AI upcycle fundamentals have not weakened, and the pullback in Asian technology stocks offers an opportunity to position

Overview

The report argues that the AI-driven technology upcycle that began in late 2022 is not yet over. Although Asian technology stocks and the SOX index have undergone a 25% to 30% correction, with investors worried about earnings downgrades, demand destruction caused by component price increases, and the sustainability of hyperscaler capex, AI model capabilities are still evolving rapidly, inference token consumption and public cloud revenue continue to accelerate, and unit economics across the AI ecosystem have also improved. The report recommends buying Asian technology stocks at current levels and shifting allocation focus toward the next-stage beneficiaries, including semiconductor manufacturing equipment, IC substrates, advanced packaging, and interconnect.

Core views

First, AI Scaling Laws remain valid; competition among frontier models and the development of open-source models will increase training and inference compute demand rather than weaken hardware investment. Second, public cloud revenue and backlog continue to grow, giving cloud service providers confidence to maintain high capex; even if free cash flow turns negative, leaders can still use equity and debt markets for financing. Third, GPUs, ASICs, and memory have not yet shown the inventory buildup commonly seen at cyclical peaks, and AI chip demand over the next 12 months is still expected to exceed the supply that supply chains and data center power budgets can support. Fourth, Asian technology earnings estimate upgrades have spread from core AI beneficiaries to analog chips, second-tier foundries, wafers, and MLCC. Fifth, future bottlenecks may gradually shift from chips to equipment, cleanrooms, packaging, interconnect, and power, and the investment themes will be reordered accordingly.

Analysis framework

The report uses a top-down AI compute cycle analysis and cross-validates it with public cloud revenue and backlog, hyperscaler capex forecasts, the breadth of Asian technology earnings estimate revisions, supply-chain inventories, historical valuation ranges, and supply-demand structures in each subsector. It also screens key industries and stocks based on the potential path of bottlenecks migrating from chips to equipment, interconnect, and power.

Methodology notes

• Technological progress and demand analysisAI Scaling Laws

  Use the empirical relationship between compute input and model capability improvement to assess the sustainability of AI training demand.

  The report applies the rule of thumb that roughly a 10x increase in training compute can bring about a roughly 2x improvement in intelligence, and combines this with performance advances in frontier models every few months and the direction of recursive self-improvement to judge that model evolution has not materially slowed.

• Demand proxy indicatorPublic cloud revenue barometer

  Use the revenue growth of large public cloud platforms as an approximate indicator of AI compute consumption intensity.

  Because much of the compute used by AI labs and other vendors is consumed through hyperscale cloud platforms, the report uses the incremental public cloud revenue and backlog of Google, Microsoft, Amazon, and Oracle to verify inference demand and capex sustainability.

• Technology cycle assessmentCyclical peak indicator framework

  Assess whether the technology cycle is approaching a peak through inventory buildup, slowing demand growth, and rapid supply catch-up.

  GPUs, ASICs, and memory currently show no significant inventory accumulation, demand is still accelerating, and supply has not rapidly caught up, so traditional cyclical peak signals have not yet flashed.

• Earnings momentum analysisBreadth of earnings estimate revisions

  Observe both the direction of earnings estimate changes and the scope of the supply chain participating in upgrades.

  Earnings upgrades have spread from direct AI beneficiaries to analog chips, second-tier foundries, wafers, and MLCC, indicating improvement in both the health of the recovery and the breadth of supply-chain coverage.

• Valuation analysisHistorical P/E range comparison

  Compare forward P/E with long-term historical averages and standard-deviation ranges.

  Excluding memory, Asian technology stock valuations are about 1 standard deviation above the 10-year average forward P/E; combined with structural improvements in margins, returns, and industry concentration, the report believes valuations are not excessive.

Asset mapping & comparison

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

• Asian technology stocks

  Overall buy direction

  Strengths

  AI demand, earnings upgrades, and cloud capex remain strong, and valuations are more reasonable after the recent correction.

  Weaknesses

  Investor positioning is crowded, and the market is highly sensitive to component price increases and capex financing risks.

  Comparison

  Compared with the previous two drawdowns of more than 20%, concerns are higher this time, but leading fundamental indicators are not weaker.

