AI cycle fundamentals remain strong, and the correction in Asian technology stocks offers an opportunity to position for the next leg

Overview

The report argues that the recent correction in Asian technology stocks mainly reflects investor concerns about earnings downgrades, demand destruction caused by high AI chip prices, and the sustainability of hyperscaler capital expenditure, rather than fundamentals having already peaked. AI model capabilities are still evolving rapidly, inference demand and public cloud revenue continue to grow, and unit economics across the AI ecosystem are improving. At the same time, inventories of key components have not accumulated abnormally, and demand remains above the supply capacity that the supply chain and data center power budgets can support. Therefore, the report concludes that the AI upcycle is not over, and that the coming quarters are more likely to see broader earnings estimate upgrades and further capex increases.

Core views

First, AI scaling laws remain valid, with multiple frontier model labs continuing to compete and drive improvements in model capability; open-source models intensify competition but are instead favorable for technology hardware and AI compute investment. Second, AI inference token consumption and public cloud revenue growth are accelerating, and expanding cloud provider backlogs provide a demand foundation for high levels of capital expenditure. Third, AI components have not shown the inventory buildup, sudden demand growth slowdown, or rapid supply catch-up commonly seen at cyclical peaks. Fourth, earnings upgrades are spreading from leading segments to areas such as analog chips, second-tier foundries, wafers, and MLCCs. Fifth, the report is most positive on semiconductor production equipment and IC substrates over the next 12 months; memory fundamentals are solid, but the valuation narrative is affected by content reductions caused by high prices; the medium-term bottleneck may shift from chips to interconnects, and then further to data center deployment and power.

Analysis framework

The report cross-validates AI model capability and token demand, public cloud revenue and backlog, hyperscaler capex forecasts, supply chain inventories, the breadth of earnings estimate revisions, historical valuation ranges, and key infrastructure bottlenecks, and uses this to rank Asian technology sub-industries and select key companies.

Methodology notes

• Technological progress analysisAI scaling laws

  An approximately 10-fold increase in training compute input typically delivers about a 2-fold improvement in intelligence level.

  The report uses whether model capabilities can continue to scale with compute as the core criterion for judging the continuation of the AI cycle, and believes that competition among multiple frontier labs and exploration of recursive self-improvement have not yet shown a slowdown in model evolution.

• Demand validationPublic cloud revenue proxy indicator

  Use revenue growth at large cloud platforms as an approximate measure of AI compute consumption intensity.

  Because a large amount of compute from AI labs and other vendors is consumed through hyperscale cloud platforms, accelerating public cloud revenue and expanding order backlogs can be used to validate the sustainability of inference demand and capital expenditure.

• Cycle analysisTechnology cycle peak signal check

  Check inventories of bottleneck components, demand growth, and the speed at which supply catches up.

  The report has not observed abnormal inventory accumulation in key AI components such as GPUs, ASICs, or memory, nor has it seen the typical cyclical peak signals of a clear demand slowdown or supply rapidly catching up.

• Earnings and valuation analysisEarnings estimate revisions and historical valuation comparison

  Assess risk-reward by combining the direction of earnings estimates, the breadth of upgrades, and valuation levels relative to long-term averages.

  Earnings estimate upgrades for Asian technology are spreading to more industry segments; excluding memory, valuations are about one standard deviation above the 10-year average P/E ratio, which the report does not consider excessive.

• Industry chain analysisInfrastructure bottleneck migration

  Identify changes in constraints formed by chips, interconnects, deployment progress, and power during the expansion of compute capacity.

  The report expects chips to remain the main bottleneck for most of 2026 to 2027, after which the importance of interconnect efficiency will rise; over the next 18 to 24 months, data center construction delays and power availability may replace chips as the main constraints.

Asset mapping & comparison

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

• Asian technology stock portfolio

  Core long allocation in the AI upcycle

  Strengths

  AI demand growth, continued earnings estimate upgrades, valuations more reasonable than before the correction, and industry chain profitability and returns on capital above prior cycles.

