The AI upcycle is not over, and the Asian technology pullback offers an opportunity to reposition

Overview

The report judges that the AI-driven technology upcycle that began at the end of 2022 is not yet over. Although Asian technology stocks and the SOX index have experienced their third drawdown of more than 20% in this cycle, and investors are concerned that the hardware price-hike narrative has become overly crowded and that the sustainability of AI capex is declining, indicators such as AI model capabilities, inference demand, cloud revenue, capex plans, component inventories, and earnings estimate revisions do not show that the cycle is about to peak. The recent correction has instead brought valuations back to more reasonable levels, offering investors an opportunity to reposition toward beneficiaries of the next phase.

Core views

First, the market is pricing in earnings estimate cuts or hyperscaler capex reductions over the coming months, but the report expects earnings estimates to continue rising over the next few quarters and to broaden from core AI hardware into areas such as analog chips, second-tier foundries, wafers, and MLCCs. Second, AI scaling laws remain valid, and competition among frontier models, the development of open-source models, and agentic AI applications will jointly drive up training and inference compute demand. Third, cloud providers generally have healthy balance sheets; even if free cash flow comes under pressure in the second half of 2026 and in 2027, they are more likely to maintain infrastructure expansion through equity and debt financing. Fourth, semiconductor equipment and IC substrates have the best supply-demand dynamics and earnings leverage, while high-speed interconnects and advanced packaging may become the next bottlenecks. Fifth, memory fundamentals are solid, but customers reducing memory configurations to control costs may cap valuation upside.

Analysis framework

The report combines a top-down assessment of the technology cycle with bottom-up comparisons across the industry chain: it measures end-market compute demand through AI model scaling, token consumption, and public cloud revenue; evaluates investment sustainability through cloud provider capex, financing capacity, and backlog; judges cycle position through inventories, supply expansion, and bottleneck migration; and then screens subsectors and stocks based on the breadth of earnings estimate revisions, margins, returns, and historical valuation ranges.

Methodology notes

• Fundamental analysisAI scaling laws and demand validation

  Assess whether AI compute demand is still in an expansion phase through improvements in model capabilities, training compute investment, and inference token consumption.

  The report believes the empirical pattern that roughly 10 times more compute resources bring about roughly a 2 times improvement in intelligence still broadly holds, and that competition among frontier models, open-source models, and agentic AI continues to push up training and inference demand.

• Demand analysisPublic cloud revenue and backlog indicators

  Use public cloud revenue growth as an approximate observable indicator of AI compute consumption.

  Because many AI labs and application vendors use computing power through large cloud platforms, cloud revenue growth, quarterly incremental revenue, and order backlog can be used to validate compute demand and confidence in capex.

• Capital expenditure analysisAssessment of hyperscaler investment and financing capacity

  Assess the sustainability of AI infrastructure investment by combining capex forecasts, free cash flow, leverage levels, and equity and debt financing channels.

  The report expects major cloud service providers not to exit AI compute investment in 2027, as healthy balance sheets and external financing capacity can buffer the pressure from free cash flow turning negative.

• Cycle analysisInventory and supply-demand bottleneck monitoring

  Identify the peak of the technology cycle through inventories of key components, demand growth, and the speed at which supply catches up.

  Key AI components such as GPUs, ASICs, and memory have not yet shown abnormal inventory build-ups, and demand still exceeds the supply capacity that the supply chain and data center power budgets can support, so the typical characteristics of a cycle peak are not yet present.

• Valuation analysisEarnings estimate revisions and historical valuation comparison

  Assess risk-reward by combining the direction and breadth of earnings estimate revisions, margins, and historical P/E ranges.

  After the pullback in Asian technology stocks, valuations excluding memory are roughly one standard deviation above the average P/E over the past decade; the report believes that broader earnings upgrades and structural margin improvement can support valuations above those of past cycles.

