CITIC SEC Technology October Investment Strategy: RSI Accelerates, Four Clues for Positioning in Q4 Market Trends

CITIC SEC released a research report stating that the next tech market rally window is highly likely to begin in 26Q4, and still requires validation from three conditions: a marginal reversal in rate hike expectations, new narratives coalescing into consensus, and the completion of capital rebalancing. It recommends adhering to structure over position, with priority attention to: 1) Hong Kong model stocks, whose valuation bubbles have been substantially squeezed outfocus on their new version iterations, RSI capability improvements, and ARR realization; 2) internet platforms, which possess C-end distribution gateways and monetization capabilitiesfocus on progress in personal agent products; 3) on the hardware side, prioritize technology upgrade directions such as NPO, OCS, thin-film lithium niobate, 1.6T optical interconnect, and high-end PCB; 4) equipment investment opportunities brought by domestic storage and advanced logic capacity expansion. CITIC SEC's main views are as follows: Market: This rate hike cycle should not be simply analogized to 2022, and the tech sector is gradually meeting conditions for left-side positioning. At the macro level, unlike the previous inflation cycle driven jointly by demand recovery, supply constraints, and energy prices, the current inflation pressure mainly stems from energy supply shocks. US economic activity and AI capital expenditure remain resilient. If geopolitical disruptions ease and oil prices decline, there is a possibility of a marginal reversal in rate hike expectations. At the valuation level, in 2022, the Nasdaq index's PE-TTM and implied EPS fell by approximately 28% and 8% respectively, with valuation contraction and earnings downgrades forming a double suppression; the current tech sector pullback is mainly valuation multiple contraction, while EPS and cash flow remain the main threads of market pricing, and the basis for continued systematic sell-offs is insufficient; domestic tech stock valuations have also declined significantly from their year-to-date highs, and some companies with verifiable performance are gradually meeting conditions for left-side positioning. At the capital level, overseas funds have begun giving positive feedback to new AI narratives, with attention drawn to the capital reallocation opportunity brought by Anthropic's potential IPO window; domestic tech capital leverage pressure has been somewhat released, with A-share TMT margin financing balance falling about 20.5% from its July high, and phased bottoming characteristics are gradually emerging. Earnings realization: CAPEX high prosperity continues, and the visibility of EPS upside for the overseas computing power chain in 2027 is relatively high, which will provide solid support for subsequent market trends. Looking ahead, it is judged that the four major CSPs will continue to view AI infrastructure as a strategic investment in competing for model capabilities and cloud market share. Reviewing the CAPEX investment strategies of CSPs over the past two years: 1) On the eve of the Coding explosion at the end of 2025, OpenAI and Anthropic's combined ARR was $30 billion, yet the four major CSPs still invested firmly. 2) The four major CSPs' AI cloud business external revenue in 2026 is approximately $70 billion, while CAPEX is approximately $750 billion. These two points indicate that CSPs do not determine capital expenditure based on current AI revenue and short-term financial returns, but rather make strategic investments based on predictions of future model capabilities, computing power scarcity, cloud business market share, and long-term competitive position. Taking into comprehensive consideration the 2028 RSI expectations, 2027 backlog orders, and 2026 prosperity signals, it is judged that CAPEX will continue to be a strategic investment for the four major cloud providers in the future, and computing power supply remains the primary constraint on CAPEX. The four major CSPs' capital expenditure in 2027 is nearly $1.4 trillion, up about 85% year-over-year; in 2028 it is nearly $1.9 trillion, up about 35% year-over-year. Outlook: RSI is accelerating, and scenarios such as enterprise-level, AI Research, and Cowork are expected to be the first to contribute commercialization increments. RSI has already shown initial results in tasks that can be formally verified, such as Coding and scientific research: the proportion of R&D work led by Claude internally at Anthropic increased from 1% in March to 26% in August; OpenAI's agent research capability has reached a level of 3.1 Agents per human workday; domestic models such as Alibaba Qwen, Z.AI GLM, and Xiaomi MiMo have also successively announced RSI progress. However, in areas such as cybersecurity, long-horizon tasks, and open-ended scientific research, RSI still needs time: OpenAI's internal success rate for 32-hour+ long-horizon tasks with zero human intervention is only about 20%, and about 700 internal agents autonomously coordinated to intrude into Hugging Face infrastructure; Anthropic Mythos 5 once uploaded a malicious software package to PyPI. Before RSI achieves an autonomous closed loop, it is recommended to focus on enterprise-level scenario applications such as AI for Science, AI Research, AI for Chip Design, Cowork, and finance. In April, OpenAI launched its first life sciences-specific model GPT-Rosalind; in June, Anthropic released Claude Science; in August, OpenAI released its first self-developed inference chip Jalapeno; in September, OpenAI and Kimi successively released enterprise-level products for legal, financial, and other sectors. 