Demand for AI and general-purpose servers is strengthening in tandem, with supply constraints becoming the key variable determining growth realization

Overview

The report focuses on two diverging themes in the Asian hardware industry: demand for AI and general-purpose servers continues to strengthen due to cloud capital expenditure, inference workloads, and agentic AI, driving capacity expansion in server ODMs, CPUs, power, cooling, and networking equipment; meanwhile, after PC shipments are supported by pull-ins during the first half of 2026, sales face a decline in the second half. The report believes that whether final shipments can be realized primarily depends on components, capacity, and other supply-chain constraints.

Core views

First, the funding foundation for AI infrastructure demand has strengthened further. J.P. Morgan's U.S. hardware team raised its data center capital expenditure growth forecasts from the previous 63% for 2026 and 40% for 2027 to 80% and 50%, respectively, marking the third upward revision this year. Combined with the latest guidance from major hyperscale cloud service providers, the report believes current estimates still have approximately 10% upside. Although negative free cash flow among major cloud service providers is difficult to avoid over the coming quarters, their core businesses remain solid, AI demand indicators are strong, and returns from AI businesses continue to improve; therefore, the report concludes that investment is unlikely to be interrupted soon. Tier-two data centers and emerging cloud providers are also becoming important incremental contributors, with their capital expenditure expected to grow 65% and 39% in 2026 and 2027, respectively, reaching US$90 billion and US$130 billion. The report estimates aggregate 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 adoption validates the demand underlying this capital expenditure. J.P. Morgan's 2026 Asia-Pacific AI implementation survey shows that 88% of respondents incurred AI spending during the past 12 months, while only approximately 3% made no investment and approximately 7% were uncertain. AI investment is primarily intended to improve productivity and generate growth, with cost reduction remaining a secondary motive. Most respondents are optimistic about the impact on profits, although execution capabilities will increasingly differentiate companies. Average AI spending as a share of combined operating expenses and capital expenditure is expected to increase from 4.5% over the past 12 months to 5.8% over the next 12 months. The report also observes accelerating cloud revenue growth at the three largest U.S. cloud service providers and improving cloud operating margins even as capital expenditure rises. 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 chips. ASIC chip shipments are expected to grow 77% and 88% year over year in 2026 and 2027, respectively, primarily driven by demand for AWS Trainium and Google TPU. As inference activity expands, more end customers may develop proprietary ASICs, further expanding the ASIC server market. Nvidia Vera Rubin systems are expected to ramp on schedule from late in the third quarter through the fourth quarter of 2026, with recent yield issues at major PCB manufacturers having only a limited effect on progress; the Kyber architecture, however, still faces challenges. The report expects 70,000 to 80,000 NVL72 rack shipments in 2026, comprising 60,000 to 65,000 GB300 racks and approximately 10,000 Vera Rubin racks. Although Rubin board-level chip production is expected to reach approximately 1.5 million units that year, production cycles and component shortages will limit Vera Rubin NVL72 rack shipments to approximately 10,000 units. NVL72 shipments are expected to increase to 85,000 to 95,000 units in 2027, mainly comprising VR200. New platforms increase value per system but may not raise ODM margins in tandem. ODM prices for GB300 and VR200 racks are expected to be approximately US$4.1 million and US$6.5 million, respectively, around 20% and 90% higher than GB200, primarily due to higher GPU and memory prices and upgrades to networking, power, and cooling content. ODM gross margins are expected to decline from 4.2% for GB200 to 3.6% for GB300 and 2.7% for VR200. However, because the VR200 architecture is more complex and requires longer testing cycles, gross profit per rack still increases. The report estimates manufacturing value added per GB300 and VR200 rack at approximately US$148,000 and US$175,000, respectively, above approximately US$145,000 for GB200. The competitive landscape may remain concentrated: even if tier-two ODMs such as Pegatron and Compal enter the market, complex design capabilities, frequent iterations, and high working-capital requirements remain barriers. Hon Hai and Quanta are named as the preferred NVL72 supply-chain plays, while Celestica and Wiwynn lead in ASIC servers. AI inference is also spreading growth from accelerators to general-purpose servers and CPUs. The report divides CPU demand into three categories: head-node CPUs for AI servers, conventional server CPUs, and agentic AI CPUs. Agentic AI requires CPUs for orchestration, tool and skill execution, storage, security, and other tasks. In addition to being configured in NVL72, Vera CPUs are also used in standalone CPU servers and storage servers on the Vera Rubin platform. The report attributes approximately 30%–40% year-over-year growth in general-purpose server demand in 2026 primarily to agentic AI and expects agentic AI CPUs and head-node CPUs to jointly create a “super” server cycle as applications shift from training to inference. Total server CPU shipments are expected to increase from 26 million units in 2025 to 68 million units in 2028, representing a 38% CAGR. Over the same period, demand for head-node CPUs, conventional server CPUs, and agentic AI CPUs is expected to grow at CAGRs of 74%, 5%, and 155%, respectively. The server CPU revenue TAM is expected to grow at a 53% CAGR, including an approximately 