipments has been raised from approximately 30% to approximately 45%, while accelerator TDP is expected to grow at an approximately 30% CAGR. As ASICs are used for both training and inference, their TDP may increase even faster. Data center networking is also viewed as a new growth market. The report estimates that Accton's serviceable market across the four major U.S. cloud service providers, including general-purpose Ethernet switches and AI front-end and back-end networking, will expand from approximately US$6 billion in 2025 to approximately US$33 billion in 2028, representing a CAGR of approximately 75%. Drivers include growth in general-purpose server shipments from AI inference, front-end and back-end network upgrades, and shorter upgrade cycles. The report expects cloud service provider general-purpose server shipments to grow 40%, 25%, and 20% in 2026, 2027, and 2028, respectively, with agentic AI potentially providing further upside. Over the past two to three years, cloud service providers prioritized investment in AI servers and data center construction, while network investment lagged. Gartner data show that spending on directly procured ODM servers grew 207% and 127% in 2024 and 2025, respectively, while switch spending increased only 18% and 103% over the same periods. Supplemental investment required to reduce latency and increase transmission speeds is therefore expected to support white-box Ethernet switches. AI data volumes and bandwidth requirements are not only driving upgrades to traditional front-end networks but also creating incremental GPU back-end networking demand. The report expects the GPU back-end Ethernet ToR switch TAM to grow at an approximately 164% CAGR from 2025 to 2028. Specific upgrades include the migration from BlueField 3 400G used in GB200/GB300 to BlueField 4 800G in VR200, as well as cloud service providers upgrading proprietary network interface cards from 200G to 400G. The network iteration period has shortened from seven years for the migration from 100G to 200G to less than two years for the migration from 800G to 1.6T. A higher proportion of high-speed ports should support increases in average selling prices. Accton is viewed as a key beneficiary due to its position in white-box data center switches. The related business is expected to account for approximately 40% of its revenue in the third quarter of 2026, while its average selling price per switch port is expected to grow at an approximately 20% CAGR from 2025 to 2028. The report's view on the PC segment contrasts with its view on servers. PC shipments in the first half of 2026 exceeded expectations, potentially due to Windows 10 replacement demand and orders being pulled forward ahead of memory price increases. The report remains cautious about the second half because demand has already been brought forward, price increases for key components are generating negative price elasticity, and the Windows 10 replacement cycle is entering its later stages. The report expects PC shipments in the second half of 2026 to decline 3% versus the first half and full-year shipments to fall 8%, including a 14% decrease in consumer PCs and a 4% decrease in commercial PCs. Commercial demand is relatively more resilient due to replacement demand and lower price sensitivity. Although PC sales volumes are under pressure, brand vendors' value capture and margins may exceed expectations. Margins at Asian PC brands in the first half of 2026 were supported by demand outperforming concerns, price increases, and product mix adjustments. PC average selling prices rose 20% to 30% year over year, and together with lower-cost inventory, drove substantial margin expansion. The bill-of-materials cost for high-end notebook computers increased by approximately 30% over the past year. DRAM and NAND prices rose three- to fivefold, increasing memory costs by US$200 to US$250 and lifting memory's share of the total system bill of materials from the high-single-digit percentage range a year earlier to 20% to 30%. The report separately assumes a 5% price increase for other components and non-core semiconductors. Brand vendors consequently implemented product price increases of 20% to 30% in the first half of 2026 and mitigated costs through specification reductions and product mix adjustments. ODM gross margins were diluted by higher system selling prices, but profit per unit remained broadly stable. Even if demand weakens in the second half, leading brands may maintain relatively resilient margins through pricing, product mix, and supply chain capabilities.

Analysis framework

The report first validates aggregate demand through the capital expenditure and financing capabilities of U.S. cloud service providers and tier-two data center operators, as well as an Asia-Pacific enterprise AI survey, and then projects the server cycle based on GPU, ASIC, and CPU shipments. It subsequently maps demand to the ODM, power, cooling, and networking markets using rack and notebook computer bill-of-materials breakdowns, shipment and average selling price decompositions, technology roadmaps, and supply chain research, while cross-validating server CPU shipments through top-down forecasts and bottom-up estimates. Finally, the report compares PC sales volumes, cost pass-through, and margins for brands and ODMs on a segmented basis.

Methodology notes

• Industry/Sector Analysis FrameworkSupply-demand framework

  Demand and supply constraint analysis

  The report assesses demand through cloud capital expenditure, enterprise AI adoption, and order backlogs, while treating component shortages, capacity, and production cycles as key variables determining actual delivery volumes.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Decomposition of shipments, power consumption, configurations, and average selling prices

  The report separately estimates accelerator and CPU shipments, TDP, server CPU average selling prices, and average selling prices per switch port to explain compound growth in the power, CPU, and networking markets.

