CITIC SEC: Certainty of AI applications improves; focus on alpha opportunities brought by application diffusion, technological upgrading, and high earnings certainty.

CITIC SEC released a research report stating that new AI application directions and commercialization progress, along with changes in US Treasury yields, are the two most critical variables for the A-share technology sector in 2027. On valuation, both China and US technology sectors have reached relatively low levels, while EPS remains in an upward trajectory. On timing, the A-share technology sector in 2027 is likely to bottom first and then rise, with bottoming from this year through early next year. The certainty of AI applications is improving, and attention should be paid to alpha opportunities brought by application diffusion, technological upgrading, and high earnings certainty. Suggested areas of focus: 1) Physical AI: embodied intelligence supply chain orders and mass-production validation of L4 and L3 autonomous driving. 2) AI application diffusion: AI Coding has already crossed the critical threshold first, with AI film and television, AI4S, and enterprise-level and personal agents successively entering the task execution stage; focus on the diffusion of application scenarios and commercialization realization; 3) Opportunities brought by domestic model iteration and capability enhancement; focus on internet platform companies and model companies; 4) Domestic computing power chain: volume ramp of domestic AI chips, commercialization of super nodes, and expansion of domestic storage capacity driving equipment orders, bringing earnings realization opportunities; 5) Overseas computing power chain: focus on structural upgrade directions such as OCS and NPO technologies. CITIC SEC's main views are as follows: Market: Model progress and commercialization ARR are the main pricing themes for the 2026 technology rally. In the second half of the year, factors such as industry narrative disruptions, macroeconomic changes, and sector trading crowding led to market adjustments, but overall sector valuations are now at historical lows. On the narrative level, AI Coding and Agents are accelerating deployment. Anthropic's ARR rose from approximately US$9 billion at end-2025 to US$30 billion in April, US$65 billion in July, and US$76.8 billion by end-September, but month-over-month growth slowed to 1.8%, and market attention to the pace of AI investment return realization has clearly increased. On price performance, as of October 27, the Philadelphia Semiconductor Index had risen 62.11% year-to-date, outperforming the Nasdaq Index's 12.60%; the STAR 50 rose 64.25% cumulatively in the first half before falling 25.90% in July alone, turning to a 6.84% decline year-to-date, while the Hang Seng Tech Index fell 31.10% year-to-date. On valuation, the STAR 50 TTM PE is at the 18.6% historical percentile since the "924" stimulus, while the ChiNext Index has returned to levels similar to the lows of previous rate-hike cycles in 2017-2018 and 2023-2024, at a low historical percentile of 4.5% on a ten-year basis. The Nasdaq TTM PE is close to the lows of previous rate-hike cycles, at the 9.1% historical percentile since 2016. Hang Seng Tech Index valuations are likewise at low levels, with TTM PE already one standard deviation below the average since 2020. On positioning, public fund technology holdings fell 7.7ppts from Q2 to 56.3% in 26Q3, but remain at an extremely high historical percentile of 98.8% since 2005; A-share TMT margin financing balances fell 28.5% from the July peak to RMB773.9 billion, with leverage pressure entering the mid-to-late clearing stage. In summary, this round of adjustment is more a rebalancing of valuations and positioning driven by changes in long-term industry narratives and expectations; computing power supply and demand remain tight, industry profit trends remain strong, and overall sector valuations are already at historical lows. Outlook for 2027: Technological breakthroughs and commercialization progress are core variables. Attention should be paid to narrative changes in suppressing factors such as the interest rate environment, geopolitical games, and Capex funding gaps during the 27Q1 window. On timing, A-share technology is likely to bottom first from this year through Q1 next year and to bottom first and then rise over the full year. On the macro level, a review of historical rate-hike cycles shows that for US technology stocks, earnings growth can offset valuation pressure from moderate rate hikes, and technological breakthroughs and a shift in rate expectations can drive recovery before rate cuts materialize, similar to the Nasdaq's trajectory in 2017 (moderate rate hikes + steady EPS growth) and 2024 (generative AI technological breakthrough + rate-cut expectations). Note, however, that for A-share technology, the late stage of a rate-hike cycle still exerts sustained pressure on the sector, and the A-share technology sector is more sensitive to liquidity inflection points. After the adjustment since mid-year, valuation pressure on the STAR and ChiNext indices has been significantly released, but the tight balance of high AI Capex investment places higher demands on earnings realization and commercialization loops; if long-end rates remain elevated, this will still exert periodic pressure on technology valuations. On the policy level, the US political cycle and geopolitical games do not change short-term industry fundamentals trends but increase long-term uncertainty. The FCC ban on Chinese optical modules that the market worries about remains at the draft and proposal stage; converting political posturing into a substantive commercial ban faces a high institutional threshold, and commercial cloud procurement has not been hard-constrained, but the US election results and subsequent FCC personnel composition need to be tracked. At the most core Capex level, the bank believes global AI Capex will still maintain high growth in 2027, but the pace of investment exceeds the pace at which cloud providers generate revenue; it is foreseeable that AI capital expenditure's reliance on external financing will increase. On the