AI capital expenditure continues to be revised upward, and the investment focus in Asian semiconductors is shifting from price increases to shipment volume and investment returns

Overview

This report summarizes investor feedback after Morgan Stanley’s three-day Greater China semiconductor roadshow in Hong Kong, and combines cloud service provider capital expenditure, AI investment returns, TSMC CoWoS capacity, and implied power consumption to analyze Asian semiconductor demand in 2027. Investors’ debate has shifted from “whether AI demand is real” to “whether incremental capital investment can generate sufficient returns,” while they are also paying closer attention to power supply, memory’s share of cloud budgets, and the transition from a price-increase logic to a shipment-volume logic.

Core views

The report believes AI infrastructure investment remains in a strong expansion cycle, and cloud service providers’ feedback on the commercialization of new capacity is positive; therefore, it is not yet concerned about overproduction of AI GPUs and ASICs in 2027. Chip output corresponding to TSMC CoWoS is expected to grow 70% to 80% YoY, making wafer foundry and advanced packaging the core shipment-volume beneficiaries. Memory remains the center of market attention, but its share of cloud capital expenditure is too high and significantly affected by pricing; if memory prices decline, budgets may be reallocated to GPUs, ASICs, and server CPUs, thereby increasing the amount of compute that can be purchased with the same capital expenditure.

Analysis framework

The research combines feedback from Hong Kong investor roadshows, tracking of global cloud service provider capital expenditure, management commentary on investment returns, breakdowns of cloud budget structure, and mapping of CoWoS wafers to chip quantities and power consumption. It also uses a SpaceX compute-expansion scenario to test advanced packaging demand and power constraints.

Methodology notes

• Demand analysisCloud capex tracking

  Assess AI hardware demand through capital expenditure budgets and revision magnitudes of major global cloud service providers.

  The report tracks 14 listed cloud service providers and combines post-quarterly earnings guidance changes to evaluate 2027 capex growth and room for further upward revisions.

• Investment return analysisAI capex ROIC framework

  Use AI revenue, capacity utilization, and management commentary to test whether capital expenditure can generate reasonable returns.

  The report views recurring revenue from large-model companies, the pace of cloud business commercialization, and absorption of new capacity as key indicators of whether future hardware demand can be sustained.

• Industry-chain mappingCoWoS capacity and power-consumption mapping

  Convert CoWoS wafer capacity into the number of GPUs or ASICs, then estimate incremental power demand based on per-chip power consumption.

  This method is used to identify the scale of advanced packaging demand and power supply bottlenecks, but the results are sensitive to assumptions about average chip power consumption, product mix, and capacity utilization.

• Scenario analysisReallocation of cloud capex structure

  Analyze the crowding-in or crowding-out effects of memory price changes on budgets for compute, networking, and server CPUs.

  If memory prices normalize, the same total capital expenditure can purchase more GPUs, ASICs, and CPUs, thereby improving the compute purchasing power of cloud capex.

• Market behavior analysisInvestor positioning and style divergence

  Identify differences in focus metrics between momentum investors and value investors based on roadshow feedback.

  Momentum investors focus more on YoY changes in DRAM prices, the breadth of earnings forecast revisions, and spot trends, while value investors emphasize valuation, buybacks, and potential cash returns.

Asset mapping & comparison

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

• TSMC

  Core Overweight, benefiting from growth in AI accelerator and CoWoS shipments

  Strengths

  Advanced process technology and CoWoS capacity occupy critical positions in the supply chain, with AI chip output expected to grow significantly in 2027.

  Weaknesses

  Capacity expansion needs to be matched simultaneously with customers’ deployment capabilities and power infrastructure.

  Comparison

  Compared with price-driven memory, TSMC is closer to shipment-volume-driven wafer and advanced packaging demand.

  Risks

  Cloud capex returns falling short of expectations, power shortages, customer project delays, and AI chip demand below forecasts.

• ASE Technology Holding Co. Ltd.

  Overweight, a beneficiary of advanced packaging and server chip ramp-up

  Strengths

  It has exposure to both packaging operations and server CPU-related areas, benefiting from higher AI infrastructure shipments.

  Weaknesses

  Demand is closely linked to the product cycles and packaging orders of major chip customers.

  Comparison

  Compared with pure wafer foundries, ASE provides complementary exposure to the packaging segment.

  Risks

  Advanced packaging competition, utilization fluctuations, and delays in AI hardware orders.

• ASMPT Ltd

  Key Overweight, benefiting from demand for semiconductor packaging equipment

  Strengths

  AI chip and advanced packaging capacity expansion can drive equipment demand.

  Weaknesses

  Equipment orders are highly cyclical, and changes in capital expenditure may amplify earnings volatility.

  Comparison

  Compared with chip manufacturers, ASMPT is more sensitive to industry capital expenditure and packaging capacity expansion.

  Risks

  Customer capex cuts, order deferrals, and industry cycle reversal.

• MediaTek

  Overweight and listed among top picks

  Strengths

  If ASIC solutions are successfully executed, they can help customers reduce AI infrastructure budgets and improve efficiency.

  Weaknesses

  Value realization depends on customer project implementation, design execution, and mass-production progress.

  Comparison

  Compared with companies that rely solely on industry-wide price increases, MediaTek has technology value-add and business-model drivers.

  Risks

  ASIC project execution falling short of expectations, customer concentration, and intensifying competition.

• Aspeed Technology, Hygon, and Montage Technology

  Key Overweight names, as a beneficiary basket of server BMCs, CPUs, and related AI infrastructure

  Strengths

  Agentic AI demand may increase demand for server CPUs, management chips, and platform supporting products.

  Weaknesses

  The companies differ in product cycles, customer structures, and technological barriers.