  Weaknesses

  Positioning is relatively crowded and sensitive to changes in the narratives around capital expenditure, earnings estimates, and supply bottlenecks.

  Comparison

  Compared with recent price performance, the report believes fundamentals are stronger and that the market has over-discounted near-term earnings downgrade risk.

  Risks

  Cloud providers cutting capital expenditure, slowing AI adoption, component supply rapidly catching up, or data center construction being hindered.

• Semiconductor production equipment

  The most favored Asian technology sub-industry over the next 12 months

  Strengths

  Expectations for wafer fab equipment spending may be raised, and leading manufacturers, challengers, and Chinese domestic semiconductor companies all have incentives to expand capacity and invest in localization.

  Weaknesses

  Equipment orders are cyclical and affected by wafer fab construction progress, export restrictions, and customer capital discipline.

  Comparison

  Compared with other Asian technology sub-industries, the visibility of capex upgrades and the scope of beneficiaries are stronger.

  Risks

  Delays in wafer fab capacity expansion, policy restrictions, bottlenecks in equipment and cleanroom capacity, and capital expenditure falling short of expectations.

• IC substrates

  The most fundamentally attractive direction within the components sector

  Strengths

  Benefits from the expansion of AI accelerator package area, server CPU demand, EMIB-T package penetration, and incremental CPO-related demand, with concentrated supply and limited new capacity over the next two years.

  Weaknesses

  Demand is relatively dependent on high-end computing and advanced packaging expansion, and margin recovery still needs time.

  Comparison

  Compared with other components, the supply-demand balance is tighter, margins are still below historical peaks, and there is greater room for earnings upgrades.

  Risks

  Advanced packaging penetration slower than expected, customer capacity expansion adjustments, or new supply exceeding expectations.

• Memory chips

  A trading opportunity with solid fundamentals but a damaged valuation narrative

  Strengths

  Supply is still meaningfully below demand over the next 2 to 3 years, and after a correction of more than 40%, there is rebound potential over the next six months.

  Weaknesses

  High prices are prompting some AI accelerators and CPUs to reduce HBM or DRAM configurations, weakening the market assumption of price-inelastic demand.

  Comparison

  Supply-demand fundamentals are not weak, but compared with semiconductor equipment and IC substrates, valuation upside is more vulnerable to concerns about demand destruction.

  Risks

  Customers further reducing memory content, price increases suppressing demand, supply expansion, or stocks being unable to return to previous highs.

• High-speed interconnects, optics, and advanced packaging

  Next-stage beneficiary directions from AI cluster efficiency improvements

  Strengths

  Low cluster utilization and a large share of instruction cycles used for data transmission drive the adoption of CPO, optical connectivity, networking equipment, 3D SoIC, and on-package high-bandwidth memory solutions.

  Weaknesses

  Some technologies are still in the introduction phase, with uncertainty around commercialization timing and technology roadmaps.

  Comparison

  As pure compute expansion is constrained by power and chip supply, interconnect optimization may have higher marginal efficiency than continuing to stack chips.

  Risks

  Delayed technology adoption, changes in standard roadmaps, capital expenditure shifts, or software algorithm improvements in cluster efficiency reducing incremental hardware demand.

• Data center power and deployment infrastructure

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

  Strengths

  AI compute expansion creates long-term demand for grid-side and behind-the-meter power, cooling, power distribution, and data center construction.

  Weaknesses

  Project cycles are long and constrained by approvals, grid connection, equipment delivery, and construction progress.

  Comparison

  Chips remain the main near-term constraint, but over the next 18 to 24 months, the importance of power and data center delivery capacity may exceed semiconductor supply.

  Risks

  Construction delays may in turn depress the pace of AI chip deployment and cloud revenue realization.