  Risks

  Unexpected earnings estimate downgrades, capex cuts by cloud service providers, rising financing costs, and data center deployment delays.

• Semiconductor manufacturing equipment: Tokyo Electron, Advantest, ASMPT, GPTC, Chroma, Hon Precision, Naura, AMEC, ACM Research

  Best-positioned subsector over the next 12 months

  Strengths

  TSMC, memory leaders, and catch-up players all have incentives to raise capex; equipment and cleanrooms may become new bottlenecks in 2027 to 2028, and semiconductor localization in China also provides incremental demand.

  Weaknesses

  Orders are highly dependent on the realization of foundry capex, and volatility may be amplified after industry expectations are revised upward.

  Comparison

  Compared with other Asian technology subsectors, semiconductor manufacturing equipment has more direct exposure to wafer manufacturing expansion and advanced process investment.

  Risks

  Capex delays, export restrictions, project execution delays, and easing supply bottlenecks.

• IC substrates: Ibiden, Unimicron

  Top pick among components

  Strengths

  Benefiting from larger AI accelerator package areas, server CPU demand, EMIB-T penetration, and incremental CPO demand; new capacity will be limited over the next two years, and margins still have room to recover.

  Weaknesses

  Demand is closely tied to advanced packaging roadmaps, the pace of server investment, and major customer product cycles.

  Comparison

  It has the strongest supply-demand fundamentals among component subsectors and still has substantial room for margin and earnings upgrades compared with historical peaks.

  Risks

  Changes in packaging technology roadmaps, customer concentration, capacity expansion exceeding expectations, and slowing server demand.

• TSMC, MediaTek

  Core semiconductor platform top picks

  Strengths

  Benefiting from demand for AI chips, advanced processes, and edge devices, while TSMC capex still has upside revision potential.

  Weaknesses

  Valuation and earnings are sensitive to AI demand expectations, advanced process utilization, and customer product timing.

  Comparison

  Compared with single-component suppliers, platform companies benefit from broader exposure, but are also more affected by the overall semiconductor cycle.

  Risks

  AI chip demand below expectations, customer concentration, competition, and geopolitical risks.

• Interconnect and networking: Accton, Delta, Aspeed

  Potential beneficiary of next-stage bottlenecks

  Strengths

  Low AI cluster utilization and the large share of instruction cycles consumed by data transfer will drive upgrades in higher-speed networking, CPO, optical interconnects, and system efficiency.

  Weaknesses

  Some technologies are still in the introduction phase, and the pace of demand release may lag market expectations.

  Comparison

  As compute chip supply increases, the importance of interconnects is expected to rise relative to pure computing.

  Risks

  Delayed CPO commercialization, insufficient network investment, changes in technology roadmaps, and customer capex volatility.

• Advanced packaging and testing

  Structural growth area

  Strengths

  2.5D packaging is spreading to CPUs and edge devices, TSMC's 3D packaging investment cycle has begun, and CPO and NPO bring additional growth.

  Weaknesses

  Capacity expansion cycles are long and technically complex, and near-term supply may be constrained by equipment and cleanrooms.

  Comparison

  Compared with traditional packaging, advanced packaging has higher value content and directly benefits from increasing AI chip complexity.

  Risks

  Slow yield improvement, customer product delays, constrained equipment delivery, and the emergence of alternative technologies.

• Memory

  Fundamentals are solid but strategy is relatively cautious

  Strengths

  Supply is still materially below demand over the next 2 to 3 years, and a correction of more than 40% may support a strong rebound over the next 6 months.

  Weaknesses

  High prices have already prompted some AI accelerators and CPUs to reduce HBM or DRAM configurations, weakening the market narrative of price-inelastic demand.

  Comparison

  Compared with equipment and IC substrates, memory supply-demand data are not poor, but valuation upside is more easily constrained by concerns about demand destruction.

  Risks

  Further customer reductions in memory content, price increases suppressing demand, cyclical capacity expansion, and share prices failing to return to prior highs.

