AI capex continues to be revised upward, while the focus in Asian semiconductors shifts from price increases to shipment volumes and investment returns

Overview

The report summarizes investor feedback gathered by Morgan Stanley after a three-day Greater China semiconductor roadshow in Hong Kong, and analyzes global cloud service provider capex, AI investment returns, TSMC CoWoS capacity, chip power consumption, and memory budget structure. Investor doubts about AI demand itself have declined, while the discussion focus is gradually shifting to whether capex can generate sufficient returns, whether power can support chip deployment, and how future budgets should be allocated between memory and compute chips.

Core views

The report believes global cloud service providers are still increasing investment in AI infrastructure, and new capacity is being rapidly absorbed by demand, so there is no obvious risk of overproduction in AI GPUs and ASICs in 2027 for now. If semiconductor price-increase momentum slows, market attention may shift from price elasticity to shipment volumes, business models, and infrastructure efficiency; therefore, TSMC foundry, ASE advanced packaging, and ASIC-related names such as MediaTek are more attractive. Memory remains the sector most closely watched by Hong Kong investors, but momentum investors focus more on price trends and upward earnings revisions, while value investors pay more attention to valuation, buybacks, and cash returns. If memory prices fall back, cloud service providers may shift more budgets to GPUs, ASICs, and server CPUs, which would instead increase overall AI computing power purchasing capacity.

Analysis framework

The report combines Hong Kong investor interviews with top-down cloud capex forecasts, bottom-up derivations of CoWoS wafer and chip volumes, single-chip power consumption conversions, cloud service provider management commentary, and scenario analysis of memory and compute chip budgets, and tests capacity and power constraints through the case of SpaceX adding 8GW of compute capacity.

Methodology notes

• Investor surveySummary of Hong Kong roadshow feedback

  Investor attention and positioning preferences

  Identifies major investor divergences on memory, logic semiconductors, AI capex returns, and power constraints through a three-day Hong Kong roadshow, and observes a migration in focus from the price-increase thesis to the shipment-volume thesis.

• Fundamental analysisAI capex–ROIC framework

  Whether revenue growth can cover capital investment

  Assesses the sustainability of AI hardware demand and the risk of overinvestment by combining cloud service provider capex guidance, management commentary on AI commercialization, and recurring revenue growth at large model companies.

• Supply-demand analysisCoWoS capacity implied chip volume model

  Mapping advanced packaging capacity to AI chip output

  Derives about 19 million AI GPUs or ASICs in 2027 from about 2 million CoWoS wafers, and uses this to evaluate demand for foundry, advanced packaging, and related supply chains.

• Constraint analysisChip volume–power consumption conversion

  Power feasibility of computing power deployment

  Converts about 19 million chips into about 38GW of power demand based on average power consumption of 2 kilowatts per GPU or ASIC, to assess whether power supply will limit actual AI hardware deployment.

• Scenario analysisReallocation of cloud capex structure

  Impact of memory price changes on computing power procurement capacity

  Analyzes the possibility that capex shifts from memory to GPUs, ASICs, and server CPUs after memory prices normalize, and scenarios in which the same budget can support more compute capacity.

Asset mapping & comparison

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

• TSMC

  Core beneficiary of advanced nodes and CoWoS expansion

  Strengths

  Has leading wafer manufacturing and advanced packaging capabilities, and can directly benefit from growth in AI GPU, ASIC, and server CPU shipments.

  Weaknesses

  Demand realization depends on customer capex, data center construction, and power supply, while massive expansion also increases cyclical volatility risk.

  Comparison

  Compared with memory companies that rely on price increases, TSMC is more oriented toward a shipment-volume thesis driven by chip quantities and advanced-node penetration.

  Risks

  AI investment returns below expectations, CoWoS order adjustments, power bottlenecks, or changes in customer in-house chip development progress.

• ASE Technology Holding

  Beneficiary of advanced packaging, ABF substrates, and the server processor supply chain

  Strengths

  Can gain multiple drivers from growth in packaging demand for AI accelerators, ASICs, and server CPUs.

  Weaknesses

  The packaging business is relatively sensitive to customer product schedules and capacity utilization.

  Comparison

  Compared with pure-play foundries, ASE is more sensitive to rising packaging complexity and the trend toward heterogeneous integration.

  Risks

  Customer order deferrals, packaging capacity oversupply, changes in technology routes, and price competition.

• MediaTek

  Key name in the ASIC and custom compute chip theme

  Strengths

  If custom ASICs are successfully executed, they can help cloud service providers improve budget efficiency and benefit from the expansion of AI compute demand.

  Weaknesses

  Project revenue has relatively high customer concentration and execution uncertainty.

  Comparison

  Compared with general-purpose GPU suppliers, its investment thesis relies more on the cost efficiency of custom chips and customer adoption.

  Risks

  Tape-out or mass production delays, customer project cuts, intensified competition, and margins below expectations.

• Aspeed, Hygon, and Montage

  Beneficiaries in server management chips, China CPUs, and AI infrastructure supporting components

  Strengths

  Can benefit from growth in server volumes, agentic AI demand, and expansion of local computing platforms.