  Comparison

  These represent shipment-volume opportunities spreading from the GPU main line to server CPUs and supporting chips.

  Risks

  Agentic AI penetration slower than expected, server deployment delays, and technology substitution.

• Nanya Technology and Winbond Electronics Corp

  Preferred names within memory subsegments

  Strengths

  DDR4 DRAM supply-demand dynamics and potential cash returns can provide cyclical and valuation support.

  Weaknesses

  They remain highly exposed to memory prices, spot trends, and earnings forecast revisions.

  Comparison

  The report relatively prefers DDR4 DRAM over NAND module makers.

  Risks

  Slowing price increases, declining contract prices, specification adjustments, and supply recovery.

• Phison Electronics Corp and Shenzhen Longsys Electronics Co Ltd

  Relatively cautious, rated Equal-weight

  Strengths

  They can participate in growth in NAND and memory demand.

  Weaknesses

  Compared with DDR4 DRAM names, NAND module pricing and competitive pressure warrant more caution.

  Comparison

  The report prefers Nanya Technology and Winbond over the above NAND module companies.

  Risks

  Chinese consumer-grade NAND module prices may slow in the fourth quarter of 2026 and could affect 2027 contract prices.

• GlobalWafers Co Ltd and Silergy Corp.

  Recommended to avoid or Underweight

  Strengths

  If the industry pricing environment improves significantly, earnings may have cyclical flexibility.

  Weaknesses

  Low-end commoditized businesses have weaker pricing power, making it difficult to fully pass through costs or capture price benefits.

  Comparison

  Compared with companies with advanced technology, shipment growth, or efficiency-improvement logic, their competitive advantages are weaker.

  Risks

  Failure to raise prices, intensified competition in low-end markets, and margin pressure.

Key data

• 2027 global cloud capital expenditureApproximately US$1.5 trillionScenario assumption including sovereign AI investment; estimates for 14 listed global cloud service providers excluding sovereign AI are close to US$1.3 trillion.
• 2027 cloud capex growth29% YoY growthRaised by 15 percentage points from the forecast of 14% about one month ago.
• 2027 capex forecast revisionRaised by over 70% year-to-date, increasing by about US$500 billionThe report believes forecasts may continue to be revised upward as long as qualitative guidance on cloud spending remains strong.
• Memory share of cloud capexApproximately 53%Based on a total budget scenario of US$1.5 trillion; high memory prices and supply shortages have pushed up the budget share.
• Compute chip budget shareApproximately 20%Covers GPUs, ASICs, and CPUs; the report also estimates that GPUs or ASICs excluding HBM account for about 15% to 20%.
• Implied CoWoS output in 2027Approximately 2 million wafers and 19 million AI GPUs or ASICsCorresponding chip output is expected to grow 70% to 80% YoY.
• Implied power demand from AI chipsApproximately 38GWA rough estimate based on 19 million GPUs or ASICs and an average thermal design power of 2kW per chip.
• SpaceX incremental compute planApproximately 8GWIf entirely using Vera Rubin, it could correspond to about 4 million GPUs and about 400,000 CoWoS wafers, accounting for about 20% of TSMC’s 2027 GPU or ASIC output.
• AWS AI revenue run rateOver US$25 billionYoY growth reached triple digits, supporting the view that AI capital expenditure is being underpinned by visible commercialization demand.

Impact & implications

Continued upward revisions to capital expenditure are positive for wafer foundry, advanced packaging, ASICs, server CPUs, BMCs, memory, networking, and power infrastructure supply chains. If investor focus shifts from price increases to shipment volume, volume-sensitive companies such as TSMC and ASE, as well as chip design companies that can reduce AI infrastructure costs or improve efficiency, may gain relative advantages. Memory may still benefit from tight supply and HBM demand, but its high budget share also means there is risk of capital and order-structure reallocation after prices normalize.

Risks

• The return on investment of AI capital expenditure fails to keep pace with the scale of investment, causing cloud service providers to cut subsequent budgets.
• Global power supply, data center grid connection, cooling, and network infrastructure become bottlenecks for AI chip deployment.
• Memory prices normalize from high levels, leading to downward revisions to revenue and earnings forecasts.
• HBM, DRAM, and consumer electronics memory specification adjustments weaken memory demand per device.
• Large projects such as SpaceX fail to materialize as planned, causing expectations for CoWoS and GPU demand to fall short.
• Delays in execution of advanced packaging, ASIC, or server CPU projects affect shipment realization.
• The research institution and some covered companies have shareholding, investment banking, or other commercial relationships; investors should make cautious judgments in conjunction with conflict-of-interest disclosures.

What to watch

• Whether the quarterly or monthly recurring revenue of large-model companies can cover the continuously increasing costs of capital expenditure.
• The latest guidance from Meta, Alphabet, Amazon, and Microsoft on AI investment returns, capacity utilization, and capital expenditure.
• Whether consensus expectations for 2027 cloud capital expenditure continue to be revised upward, and whether 29% YoY growth can be maintained.
• Progress in realizing TSMC CoWoS capacity, wafer input, and the assumption of about 19 million AI chips in output.
• Progress in power supply, grid connection, and data center construction corresponding to the incremental approximately 38GW of power demand.
• The construction pace of SpaceX’s expansion from 2GW to about 10GW and its Vera Rubin procurement commitments.
• DRAM spot and contract prices, YoY price increases, and the breadth of earnings forecast revisions.
• Whether memory’s budget share in cloud capex is reallocated toward GPUs, ASICs, and server CPUs.
• Whether Rubin Ultra’s HBM specifications and changes in DRAM capacity for new iPhones are ultimately due to shortages or customer cost reductions.