Asset mapping & comparison

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

• Asian technology equities

  Assets that broadly benefit from AI compute demand, cloud capex, and the diffusion of earnings estimate upgrades

  Strengths

  Earnings estimates continue to be revised up, valuations are more reasonable after the pullback, and industry-chain margins and returns on capital have improved versus past cycles.

  Weaknesses

  Positioning is crowded and highly sensitive to capex, the financing environment, and expectations for price increases.

  Comparison

  Compared with the earnings downgrade cycle expected by the market, the report is more inclined to expect the breadth of earnings upgrades to expand over the coming quarters.

  Risks

  Cloud providers cut capex, model progress or application adoption slows, or supply of key components quickly catches up with demand.

• Semiconductor equipment

  The report believes it is the best subsector within Asian technology over the next 12 months

  Strengths

  Capex at TSMC and major memory makers may be revised up, while advanced packaging, cleanrooms, and localization create multiple demand drivers.

  Weaknesses

  Equipment delivery cycles are long, and revenue realization depends on customers actually executing capacity expansion plans.

  Comparison

  Compared with other hardware subsectors, wafer fabrication equipment has greater room for expectation upgrades and a broader range of beneficiaries.

  Risks

  Fabs delay capacity expansion, export controls, or investment in advanced nodes falls short of expectations.

• IC substrates

  The most fundamentally attractive area within components

  Strengths

  Benefits from larger AI accelerator packages, server CPUs, EMIB-T, and co-packaged optics demand, with concentrated supply and limited capacity expansion over the next two years.

  Weaknesses

  Demand is closely tied to advanced packaging roadmaps and high-end computing product cycles.

  Comparison

  Compared with memory, substrates have a more concentrated supply structure, margins are further below historical peaks, and there is more room for earnings upgrades.

  Risks

  Changes in packaging roadmaps, customer destocking, or new capacity exceeding expectations.

• Memory

  A cyclical asset with solid supply-demand dynamics but a weakening market narrative

  Strengths

  Supply is expected to remain below demand over the next 2-3 years, and after a deep correction there is room for a rebound over a six-month horizon.

  Weaknesses

  High prices are prompting customers to reduce HBM and DRAM configurations, weakening the narrative of price-inelastic demand.

  Comparison

  Compared with equipment and substrates, memory fundamentals are not poor, but valuation expansion is more easily constrained by concerns about demand destruction.

  Risks

  AI chips reduce memory density, end demand is suppressed by high prices, or supply recovers faster than expected.

• High-speed interconnects, optical connectivity, and advanced packaging

  Beneficiary areas that improve AI cluster efficiency and alleviate data transmission bottlenecks

  Strengths

  Low cluster utilization and a large share of instruction cycles used for data transmission are driving adoption of CPO, optical connectivity, networking equipment, and 3D packaging.

  Weaknesses

  Some technologies are still in the introduction phase, and the pace of commercialization and technology roadmaps remain uncertain.

  Comparison

  As the marginal efficiency of simply adding compute chips declines, the importance of interconnects and packaging will rise relatively.

  Risks

  Algorithm optimization replaces some hardware demand, technology mass production is delayed, or customer capex is reprioritized.

• Data center power and deployment infrastructure

  May replace chips as the main bottleneck for AI compute from the second half of 2027 to 2028

  Strengths

  Expansion in AI cluster scale will continue to increase demand for grids, behind-the-meter power, cooling, and data center deployment.

  Weaknesses

  Project approval and construction cycles are long, and demand realization is more delayed than chip investment.

  Comparison

  Chips will remain the main bottleneck for most of 2026 to 2027, but constraints may gradually shift toward power and deployment over the next 18-24 months.

  Risks

  Data center project delays, restricted grid access, policy approvals, and rising cost of capital.