• Competition and Strategy FrameworkValue chain analysis

  Server and notebook computer bill-of-materials analysis

  The report breaks down the costs of GPUs, memory, networking, power, cooling, and other components to assess rack prices, ODM gross profit per rack, and changes in PC brand pricing and margins.

• Competition and Strategy FrameworkMoat / competitive advantage

  Analysis of competitive barriers for AI server ODMs

  The report uses complex design capabilities, frequent upgrades, and working capital requirements to explain why the market remains concentrated among a small number of major ODMs despite the entry of tier-two manufacturers.

• (Out-of-Vocabulary Method)

  Top-down and bottom-up cross-validation

  Server CPU shipments are first projected based on overall market drivers and then cross-checked using different CPU demand categories and supply chain estimates.

• (Out-of-Vocabulary Method)

  Asia-Pacific AI implementation survey

  The report uses enterprises' AI expenditure over the past and next 12 months, investment motivations, and profit expectations to verify whether AI demand has entered the mainstream and continues to expand.

Asset mapping & comparison

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

• Delta（2308 TT）

  The report identifies it as a key beneficiary of long-term trends in AI data center power and cooling.

  Strengths

  AC/DC server power supply revenue is expected to grow at a 75% to 80% CAGR from 2025 to 2028, broadly consistent with the approximately 80% growth rate of the server power TAM.

  Comparison

  Its revenue growth forecast broadly tracks the overall server power TAM.

  Risks

  Growth fulfillment is affected by accelerator shipments, TDP increases, power rack adoption, and supply capacity.

• Hon Hai（2317 TT）

  The report identifies it as one of the preferred companies in the NVL72 supply chain.

  Strengths

  It operates within the highly concentrated NVL72 ODM supply chain, which has substantial design and working capital barriers.

  Comparison

  It is identified alongside Quanta as a preferred name in the NVL72 supply chain.

  Risks

  AI server component shortages and declining ODM gross margins.

• Quanta（2382 TT）

  The report identifies it as one of the preferred companies in the NVL72 supply chain.

  Strengths

  It possesses competitive barriers derived from complex AI server design and upgrade capabilities.

  Comparison

  It is identified alongside Hon Hai as a preferred name in the NVL72 supply chain.

  Risks

  AI server supply constraints and pressure on rack ODM gross margins.

• Wiwynn（6669 TT）

  It covers GPU, AI ASIC, and general-purpose servers and is identified as one of the leading ASIC server manufacturers.

  Strengths

  Its business spans multiple AI computing architectures and general-purpose servers, allowing it to capture different types of server demand.

  Comparison

  The report views it and Celestica as leaders in the ASIC server segment.

  Risks

  Supply chain constraints and declining AI server ODM gross margins.

• Celestica

  The report identifies it as a leading manufacturer in the AI ASIC server segment.

  Strengths

  It benefits from growth in demand for AWS Trainium, Google TPU, and potentially proprietary inference ASICs.

  Comparison

  It and Wiwynn are leaders in the ASIC server segment.

  Risks

  ASIC shipment growth and customer-developed projects may progress more slowly than the report expects.

• ASPEED（5274 TT）

  The report believes supply releases in the second half of 2026 could provide revenue upside.

  Strengths

  Order visibility has extended into 2027, allowing it to benefit from strengthening server demand.

  Weaknesses

  Near-term revenue fulfillment depends on supply releases.

  Risks

  Supply capacity may not be released on schedule.

• Lenovo（992 HK）

  The report notes that the inflection point in its ISG margin has been confirmed and its AI project pipeline has doubled.

  Strengths

  Its AI project pipeline is expanding, and ISG margins are showing signs of improvement.

  Weaknesses

  It also faces weakening PC end demand in the second half of 2026.

  Comparison

  Leading PC brands are expected to have more resilient margins due to their stronger supply chain capabilities.

  Risks

  PC demand being pulled forward, component price increases, and negative price elasticity.

• Accton（2345 TT）

  The report views it as a key beneficiary of AI-driven white-box data center switch upgrades and the new server rack market.

  Strengths

  It has a strong position in white-box data center switches, with the related business expected to account for approximately 40% of revenue in the third quarter of 2026 and average selling prices per port expected to grow at an approximately 20% CAGR from 2025 to 2028.

  Comparison

  Its serviceable market is expected to increase from approximately US$6 billion in 2025 to approximately US$33 billion in 2028.

  Risks

  Catch-up network investment or the pace of 800G and 1.6T upgrades may fall short of expectations.

• Lotes（3533 TT）

  It benefits from strong server demand and has optionality in its AI business.

  Strengths

  Order visibility has extended into 2027.

  Weaknesses

  The report notes near-term pressure on gross margins.

  Risks

  Near-term gross margin headwinds and server supply constraints.