10% CAGR in CPU average selling prices. Intel and AMD data center CPU revenue has accelerated over the past three quarters. AMD's long-term outlook for the server CPU TAM to reach US$220 billion by 2030, representing an approximately 50%-plus CAGR from 2025 to 2030, is also cited by the report as validation of supply-chain demand. System shipment data likewise points to a multi-year cycle. U.S. cloud service provider server shipments in the first quarter of 2026 outperformed seasonality and increased sequentially. Industry shipments appear to have achieved double-digit sequential growth in the second quarter and are expected to sustain that momentum through the second half. Order signals have continued to improve over the past three to six months, while visibility for Lotes and ASPEED has extended into 2027. Memory manufacturers have received feedback from U.S. cloud service providers indicating 50%–80% demand growth in 2027. The report expects server shipment growth to accelerate to 25% in 2027, driven jointly by 2026 backlogs and strong demand. By customer, cloud service provider server shipments are expected to grow 40%–50% in both 2026 and 2027, while enterprise server shipments are expected to range from a single-digit decline to flat. Overall server shipments may continue to post double-digit growth in 2028–2029, but supply-chain constraints are the most critical swing factor. Data center power is another highly elastic growth theme. The report expects this market's TAM to reach US$50 billion in 2028, representing an approximately 80% CAGR from 2025 to 2028. Delta's AC/DC server power revenue is expected to grow at a 75%–80% CAGR over the same period. Growth comes from three overlapping factors: based on the latest CoWoS data, accelerator shipments are growing at a CAGR of approximately 40%–50%; ongoing specification upgrades are driving TDP growth at a 30%–50% CAGR; and power-supply architecture is shifting from conventional power configurations toward higher-value power racks. The report raises its forecast for AI accelerator shipment CAGR from approximately 30% previously to approximately 45% and expects accelerator TDP to grow at an approximately 30% CAGR. TDP increases may be even faster after ASICs are used for training and inference. Power-rack adoption is expected to reach 20% during the Vera Rubin cycle and exceed 50% during the VR Ultra cycle. HVDC power racks and solid-state transformers are viewed as longer-term architectural drivers. Networking equipment investment previously lagged server investment, but the report believes a catch-up cycle is beginning. For the four largest U.S. cloud service providers, Accton's serviceable data center networking market—including general-purpose Ethernet switches and AI front-end and back-end networks—is expected to grow from approximately US$6 billion in 2025 to approximately US$33 billion in 2028, representing an approximately 75% CAGR. Drivers include general-purpose server capacity expansion, front-end and back-end network upgrades, and shorter bandwidth 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 additional upside. Over the past two to three years, cloud service providers prioritized investment in AI servers and data center construction, while networking investment lagged. Data show that spending on ODM-direct servers grew 207% and 127% in 2024 and 2025, respectively, while spending on ODM-direct switches grew only 18% and 103% during the same periods. Higher bandwidth and data volumes will drive upgrades to conventional front-end networks and create new GPU back-end networks. The TAM for GPU back-end Ethernet ToR switches is expected to grow at an approximately 164% CAGR from 2025 to 2028. Front-end specifications include a migration from BlueField 3 400G in GB200/300 to BlueField 4 800G in VR200, as well as upgrades in cloud service providers' proprietary NICs from 200G to 400G. The networking transition from 100G to 200G took approximately seven years, while the transition from 800G to 1.6T is expected to take less than two years, increasing the share of high-speed ports and the selling price per port. Accton's average selling price per switch port is expected to grow at an approximately 20% CAGR from 2025 to 2028. The PC business is at a different stage. Shipments in the first half of 2026 were better than expected, potentially due to Windows 10 replacement demand and pull-ins ahead of memory price increases. These factors brought forward some second-half demand. Together with negative price elasticity caused by rising prices for key components and the Windows 10 replacement cycle entering its later stages, the report expects PC shipments to decline 3% sequentially in the second half of 2026 and 8% for the full year, including declines of 14% for consumer PCs and 4% for commercial PCs. However, brand margins may perform better than shipment volumes. Asian PC brands achieved better-than-expected margins in the first half through price increases and product-mix adjustments, with PC average selling prices rising 20%–30% year over year, while also benefiting from lower-cost inventory on hand. ODM margins were diluted by higher selling prices, but absolute profit remained intact. High-end laptop bill-of-materials costs rose approximately 30% over the past year, with DRAM and NAND prices increasing three- to fivefold and adding US$200–250 to memory costs. Memory's share of total system bill-of-materials costs increased from the high-single-digit percentage range a year ago to 20%–30%. The report also assumes a 5% price increase for other components and non-core semiconductors. Brands therefore raised prices 20%–30% in the first half of 2026 and used specification reductions and product-mix adjustments to cushion costs. Low- and mid-range products are more sensitive to the rising share of memory costs because of their lower price bases. Tier-one brands are expected to demonstrate relatively greater margin resilience in the second half of 2026 due to their supply-chain capabilities.