26Q4 is expected to be an important observation window for a new round of version iterations of domestic models, and the narrowing of model capability gaps and commercialization realization are expected to become important catalysts for the revaluation of domestic model stocks. Application deployment: Meta Muse goes viral, personal AI assistants achieve large-scale deployment, and the integration of model capabilities with the Meta ecosystem is expected to open new space for C-end commercialization. On September 8, Meta officially released its personal AI assistant Meta Muse, powered by the Muse Spark model. Post-launch data performance was strong: in less than two weeks it topped the free app charts on the US App Store and Google Play, with initial downloads and daily active users both exceeding ChatGPT. Muse's popularity comes from the integration of model capabilities with the Meta ecosystem: Muse focuses on C-end lifestyle scenarios, with conversational interaction, goal tracking, proactive suggestions, and an open connector ecosystem, deeply integrated with third-party applications such as Slack, Shopify, Stripe, and Spotify, and leveraging the approximately 3.6 billion daily active user ecosystem of Facebook, Instagram, and WhatsApp for traffic diversion. Muse opens new space for AI application commercialization: on the one hand, subscription-based, with Muse currently adopting three tiersfree, $20/month, and $100/monthand planning to explore a transaction commission system; on the other hand, Meta announced it will integrate Muse into Meta AI glasses, Muse Charm, and other smart hardware. Muse's popularity also boosts computing power demand: if the existing architecture is maintained, 100 million Muse users would correspond to approximately 200 million CPU cores, 800PB of memory, 10,000PB of solid-state drives, and approximately 1.6GW of power. It is recommended to monitor commercialization progress such as Muse user and subscription penetration, transaction commission implementation, and smart hardware launches. Hardware: Focus on Q3 earnings clues and new technology opportunities. Rubin platform upgrades, optical interconnect iteration, and domestic equipment capacity expansion constitute the main threads of computing power allocation. On the overseas chain side, the advancement of Rubin mass production will drive upgrades in HBM4, high-speed interconnect, high-end PCB materials, and liquid cooling power supply, with demand spreading from GPUs to the entire computing power system. Optical interconnect allocation should balance current profitability and technological elasticity, tracking 800G/1.6T order deliveries in the short term, and focusing on customer adoption and volume ramp for OCS, NPO, and thin-film lithium niobate in the medium term. On the domestic chain side, Huawei and Alibaba's supernode system upgrades are pushing computing power competition toward system collaboration, elevating the importance of interconnect, packaging, and heat dissipation; global semiconductor equipment sales growth and domestic manufacturers' earnings realization provide validation for advanced logic and storage capacity expansion, with Q3 earnings being a window to further verify order conversion. Physical AI: Mass production and operational validation become core catalysts; focus on supply chain orders and real-world scenario returns. Tesla Cybercab is advancing commercial operations, Optimus is entering the mass production preparation and capacity ramp-up phase, and the focus of physical AI observation is shifting from product demonstrations to delivery and usage effects. Domestic embodied intelligence manufacturers are also continuously advancing in shipments and capacity building, but high growth targets and commercialization quality still need to be distinguished; reliability, cost, and investment returns in real scenarios will determine the sustainability of demand. At the technical level, insufficient real interaction data remains a constraint, and improvements in world models, reinforcement learning, and motion control capabilities are expected to drive subsequent breakthroughs. It is recommended to focus on tracking supply chain orders, actual weekly output, autonomous driving operational data, and regulatory progress, to seize opportunities in the transmission from mass production validation to revenue realization. Risk factors: Risk of high-valuation stocks pulling back; macroeconomic recovery progress falling short of expectations; risk of related 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 weak macroeconomic growth leading to domestic government and enterprise IT spending falling short of expectations, etc.