Key data

• AI chip shipment CAGR45%2025–2028 forecast
• AI ASIC chip shipment CAGR60%2025–2028 forecast; year-over-year growth of 77% and 88% in 2026 and 2027, respectively
• U.S. data center capital expenditure growth2026 +80%, 2027 +50%Previous forecasts were +63% and +40%, respectively; the third upward revision this year, with the latest outlook indicating approximately 10% additional upside
• Tier-two data center operator capital expenditureUS$90 billion in 2026, US$130 billion in 2027Year-over-year growth of 65% and 39%, respectively
• AI capital expenditure and debt financingUS$5.5 trillion; of which US$4.1 trillion will be debt-financedCumulative estimate for 2026–2030
• AI expenditure adoption rate among Asia-Pacific enterprises88%Respondents incurred AI expenditure over the past 12 months; approximately 3% had not invested and approximately 7% were uncertain
• Average AI spending intensityIncrease from 4.5% to 5.8%As a percentage of operating expenses plus capital expenditure, rising from the past 12 months to expectations for the next 12 months
• 2026 NVL72 rack shipments70,000 to 80,000 unitsIncluding 60,000 to 65,000 GB300 units and approximately 10,000 VR units
• 2027 NVL72 rack shipments85,000 to 95,000 unitsVR200 is expected to account for the majority
• GB300 and VR200 rack ODM pricesApproximately US$4.1 million and US$6.5 millionApproximately 20% and 90% higher than GB200, respectively
• GB300 and VR200 ODM gross margins3.6% and 2.7%Below GB200's 4.2%
• Server CPU shipmentsIncrease from 26 million to 68 million unitsFrom 2025 to 2028, representing a 38% CAGR
• Server CPU revenue TAM CAGR53%2025–2028, including an approximately 10% CAGR in server CPU average selling prices
• 2027 server shipment growth25%Driven by backlogged orders and strong server demand
• Data center power TAMApproximately US$50 billion2028 forecast; 80% CAGR from 2025 to 2028
• Accton data center networking serviceable marketIncrease from approximately US$6 billion to approximately US$33 billionFrom 2025 to 2028, representing an approximately 75% CAGR
• GPU back-end Ethernet ToR switch TAM CAGRApproximately 164%2025–2028 forecast
• 2026 PC shipmentsDown 8% for the full yearConsumer PCs down 14% and commercial PCs down 4%; second half down 3% versus the first half
• High-end notebook computer bill-of-materials costIncrease of approximately 30%Over the past year; memory costs increased by US$200 to US$250
• PC brand product price increases20% to 30%In the first half of 2026, to offset rising bill-of-materials costs

Impact & implications

The report believes that AI investment is spreading from GPU servers to ASICs, general-purpose server CPUs, power, cooling, and networking. The Asian hardware supply chain therefore faces not a single-product cycle but a multi-year demand cycle in which multiple infrastructure segments reinforce one another. Near-term revenue fulfillment will depend on whether component supply and capacity can be released. Meanwhile, AI server ODMs may experience declining gross margins but higher gross profit per rack. The PC industry shows a divergence between weak volume and strong pricing: sales volumes and end demand face pressure, while leading brands may maintain relatively resilient margins through price increases, product mix, and supply chain capabilities.

Risks

• Whether server demand can translate into actual shipments depends on components, capacity, and production cycles, making supply chain constraints the most critical swing factor.
• Vera Rubin-related PCB yield issues currently have a limited impact, but their potential effect on the subsequent ramp-up timetable still requires monitoring.
• The Kyber architecture faces challenges that may affect the related future product roadmap.
• The report views negative free cash flow at leading cloud service providers over the coming quarters as inevitable, while continued investment still depends on core businesses, AI demand, and improving returns.
• AI server ODM gross margins for GB300 and VR200 are expected to be lower than for GB200.
• PC demand in the second half of 2026 may weaken due to orders being pulled forward into the first half, negative price elasticity caused by component price increases, and the later stages of the Windows 10 replacement cycle.
• Rising memory and other component prices have a greater impact on low- and mid-range PCs, while PC ODM gross margins may also continue to face dilution.

What to watch

• Monitor server component supply, capacity releases, and whether backlogged orders can support 25% shipment growth in 2027.
• Track the Vera Rubin ramp-up from late in the third quarter through the fourth quarter of 2026 and progress toward VR200 becoming the primary NVL72 product in 2027.
• Track whether capital expenditure increases by leading and tier-two U.S. cloud service providers in 2026–2027 can continue to be fulfilled.
• Monitor the incremental contribution of AI inference and agentic AI to head-node CPUs, general-purpose servers, and network investment.
• Track whether power rack adoption can increase from approximately 20% during the Vera Rubin cycle to more than 50% during the VR Ultra cycle.
• Monitor 800G-to-1.6T switch upgrades, GPU back-end Ethernet construction, and growth in Accton's average selling price per port.
• Observe the differing effects of PC shipments, brand price increases, and product mix adjustments on brand and ODM margins in the second half of 2026.