Analysis framework

The report first validates aggregate demand using capital expenditure forecasts for U.S. hyperscale cloud service providers and tier-two data center operators, then confirms the commercial foundation for AI investment through an Asia-Pacific enterprise AI survey and trends in cloud revenue and margins. It subsequently dissects GPU and ASIC server shipments, the cadence of Nvidia platforms, rack bill-of-materials costs, and the ODM competitive landscape across successive layers of the supply chain, while cross-checking top-down CPU demand forecasts against bottom-up supply-chain intelligence. Finally, it translates server volumes, chip power consumption, and architectural changes into power and networking TAMs, while decomposing the volume and price effects on the PC business through shipments, selling prices, bill-of-materials costs, and margins.

Methodology notes

• Industry/Sector Analysis FrameworkUpstream, midstream, and downstream industry-chain transmission

  Industry-chain transmission of AI capital expenditure to servers, CPUs, power, cooling, and networking equipment

  Starting from cloud service provider capital expenditure and AI application demand, the report sequentially derives accelerator and server shipments and then analyzes their impact on the revenue and value content of Asian ODM, CPU, power, and switch suppliers.

• Industry/Sector Analysis FrameworkSupply-demand framework

  Demand forecasting and supply-chain constraint analysis

  The report treats capital expenditure, orders, and end-market adoption rates as demand-side evidence, and component shortages, production cycles, yields, and capacity as supply-side constraints, to assess whether shipment forecasts can be realized.

• Industry/Sector Analysis FrameworkVolume-price decomposition

  Decomposition of shipments, average selling prices, and margins

  Servers, power equipment, switches, and PCs are each decomposed into volumes, selling prices or value per system, and margin changes to explain why revenue growth and profit performance may diverge.

• Corporate Fundamentals and Financial FrameworkOperating/financial leverage analysis

  Analysis of ODM gross margins and gross profit per system

  The report notes that rising prices and complexity of next-generation racks will reduce ODM gross margins, but higher value per system and longer testing times can still increase gross profit per rack.

• (Method outside the vocabulary)

  Top-down forecasting cross-validated with bottom-up estimates

  Total server CPU demand is first estimated top-down based on applications and market size, and then triangulated against bottom-up estimates at the chip, server, and supply-chain levels.

• (Method outside the vocabulary)

  Bill-of-materials cost analysis

  The report decomposes the costs of GPUs, memory, networking, power, cooling, and other components in server racks and laptops to explain changes in system prices, brand price increases, and margins.

• (Method outside the vocabulary)

  TAM and SAM market-sizing analysis

  The report combines equipment shipments, power consumption, configuration adoption rates, and per-port selling prices to estimate the total or serviceable addressable markets for server CPUs, power equipment, and data center networking.

Asset mapping & comparison

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

• Delta（2308 TT）

  The report views it as a key beneficiary of the long-term AI data center power and cooling trend; the valuation table lists an OW rating, a current price of 1745.00, and a target price of 2500.00.

  Strengths

  AC/DC server power revenue is expected to grow at a 75%–80% CAGR from 2025 to 2028, broadly tracking the expansion of the server power TAM.

  Comparison

  The report expects its growth to broadly match the approximately 80% CAGR of the server power TAM.

  Risks

  Supply-chain constraints may affect end-server shipments and realization of power demand.

• ASPEED（5274 TT）

  The report believes supply releases in the second half of 2026 will drive revenue upside; the valuation table lists an OW rating, a current price of 15750, and a target price of 23000.

