Author: CITIC Securities
On November 12, U.S. time, NVIDIA’s board announced that Jensen Huang has stepped down as President, CEO, and director, effective immediately; CFO Colette Kress has been appointed interim CEO. Since November 5, NVIDIA’s stock has declined by over 22%. We believe that after management changes at leading overseas companies, the likelihood of shifts among their customers, supply chains, and overall strategies has increased. Additionally, according to The Information, the U.S. Department of Commerce is finalizing new export control rules for AI chips, aiming to bring countries such as those in Southeast Asia under stricter export licensing requirements, making it increasingly difficult for NV-series cards to enter the domestic market, which may cause short-term volatility in market expectations regarding Chinese CSPs’ access to computing power. In the short term, stricter export controls are expected to slow the progress of overseas data center construction and reduce visibility into CSP growth rates. In the medium to long term, we expect the proportion of domestic CSPs procuring domestic computing power to further increase; as domestic products iterate and supply capacity strengthens, domestic computing power is poised to become a cornerstone of domestic AI-related Capex and capture a significant mid-term market share.
Item:
Jensen Huang has stepped down from all positions at NVIDIA, with the CFO assuming the role of interim CEO. On November 12 (U.S. time), NVIDIA’s board of directors announced that Jensen Huang has resigned as President, Chief Executive Officer, and director, effective immediately; Colette Kress, Executive Vice President and CFO, has been appointed interim CEO while continuing to serve as CFO. The board has initiated a search for a new CEO and has appointed Harvey C. Jones as the company’s first Chairman of the Board. As a co-founder who led NVIDIA for 33 years, Huang’s deep personal ties to the company’s strategy and customer relationships mean this leadership transition will have far-reaching implications beyond a typical personnel change.
Domestic semiconductor and computing power chain: The substitution logic is significantly strengthened, with the deployment of super nodes driving the domestic chain into a period of performance realization.
Following management adjustments at leading overseas companies, coupled with iterative improvements in domestic products and enhanced supply capabilities, domestic compute power is poised to become the cornerstone of domestic AI-related Capex, capturing a significant mid-term market share. On the demand side, the difficulty of introducing NVIDIA series cards into China continues to rise; our estimates project China’s existing compute load to reach 30.4 GW by 2026 and 97.4 GW by 2030. Favorable conditions—including state-owned enterprise support, lower credit costs, and robust infrastructure capacity—confirm that China’s compute infrastructure development remains in a clear acceleration phase. Under policy mandates for domestic localization, breakthroughs in performance and increased production capacity of domestic AI accelerator cards are critical. In this shifting landscape, domestic AI chip manufacturers are likely to receive greater supply support. Additionally, the ramp-up in production capacity and improved yields of domestic advanced process lines provide a solid foundation for the scaling of domestic AI compute cards. In terms of industry timing, domestic ultra-nodes led by companies such as Alibaba and ByteDance are expected to accelerate significantly by the end of this year and into next year, driving clear increases in both supply and demand for domestic cards. The interconnection segments of ultra-nodes are anticipated to experience a substantial surge in value, with the domestic supply chain expected to gradually realize performance gains starting in Q4 2026.
Overseas hashpower chain: Investment and financing pressures were already present; management changes have added further disruption—maintain caution regarding the short-term outlook for overseas chains.
We estimate that the global (excluding China) existing computing power load will reach 69.3 GW by 2026, and at the current pace, it is projected to rise to 322.4 GW by 2030. However, under this rapid growth, North American computing power construction faces challenges on both the investment and financing fronts: On the investment side, macroscopically, AI capital expenditures are set to reach a peak share of GDP for a single industry; microscopically, CSP end revenues are only one-seventh of their investments, making breakthroughs on the application side critical; on the financing side, considering CSP funding gaps and demand from new cloud factories, capital pressure is expected to remain relatively manageable through 2027, but deficits will become significant after 2028. Against this backdrop, Jensen Huang’s departure will weaken NVIDIA’s ability to integrate its ecosystem and provide credit guarantees, significantly increasing the likelihood of changes among NVIDIA’s customers, supply chain, and overall strategy, potentially slowing technological iteration in the overseas computing power chain.
Domestic CSP: Increased difficulty in acquiring computing power disrupts short-term performance expectations
According to The Information, the U.S. Department of Commerce is finalizing new export control rules for AI chips, planning to bring countries such as Southeast Asian nations under stricter chip export licensing requirements.