Key data

• Recent correction in Asian technology stocks and the SOX index25%-30%The report believes current share prices have priced in a relatively high probability of short-term earnings downgrades or capex cuts.
• Forecast capex growth for major cloud service providers in 2027About 65%Expected to remain strongly expansionary after growth of more than 100% in 2026.
• Incremental public cloud revenue in the second quarter of 2026About US$15 billionClose to twice the incremental revenue in the first quarter of 2026, showing acceleration in cloud computing and AI token consumption.
• Net debt-to-equity ratio of major cloud providersAbout 12%As of the second quarter of 2026, overall balance sheets remain relatively healthy.
• Model FLOP utilization in AI clustersAbout 20%-40%Low utilization indicates substantial room for improvement in interconnects, data transmission, and cluster efficiency.
• Share of AI accelerator instruction cycles used for data transmissionAbout 50%-60%Supports high-speed interconnects, optical connectivity, and advanced packaging becoming key focuses in the next phase.
• Valuation level of Asian technology stocksAbout one standard deviation above the ten-year average P/EThe statistical scope excludes memory, and the report believes post-correction valuations are not excessive.
• Recent correction in memory stocksMore than 40%They may rebound significantly over the next six months, but the report does not expect them to return to their May 2026 highs in the near term.
• Outlook for the memory supply-demand gapSupply below demand over the next 2-3 yearsFundamentals are solid, but high prices may prompt customers to reduce memory configurations and limit valuations.
• Potential window for bottlenecks to migrate from chips to powerThe next 18-24 monthsBy the second half of 2027 and in 2028, data center deployment delays and power availability may replace chips as the main constraints.

Impact & implications

At the investment level, investors should shift from simply chasing hardware beneficiaries of price increases toward areas with sustainable earnings upgrades, concentrated supply, and characteristics of the next bottleneck. Semiconductor equipment benefits from upward revisions to wafer fabrication equipment spending, advanced packaging, and semiconductor localization in China; IC substrates benefit from larger AI accelerator package sizes, server CPUs, EMIB-T, and co-packaged optics demand; and high-speed interconnects, optical connectivity, and advanced packaging benefit from the urgent need to improve cluster utilization. Memory is more suitable for a tactical rebound trade than for sustained valuation expansion. In the medium term, investors also need to monitor the possibility that AI infrastructure bottlenecks shift from chips to data center deployment and power supply.

Risks

• Free cash flow at hyperscalers turns negative in the second half of 2026 and in 2027, and equity and debt financing may trigger valuation and market volatility.
• AI capex growth or actual deployment speed falls short of expectations, leading to earnings estimate cuts across the hardware supply chain.
• AI model capability improvements, inference token consumption, or commercial adoption of generative AI falls short of expectations.
• High prices cause AI chip customers to reduce configurations of components such as HBM and DRAM, creating demand destruction.
• New supply of key components is released quickly or inventories begin to accumulate, causing the cycle to peak early.
• Interconnects, advanced packaging, data center construction, and power supply form new bottlenecks, limiting the conversion of computing power into usable services.
• Export controls, localization competition, and regional regulatory changes affect the Asian semiconductor supply chain.
• The research institution has market-making, shareholding, client, or potential investment banking relationships with some of the companies discussed, so the independence of the views should be assessed prudently in conjunction with disclosures.

What to watch

• Software and internet companies' disclosures on adoption rates and monetization progress for generative AI, large language models, and agentic workflows.
• Whether AI applications spread from software and programming automation to industries such as finance and healthcare.
• Breakthroughs by frontier model labs in recursive self-improvement and model capabilities.
• Public cloud revenue growth, quarterly incremental revenue, and changes in large cloud provider backlog.
• 2027 capex plans and actual financing arrangements for major cloud providers, SpaceX, and AI labs.
• The magnitude of Asian technology earnings estimate upgrades and the breadth of their diffusion into analog chips, second-tier foundries, wafers, and MLCCs.
• Inventories, lead times, and new capacity for GPUs, ASICs, HBM, and other key components.
• Whether investments in wafer fabrication equipment, cleanrooms, and 2.5D and 3D advanced packaging become new bottlenecks in 2027-2028.
• AI cluster utilization and the adoption speed of CPO, optical connectivity, networking equipment, and advanced packaging.
• From the second half of 2027, whether data center deployment progress, grid access, and behind-the-meter power can match growth in chip supply.