  Weaknesses

  Segment product market size, customer structure, and regional policy exposure are relatively concentrated.

  Comparison

  Compared with pure AI accelerator names, these companies provide infrastructure support or alternative computing capabilities, with more diversified demand sources.

  Risks

  Server deployment below expectations, technology iterations, supply chain constraints, and local competition.

• Nanya Technology and Winbond

  Preferred names in DDR4 DRAM and niche memory

  Strengths

  The report believes they are more attractive for allocation than NAND module makers, and may benefit from tight supply-demand, valuation support, and cash returns.

  Weaknesses

  Earnings remain highly dependent on memory prices, inventory cycles, and changes in product specifications.

  Comparison

  The report prefers the two over NAND module makers such as Phison and Longsys.

  Risks

  Slower DRAM price increases, customer specification downgrades, supply recovery, and earnings forecast downgrades.

• GlobalWafers and Silergy

  Low-end commoditized areas where the report recommends caution or avoidance

  Strengths

  If industry prices rise broadly, they may still gain some cyclical benefits.

  Weaknesses

  Low-end products have weaker pricing power and may not be able to smoothly pass costs on to customers or realize price increases.

  Comparison

  Compared with companies such as TSMC, ASE, and MediaTek that have shipment-volume or technology value-add theses, the risk-reward is weaker.

  Risks

  Failure to raise prices, capacity oversupply, product commoditization, and insufficient demand recovery.

Key data

• 2027 global cloud capex scenarioAbout USD1.5 trillionIncludes sovereign AI investment; when only the top 14 listed global cloud service providers are counted, the forecast is close to USD1.3 trillion.
• 2027 cloud capex YoY growth29%Revised up by 15 percentage points from the forecast of 14% about one month earlier.
• Year-to-date upward revision to 2027 capex forecastMore than 70%, about USD500 billionThe report believes forecasts may continue to be revised upward as long as qualitative commentary related to cloud spending remains strong.
• Memory share of 2027 cloud capexAbout 53%The high share reflects tight supply and elevated prices for high-bandwidth memory such as HBM.
• Compute chip share of 2027 cloud capexAbout 20%Includes GPUs, ASICs, and CPUs; another calculation shows GPUs or ASICs excluding HBM account for about 15% to 20%.
• 2027 CoWoS implied outputAbout 2 million wafers and 19 million AI GPUs or ASICsMorgan Stanley expects related chip output to grow 70% to 80% YoY.
• 2027 GPU or ASIC implied power consumptionAbout 38GWEstimated based on average thermal design power of 2 kilowatts per chip.
• SpaceX incremental compute capacity scenario8GWIf all uses Vera Rubin, it could imply about 4 million Rubin GPUs, about 400,000 CoWoS wafers, and about 20% of TSMC's 2027 AI accelerator output.
• AWS AI revenue run rateMore than USD25 billionAmazon said its YoY growth rate reached triple digits.

Impact & implications

Continued upward revisions to AI capex and stronger commercialization signals provide medium-term demand support for advanced nodes, CoWoS packaging, ASICs, server management chips, and high-bandwidth memory. If memory prices fall back, it may weaken memory price elasticity in the short term, but could release budgets for more GPUs, ASICs, and CPUs, thereby benefiting logic and computing semiconductors driven by shipment volumes. Meanwhile, the implied power demand of about 38GW indicates that chip capacity is not the only constraint; grid, power supply, cooling, and data center construction progress will determine the speed at which orders ultimately translate into actual deployment.

Risks

• Insufficient growth in AI revenue and recurring revenue at large model companies, causing capex ROIC to fall below expectations.
• Delays in power, cooling, grid connection, and data center construction, preventing chip capacity from being converted into actual deployment in a timely manner.
• Normalization of memory prices and customers lowering DRAM or HBM specifications, weakening memory companies' earnings elasticity.
• Cloud service providers reallocating budgets among memory, GPUs, ASICs, CPUs, and infrastructure, causing supply chain demand mismatches.
• Construction targets for large projects such as SpaceX are highly scenario-based assumptions, and actual progress and equipment configuration may fall short of plans.
• Morgan Stanley has shareholdings, investment banking business, or other commercial relationships with multiple covered companies, which may constitute potential conflicts of interest.

What to watch

• Whether cloud service provider quarterly capex guidance and 2027 consensus expectations continue to be revised upward.
• Latest commentary from Meta, Google, Amazon, and Microsoft on AI commercialization, capacity utilization, and ROIC.
• Whether quarterly or monthly growth in recurring revenue at large model companies can cover continuously increasing capex.
• TSMC CoWoS capacity, wafer demand, and shipment realization for about 19 million AI GPUs or ASICs.
• Whether power supply, data center grid connection, cooling, and network construction can match the implied demand of about 38GW.
• HBM, DRAM, and NAND spot and contract prices, breadth of earnings forecast upgrades, and memory company buyback plans.
• SpaceX's construction progress from 2GW to about 10GW, as well as actual Vera Rubin procurement and CoWoS consumption.
• Whether memory's share of cloud capex declines, and whether the released budget shifts toward GPUs, ASICs, and server CPUs.