Key data

• Recent market correction25% to 30%Recent correction magnitude for Asian technology stocks and the SOX index; this is the third drawdown of more than 20% in the current AI cycle.
• Public cloud incremental revenue in 2Q 2026Approximately US$15 billionSequential incremental revenue was close to twice that of 1Q 2026.
• Forecast capex growth for major cloud service providers in 202765%After strong growth of more than 100% in 2026, high-speed growth is expected to continue in 2027.
• Net debt-to-equity ratio of hyperscalersApproximately 12%As of 2Q 2026, balance sheets remained generally healthy, and financing can be supplemented through debt and equity markets.
• Asia technology valuationAbout 1 standard deviation above the 10-year average forward P/EExcluding the memory sector, the report believes current valuations are not extreme.
• AI cluster model FLOP utilizationApproximately 20% to 40%Some large clusters are even below 20%, indicating substantial room to improve cluster efficiency and interconnect capability.
• Share of time AI accelerators spend on data transferApproximately 50% to 60%A large share of instruction cycles is used for data movement rather than raw computation, reinforcing the investment case for high-speed interconnects and advanced packaging.
• Recent correction in memory stocksMore than 40%Low levels may support a rebound over the next 6 months, but the report does not currently expect a recovery to the May 2026 highs.
• Duration of memory supply-demand gapThe next 2 to 3 yearsSupply is expected to remain materially below demand, but high prices may prompt customers to reduce memory configurations.
• Potential window for chip bottleneck shiftThe next 18 to 24 monthsBy the second half of 2027 and into 2028, data center deployment progress and power availability may replace chips as the main constraints.

Impact & implications

The current pullback looks more like an expectations reset than a reversal in AI fundamentals. If public cloud revenue, model capabilities, and capex continue to be revised upward, earnings revisions for Asian technology stocks are likely to broaden, and equipment, substrates, packaging, testing, and interconnect may become the main leaders of the next rally. Investors should reduce concentration in narratives that rely solely on component price increases and rotate toward areas with capacity bottlenecks, technology barriers, and structural demand, while also remaining alert to volatility from capex financing, power constraints, and memory content reductions.

Risks

• Free cash flow of hyperscalers may turn negative in the second half of 2026 and in 2027, and financing needs may increase market volatility.
• Tightening debt or equity financing conditions may force cloud service providers to slow AI infrastructure build-out.
• Excessively high AI chip and memory prices may trigger configuration reductions and demand destruction.
• If competition between open-source and closed-source models does not translate into higher compute consumption, hardware capex expectations may disappoint.
• Data center construction delays and insufficient grid or behind-the-meter power may become major bottlenecks from the second half of 2027 to 2028.
• Semiconductor supply expansion faster than demand may lead to easing shortages, rising inventories, and earnings estimate downgrades.
• Commercialization of advanced packaging, CPO, and high-speed interconnect technologies may be slower than expected.
• Key markets and semiconductor supply chains face geopolitical, trade restriction, and regulatory risks.
• J.P. Morgan has disclosed interests with some of the companies discussed, including market making, shareholding, client relationships, or potential investment banking compensation.

What to watch

• Adoption of generative AI, LLMs, and agentic workflows disclosed by Software and Internet companies.
• Whether AI applications further spread from software and programming automation to the financial and healthcare industries.
• Progress by frontier AI labs in recursive self-improvement and model capability breakthroughs.
• Incremental public cloud revenue and backlog at Google, Microsoft, Amazon, and Oracle.
• 2027 capex guidance, financing arrangements, and free cash flow changes at major cloud service providers.
• Whether Asian technology earnings upgrades continue to spread to analog chips, second-tier foundries, wafers, and MLCC.
• Whether inventories of GPUs, ASICs, HBM, and other key AI components begin to accumulate abnormally.
• Whether TSMC and major memory makers materially raise capex.
• Whether equipment, cleanrooms, advanced packaging, and testing capacity become new bottlenecks in 2027 to 2028.
• Changes in AI cluster utilization, CPO penetration, optical interconnect demand, and networking equipment spending.
• Whether power availability and data center deployment cycles begin to limit the actual ramp-up of AI chips.
• Whether high memory prices prompt more customers to reduce HBM or DRAM configurations.