  Strengths

  Order visibility has extended into 2027.

  Weaknesses

  Near-term growth realization still depends on supply releases.

  Risks

  Failure to release supply on schedule may limit revenue upside.

• Lenovo（992 HK）

  The report believes the ISG margin inflection point has been confirmed and the AI project pipeline has doubled; the valuation table lists an OW rating, a current price of 29.72, and a target price of 50.00.

  Strengths

  Improving ISG margins, an expanding AI business pipeline, and strong supply-chain capabilities as a tier-one PC brand.

  Weaknesses

  The PC business faces weakening end demand in the second half of 2026.

  Comparison

  The report believes tier-one PC brands have relatively more resilient margins.

  Risks

  Pull-in orders, price elasticity resulting from rising costs, and the later stage of the Windows 10 replacement cycle may suppress PC sales.

• Accton（2345 TT）

  The report views it as a beneficiary of AI-driven network upgrades and the new server-rack market; the valuation table lists an OW rating, a current price of 2105.0, and a target price of 3700.0.

  Strengths

  It has a strong position in white-box data center switches, with the related business expected to account for approximately 40% of total revenue in the third quarter of 2026.

  Comparison

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

  Risks

  If networking capital expenditure continues to lag server investment, realization of market growth may be delayed.

• Wiwynn（6669 TT）

  The report believes it is well positioned across GPU, AI ASIC, and general-purpose servers; the valuation table lists an OW rating, a current price of 6325.0, and a target price of 7500.0.

  Strengths

  Its products cover GPU servers, ASIC servers, and general-purpose servers, and it is named as one of the leading suppliers in ASIC servers.

  Comparison

  Together with Celestica, it is named by the report as a leader in ASIC servers.

  Risks

  Supply-chain constraints and declining AI server ODM gross margins.

• Lotes（3533 TT）

  The report highlights its strong server demand and AI business optionality; the valuation table lists an OW rating, a current price of 1610.0, and a target price of 2500.0.

  Strengths

  Order visibility has extended into 2027.

  Weaknesses

  It faces near-term gross-margin pressure.

  Risks

  Near-term gross-margin headwinds and supply-chain constraints.

• Quanta（2382 TT）

  The report names it as one of the preferred NVL72 supply-chain plays; the valuation table lists an OW rating, a current price of 327.0, and a target price of 400.0.

  Strengths

  It has a competitive position in ODM manufacturing for complex NVL72 systems.

  Weaknesses

  The higher share of GPUs and memory in new platforms may dilute ODM gross margins.

  Comparison

  Together with Hon Hai, it is named as a preferred NVL72 supply-chain play.

  Risks

  Component shortages, production cycles, and lower margins on new platforms.

• Hon Hai（2317 TT）

  The report names it as one of the preferred NVL72 supply-chain plays; the valuation table lists an OW rating, a current price of 246.5, and a target price of 330.0.

  Strengths

  Complex design capabilities, scale, and working-capital requirements help major ODMs maintain a concentrated competitive landscape.

  Weaknesses

  Higher value content in next-generation racks does not mean gross margins will rise in tandem.

  Comparison

  Together with Quanta, it is named as a preferred NVL72 supply-chain play.

  Risks

  Supply shortages and compression of AI server ODM gross margins.

• Celestica

  The report names it as a leading supplier in AI ASIC servers.

  Strengths

  It benefits from demand for AWS Trainium and Google TPU, as well as more customers developing proprietary inference ASICs.

  Comparison

  Together with Wiwynn, it is named as a leader in ASIC servers.

  Risks

  ASIC demand and shipments remain affected by customer project schedules and supply-chain constraints.

• Asian PC brands and ODMs

  PC brands protect margins through price increases, lower-cost inventory, and product-mix adjustments, while ODMs face gross-margin dilution even as absolute profit remains intact.

  Strengths

  Tier-one brands have stronger supply-chain capabilities, while commercial PC demand is relatively more resilient.

  Weaknesses

  Demand in the second half of 2026 is affected by first-half pull-ins, price elasticity, and the later stage of the replacement cycle.

  Comparison

  Brand margins are performing better than ODM margins, while commercial PCs are expected to perform better than consumer PCs.

  Risks

  Rising prices for memory and other components have a greater impact on low- and mid-range products.