In the short term, stricter export controls are expected to slow the progress of advanced computing center construction, impacting growth continuity and reducing visibility into near-term CSP growth rates. However, given that demand for computing power has not significantly changed, and with supply constraints in place, CSP margins are poised for notable improvement, while the ROIC of previously purchased chips is also expected to rise further. Second, for model companies, the exploration of frontier models’ intelligence limits still heavily relies on support from advanced computing power. Under limited supply, we believe model companies will further optimize their AI infrastructure for training and inference, driving cost reduction and efficiency gains.
In the medium to long term, we believe the proportion of domestic CSPs procuring domestic computing power will continue to rise. Full-stack platforms with strong heterogeneous computing capabilities are well-positioned to achieve more efficient utilization of domestic computing power and capture greater market share.
AI application: The chip structure remains favorable; application diffusion is the primary growth driver for model commercialization by 2027.
In our annual strategy, we assess that the certainty of AI applications is increasing: AI coding has率先 crossed the tipping point and evolved into a universal foundation for agents, with top products experiencing explosive growth in ARR; OpenAI’s internal daily agent workload now equals 3.1 human workloads, and active user numbers surged after Codex was integrated into the ChatGPT desktop app. In enterprise AI, corporate AI spending and LLM adoption rates continue to rise, with on-premises deployment dominating enterprise AI demand; leading vendors are driving a “deployment–feedback–iteration” flywheel through frontline deployment (FDE) models. For personal agents, products such as Qwen Go and Meituan Xiao Tuan have entered the phase of executing real-world tasks; WeChat has opened two AI integration modes, enabling agents to call mini-program services. AI4S and vertical-specific AI are also entering their respective phases of explosive growth.
In the media sector, with domestic video model capabilities advancing and costs declining, we remain bullish on the flourishing of AI-generated content, particularly AI long-form dramas and AI films, as batches of high-quality content are expected to launch. We also closely monitor the industry catalyst expected from ByteDance’s world model release in December. On the platform side, WeChat is accelerating its entry into the short-form drama ecosystem, enabling upstream IP holders and content creators to potentially benefit from short-term gains—pay special attention to companies possessing high-quality IP and content. From the perspective of shareholder positioning and trading volume share, the media sector still has considerable upside compared to the first half of 2023 and early 2026, with a healthy shareholder structure. Volatility in other sectors, from a capital flow standpoint, supports potential upward momentum for the media sector going forward.
Investment strategy:
From a configuration perspective, the A-share AI hardware industry chain currently lacks sufficient upside potential to hedge against uncertainty; the active public fund allocation to TMT remains at the 98.4th percentile since 2010. The sector will shift from being driven by incremental capital to competing for existing or even reduced capital, requiring time for the capital ecosystem to recover, with momentum and fundamentals-based strategies weakening and rotation accelerating.
Within the technology sector, we recommend capturing structural opportunities in the following order: 1) Domestic computing power: Adjustments in overseas management and intensified export controls reinforce the logic of domestic localization; domestic supply chains are expected to gradually deliver earnings starting in Q4 2026; 2) Physical AI: Major players entering the space, with 2027 poised to be the year of commercialization; focus on orders in the embodied intelligence supply chain and mass production validation of L4 and L3 autonomous driving; 3) Expansion of AI applications: AI coding has already crossed the tipping point; AI in film and television, AI4S, enterprise AI, and personal agents are now entering the task-execution phase—monitor application expansion and commercialization milestones.
Additionally, pay attention to broad structural recovery opportunities in non-AI sectors: previously, non-AI sectors faced pressure from rising raw material costs, rising long-term interest rates, and foreign exchange losses stemming from AI-related externalities; the main phase of valuation contraction has already been completed this year. As AI investment slows, interest rates and raw material prices decline, and the RMB appreciation pace moderates, the K-shaped divergence is expected to narrow. We recommend stable return strategies represented by PB-ROE (focus on white goods, logistics, chemicals and energy, oil transportation, pharmacies, outbound-oriented securities firms, and financially sound real estate leaders), as well as contrarian positioning and expectation-driven strategies centered on domestic consumption (focus on food and beverage, catering supply chains, airport duty-free, health supplements, and medical device leaders).