Key data

• Number of drawdowns in this cycleThe third drawdown of more than 20%Refers to the significant corrections in Asian technology stocks and the SOX index during the AI upcycle that began in late 2022.
• Recent market correctionAbout 25% to 30%The report believes this magnitude already reflects a relatively high probability of near-term earnings downgrades or capex cut expectations.
• Second-quarter incremental public cloud revenueAbout US$15 billionNearly doubled from the first quarter of 2026, indicating that AI compute consumption is still accelerating.
• 2027 hyperscaler capital expenditure growth65%J.P.Morgan aggregate forecast; the growth rate exceeds 100% in 2026.
• Hyperscaler net debt-to-equity ratioAbout 12%As of the second quarter of 2026, the report uses this to judge that their balance sheets remain broadly healthy.
• AI cluster model floating-point utilizationTypically 20% to 40%Low utilization implies significant room for improvement in interconnects, data transmission, and cluster efficiency.
• Asian technology valuationAbout one standard deviation above the 10-year average P/E ratioExcludes the memory sector; the report believes post-correction valuations are not excessive.
• Duration of the memory supply-demand gapThe next 2 to 3 yearsFundamentals remain tight, but high prices may push customers to reduce memory configurations.
• Recent correction in memory stocksMore than 40%The report expects a potentially significant rebound over the next six months, but does not yet believe the stocks can return to their May 2026 highs.
• Potential bottleneck switching windowThe next 18 to 24 monthsChip constraints may gradually give way to constraints from data center deployment progress and power availability.

Impact & implications

The current correction provides an opportunity to reallocate to Asian technology stocks, but returns in the next stage may shift away from a hardware narrative based solely on price increases and scarcity toward segments driven by capex upgrades, concentrated supply, advanced packaging expansion, or efficiency improvements. Semiconductor production equipment, IC substrates, advanced packaging and testing, and interconnect beneficiaries have clearer upside. Although memory supply and demand are tight, configuration reductions caused by high prices make its valuation elasticity weaker than other beneficiary sectors. Over the medium to long term, investors also need to assess in advance the constraints that power and data center delivery capacity may impose on the pace at which chip demand is realized.

Risks

• Hyperscaler free cash flow may turn negative in the second half of 2026 and in 2027, and the need for debt and equity financing may increase market volatility.
• If AI capital expenditure or inference demand falls short of expectations, earnings estimate upgrades and supply chain orders will weaken.
• High-priced AI chips and memory may lead customers to reduce configurations, delay procurement, or adopt lower-cost alternatives.
• Rapid expansion of semiconductor supply may cause chip shortages to ease earlier than expected, pressuring prices and margins.
• Data center construction delays, grid access, and insufficient power supply may limit compute deployment after the second half of 2027.
• Improvements in model capabilities, recursive self-improvement, or the adoption speed of generative AI in industries such as finance and healthcare may be lower than expected.
• Positioning in Asian technology stocks is crowded, and changes in the macro environment, financing costs, or market risk appetite may amplify valuation volatility.
• Semiconductor equipment and China’s localized supply chain may be affected by changes in policy, trade, and export controls.

What to watch

• The adoption rate and monetization progress of generative AI, LLMs, and agentic workflows disclosed by software and internet companies.
• Whether generative AI expands beyond software and programming automation into industries such as finance and healthcare.
• Progress by frontier model labs in model capabilities, inference efficiency, and recursive self-improvement.
• 2027 hyperscaler capital expenditure guidance, financing arrangements, and changes in free cash flow.
• Whether public cloud revenue growth, incremental revenue, and order backlogs continue to expand.
• The magnitude of Asian technology earnings estimate upgrades and the breadth of their spread to analog chips, second-tier foundries, wafers, and MLCCs.
• Whether inventories of GPUs, ASICs, HBM, and other key AI components begin to accumulate abnormally.
• Capital expenditure and equipment order upgrades from TSMC, major memory manufacturers, and Chinese semiconductor companies.
• Whether HBM and DRAM configurations in AI accelerators and CPUs continue to be lowered due to rising costs.
• Changes in AI cluster utilization, CPO and advanced packaging penetration, and bottlenecks in power and data center delivery.