Key data

• AI chip shipment CAGR45%J.P. Morgan forecast for 2025–2028
• 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 report believes updated cloud service provider guidance provides 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 from 2026 to 2030US$5.5 trillionOf which US$4.1 trillion is expected to be debt-financed
• Asia-Pacific enterprise AI investment penetration88%Respondents incurred AI spending during the past 12 months; average investment intensity is expected to increase from 4.5% to 5.8%
• 2026 NVL72 rack shipments70,000–80,000 unitsIncluding 60,000–65,000 GB300 racks and approximately 10,000 Vera Rubin racks
• 2027 NVL72 rack shipments85,000–95,000 unitsExpected to consist mainly of VR200 racks
• 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%, 2.7%Below GB200's 4.2%
• Server CPU shipmentsFrom 26 million units in 2025 to 68 million units in 202838% CAGR
• Server CPU revenue TAM CAGR53%2025–2028 forecast, also assuming an approximately 10% CAGR in CPU average selling prices
• 2027 server shipment growth25%Expected to be driven by backlogs and strong server demand
• Data center power TAMUS$50 billion in 2028Expected CAGR of approximately 80% from 2025 to 2028
• Power-rack adoption rate20% during the Vera Rubin cycle, over 50% during the VR Ultra cycleChanges in power-supply architecture are one of three compound drivers of power-market growth
• Accton data center networking SAMFrom approximately US$6 billion in 2025 to approximately US$33 billion in 2028Expected CAGR of approximately 75%
• GPU back-end Ethernet ToR switch TAM CAGRApproximately 164%2025–2028 forecast
• Change in PC shipmentsFull-year 2026 -8%, second-half sequentially -3%Consumer and commercial PC shipments are expected to decline 14% and 4%, respectively, for the full year
• Increase in high-end laptop bill-of-materials costsApproximately 30%Increase over the past year; memory costs rose by US$200–250
• PC brand price increases in the first half of 202620%–30%To address rising memory and other component costs and protect margins

Impact & implications

The report believes that AI hardware growth has expanded from GPU servers to ASICs, general-purpose servers, CPUs, power, cooling, and networking equipment. The Asian hardware supply chain may therefore face a demand cycle lasting several years rather than a short-term ramp tied to a single product generation. The extent to which suppliers benefit depends on their complex-system design capabilities, financial resources, supply capabilities, and market shares in high-value power and high-speed networking segments. At the same time, revenue growth does not necessarily equate to margin improvement. The value content of next-generation racks rises significantly, but ODM gross margins are diluted by the increasing share of high-priced GPUs and memory. PC brands can protect margins through price increases and product-mix adjustments, while ODM margins remain under pressure. The report therefore regards supply constraints, product mix, and gross profit per system as more important profitability variables than shipment growth alone.

Risks

• Supply-chain and component constraints are the key swing factors determining whether server shipment forecasts can be realized.
• Production cycles and component shortages are expected to limit 2026 Vera Rubin NVL72 rack shipments to only approximately 10,000 units.
• Nvidia's Kyber architecture still faces challenges, creating execution uncertainty for the platform roadmap.
• Negative free cash flow among major hyperscale cloud service providers is difficult to avoid over the coming quarters and may increase ongoing financing pressure.
• The rising share of high-priced GPUs and memory in GB300 and VR200 is expected to reduce ODM gross margins to 3.6% and 2.7%.
• PC pull-ins during the first half of 2026, negative price elasticity caused by rising prices for key components, and the Windows 10 replacement cycle entering its later stages may cause second-half demand to be weaker than seasonality.
• Low- and mid-range PCs are more sensitive to the rising share of memory in bill-of-materials costs because of their lower selling prices.

What to watch

• Watch whether Vera Rubin systems can ramp on schedule from late in the third quarter through the fourth quarter of 2026.
• Track whether supply constraints ease and whether the supply releases in the second half of 2026 cited by ASPEED materialize.
• Monitor 2027 capital expenditure guidance from U.S. cloud service providers and whether the 50%–80% demand growth feedback received by memory manufacturers translates into orders.
• Track whether NVL72 shipments can reach 85,000–95,000 units in 2027 and transition smoothly to being predominantly VR200.
• Watch the demand stimulus from agentic AI and inference workloads for general-purpose servers, head-node CPUs, and standalone CPU servers.
• Track the power-rack adoption assumptions of 20% during the Vera Rubin cycle and more than 50% during the VR Ultra cycle.
• Watch the 800G-to-1.6T upgrade, construction of GPU back-end Ethernet networks, and increases in average selling prices per switch port.
• Track whether PC shipments decline 3% sequentially in the second half of 2026 and whether tier-one brands can preserve margin resilience through price increases and product-mix adjustments.