liquidity level, March 2027 is the peak for Hong Kong stock lock-up expirations, estimated at HK$125.8 billion in a single month, of which the information technology industry accounts for approximately HK$298.4 billion, or about 65%; attention should be paid to its disruption to liquidity at that time. Foundation models: Capabilities are steadily improving and application scenarios are accelerating diffusion. The launch of models trained on Rubin in Q1 next year is a window to watch, and domestic models will also enter a major version iteration window around the Spring Festival; attention should be paid to RSI progress. On the supply side, according to METR, the task duration models can complete has increased from about 4 minutes for GPT-4 to about 39 minutes for o1 and about 352 minutes for GPT-5.2 (high), with models' long-horizon task capabilities continuing to extend. The IPO window for model makers is approaching; Anthropic has officially filed its S-1, and referencing Z.AI's pace of intensively releasing flagships before listing, model makers have incentives to release stronger flagships before going public, and iteration cycles may further accelerate. On the demand side, combined capital expenditure by Alibaba, Tencent, ByteDance, Baidu, and Kuaishou in 2026 is expected to reach RMB790 billion, roughly double that of 2025; according to CITIC SEC's internet team forecast, the above companies' capital expenditure will reach RMB891.8 billion and RMB920.5 billion in 2027 and 2028, respectively; according to CITIC SEC's forward-looking team forecast, the four major North American cloud providers' combined capital expenditure in 2026 will be US$735 billion, up 79% year-on-year, with the four major US CSPs' Capex expected to be nearly US$1.17 trillion (+59%) in 2027 and nearly US$1.29 trillion (+10%) in 2028, with inference demand growth supporting computing power investment. On the application side, AI Coding has crossed the critical threshold first and settled into a general foundation for agents, with leading products' ARR showing explosive growth; in AI film and television, according to DataEye, the domestic AI drama and anime drama market is expected to reach RMB40 billion in 2026, up 138% year-on-year, with the overseas market exceeding US$4 billion; enterprise-level AI, vertical AI, AI4S, and Personal Agent are successively entering explosive stages, and application diffusion is expected to become the main incremental driver of model commercialization in 2027. Computing power industry: The price increase cycle is entering its latter half, and the domestic chain is entering a volume ramp and harvest period. On the domestic chain, CITIC SEC's electronics team expects domestic AI chip shipments to exceed 5 million units in 2027, with the localization rate further increasing. Driven by material costs and performance improvements, leading products' price increases are expected to exceed 20%, helping maintain high profitability. Beyond single-card performance improvements, the industry is seen shifting from competition on single-chip parameters to comprehensive capability competition in system-level interconnection, unified memory, and software-hardware synergy; domestic computing power is expected to accelerate its catch-up with system-level capabilities, and super nodes may see accelerated development. Meanwhile, domestic storage manufacturers' continued technological progress and capacity expansion will also drive the industry's continued upward cycle. On the overseas chain, 2027 is the first year of Rubin volume ramp; the bank expects optical interconnection demand to continue exceeding expectations and industry supply and demand to remain in tight balance, driven by three factors: cluster scale expansion, speed upgrades, and the shift from copper to optics in Scale-up; at the data center construction level, AIDC's high requirements for uninterrupted stable power supply bring incremental opportunities in power and energy. Physical AI: Giants are entering, and 2027 is the first year of deployment. On the embodied model level, general large-model makers and technology giants such as OpenAI and Nvidia are accelerating entry into physical AI; technological competition is no longer limited to training dedicated VLAs but is expanding to the combination of general models and Siasun Robot&Automation motion control, which is expected to accelerate the evolution of Siasun Robot&Automation brain architecture and reshape value distribution across the industry chain. On the body level, Tesla's Optimus production line broke ground in Q2 2026 and has now begun production on a large-scale production line at its Fremont factory; after the self-developed AI5 chip is integrated into vehicles next April, the on-device computing shortfall will be addressed, and the bank is bullish on Tesla Siasun Robot&Automation entering a mass-production ramp cycle in 2027; in 26H1, China's humanoid Siasun Robot&Automation complete-machine output exceeded 40,000 units, and the bank expects it to exceed 100,000 units for the full year, with domestic supply chains resonating with giants' mass production. On the autonomous driving level, Tesla's Cybercab launch and Waymo's fleet more than doubling from 2025, while domestic players such as Apollo Go and Pony.ai also entering a commercialization acceleration stage, with L4 entering a volume ramp; L3 is still awaiting market pricing, and Huawei's Yinwang expects highway-scenario L3 to achieve large-scale commercial use in 2027, with the biggest variable being the pace of expansion of road capacity where L3 can be used. Risk factors: Risks of intensified geopolitical conflicts; risk of high-valuation stocks pulling back; macroeconomic recovery progress falling short of expectations; risk of relevant industrial policies falling short of expectations; corporate core technology and product R&D progress falling short of expectations; AI application deployment speed falling short of expectations; risk of cloud providers' capital expenditure falling short of expectations; risk of sluggish macroeconomic growth leading to domestic government and enterprise IT spending falling short of expectations, etc.