CITIC SEC: Continue to be bullish on the growth potential of Chinese and US models; the investment value of domestic models is expected to gradually emerge.

CITIC SEC released a research report stating that although the current market attitude toward large models is becoming more cautious, it remains firmly bullish on the long-term commercial value and valuation growth potential of Chinese and US models. At present, the RSI closed loop is accelerating, and model iteration continues to speed up. With Anthropic's listing window approaching, coupled with the gradual implementation of enterprise-level commercialization, this brings an opportunity for model value re-rating. Overseas, the focus is on ARR growth after year-end FDE project acceptance, while domestically, the market is waiting for breakthroughs in model capabilities. Currently, domestic model capabilities lag overseas by about 6 months, while commercialization lags by about 18 months. If the capability gap remains stable, domestic models are expected to accelerate their catch-up in commercialization. With a lower revenue base, faster growth rate, and more attractive valuation, they are expected to usher in a systematic re-rating, and the investment value of domestic models and internet companies is expected to gradually emerge. The main views of CITIC SEC are as follows: Report origin: The marginal slowdown in Anthropic ARR triggered market volatility. However, the combined ARR growth of OpenAI and Anthropic remains solid, and the firm remains firmly bullish on the commercial value and valuation growth potential of Chinese and US models. Since July, market concerns about the slowdown in Anthropic ARR growth have intensified, mainly due to two reasons: first, prior expectations were overly optimistic, as the market linearly extrapolated the April single-month ARR increment of USD 17 billion to May-June; second, since July, OpenAI has continued to take share from Anthropic. The confusion among third-party data calibers has also exacerbated fluctuations in ARR expectations. However, looking at the combined scale of the two companies, ARR growth remains solid. The average monthly ARR increase in August-September was USD 15.9 billion, close to the combined increase of about USD 17 billion in April. On a full-year basis, if extrapolated at a monthly increase of USD 15 billion, ARR is expected to reach USD 180 billion by year-end, 5.9 times that at the end of 2025, faster than the 4.7 times and 3.8 times in the previous two years. Looking ahead, after Anthropic's November IPO, it may disclose more fundamental information and provide clearer guidance, at which point the model layer may see a new round of capital allocation and rebalancing. Frontier outlook: The RSI closed loop is accelerating, with new scenarios such as Cowork, AI4S, AIDD, and AI for Chip Design landing first. RSI is progressing smoothly. Within Anthropic, the share of R&D work led by Claude rose from 1% in March to 26% in August, and the share of work involving deep AI collaboration or above has exceeded 90%; OpenAI's agent research capability has reached a level of 3.1 Agents per human working day. However, fully realizing the RSI closed loop still takes time. Before RSI achieves a fully autonomous closed loop, scenarios such as Cowork, AI4S, and AI for Chip Design are expected to contribute incremental value first. In AI4S, OpenAI and Anthropic released GPT-Rosalind and Claude Science in April and June, respectively. In AI for Chip Design, OpenAI released its first self-developed inference chip, Jalapeno, in mid-year, participating in chip design through internal models and reinforcement learning algorithms; Mythos 5.1 rewrote the GPU kernels of seven learning models, increasing inference speed by up to 2.5 times. Commercial implementation: In the short term, enterprise-level implementation enters an acceptance period at year-end, while on the consumer side, Meta Muse ignites expectations for an application explosion. In the long term, the firm believes it is still in the very early stage of AI commercialization. Current model capabilities maintain linear growth, but payment and usage remain highly uneven, showing commercialization characteristics similar to the early mobile internet era. AI spending by the top 1% of enterprise employees is 10 times that of the top 10% and 576 times that of the top 50%. In terms of per capita usage, California is 6.5 times West Virginia; the United States is 21 times India. Models show commercialization characteristics highly similar to the early mobile internet. Experience from the mobile internet shows that technology adoption requires a process from concentration to equalization. In 2010, the top 1% of users generated 52% of monthly data traffic in North America; by 2019, the top 1% of users accounted for only 5% of monthly data traffic. In the future, even considering only the diffusion of demand to more enterprises, scenarios, and regions, the commercialization space remains broad. In the short term, enterprise-level scenarios such as finance are expected to lead the acceleration of commercialization. Referring to historical SaaS experience, from small-scale validation to actual commercialization implementation, it takes at least 6 months for FDEs to help enterprises achieve large-scale deployment. Based on this, 26Q4 will be a key time point for confirming large-scale enterprise-level implementation of overseas models. Once ARR re-accelerates, market sentiment is expected to reverse. Domestic models: Demand for data annotation is exploding. If the capability gap is maintained, it is expected to drive a systematic re-rating. Demand for data annotation is exploding, the post-training data bottleneck is easing, and domestic model post-training will ultimately break through in the cracks. Currently, the combined ARR of domestic models is about USD 8 billion, equivalent only to the level of overseas models 18 months ago, while valuation is only one-tenth that of overseas models. Once domestic model capabilities break through, a lower revenue base, faster growth rate, and more attractive valuation are expected to drive a systematic re-rating of domestic models. Risk factors: Risks of model technology innovation and capability breakthroughs falling short of expectations; risks of increased capital expenditure leading to depreciation eroding profits; risks of computing power supply falling short of expectations; risks of AI application commercialization falling short of expectations; risks of intensifying industry competition; risks of export controls on hardware in the computing power chain.