CITIC SEC: Overseas AI models and applications intensively catalyze the driving force, and the computing power industry chain may usher in a new round of growth.

CITIC SEC released a research report stating that recently, there has been a strong demand for overseas reasoning and training computing power, leading to price hikes at both Amazon Web Services and Google Cloud. The demand for computing resources for reasoning and training has significantly increased, supporting the demand for training computing power. Although the visibility of large-scale commercialization of AI applications is still limited, the recent intensive catalysis of overseas AI models and applications is expected to keep the overseas computing power demand on an upward trend, potentially bringing a phase of recovery to previous concerns surrounding computing power. The bank predicts that in the next 3-6 months, the overseas computing power chain is expected to return to a relatively optimistic situation. Considering the current industry evolution and model iteration pace, the report suggests focusing on three types of investment opportunities: first, the rising demand for computing power brings investment opportunities for cloud service providers, with Anthropic's related industry chain potentially benefiting as the core direction for Agent; second, investment opportunities in the overseas computing power chain brought by the demand for computing power. The report also states that the release of unexpectedly high-performing models may lead to a reassessment of model companies. CITIC SEC's main points are as follows: Reports of overseas reasoning and training computing power demand being strong, with both Amazon Web Services and Google Cloud raising prices. Following a 15% price increase for EC2 machine learning capacity blocks by AWS on January 23, Google Cloud announced on January 27 that its cloud network transmission services would undergo a price increase, with the North American region seeing a doubling in prices. The bank believes that this round of price increases for cloud services stems from the continuous rise in demand for GPU reasoning and training. Looking ahead to the next 3-6 months, the continued deployment of reasoning-based AI applications is expected, coupled with intensive model iteration on the training side. The bank forecasts that computing power demand is likely to continue to rise in the short term, with previous concerns of a "computing power bubble" in terms of sentiment and valuation potentially easing in phases. Reasoning side: Accelerated deployment of Agents expected to support reasoning computing power demand. Recently, the AI intelligent agent application MoltBot (formerly ClawdBot) has gained rapid popularity, signifying a step forward for AI agents that can obtain higher permissions for computer operations and handle more complex tasks, resulting in increased computing power consumption. Additionally, products released by Anthropic, such as Claude Code, Agent Skills, Claude Cowork, and Claude in Excel, have expanded the usage scenarios for Agent deployment. Based on actual usage data, token calls have seen rapid growth for 2-3 weeks since early January 2026, with AI application usage continuing to rise. According to The Information, Anthropic has raised its revenue forecast for 2026 and 2027 to $18 billion and $55 billion, respectively. Drawing a comparison to the prosperity that Gemini brought to Google Cloud in the previous round, cloud service providers associated with Anthropic are also likely to see a rapid increase in reasoning demand. Training side: Continuous exploration of scaling limits in the industry supports the demand for training computing power. In terms of models, a new round of models is expected to be intensively released in Q1 2026. In the realm of language models, models like Grok-5 and GPT-6 continue to explore scaling effects, expected to be trained on the new generation computing platform GB300 with more data and larger parameter scales, thereby supporting the demand for training computing power. For multimodal models, video generation models like Veo-4 are still in a phase of rapid iteration, posing higher demands for computing power. The exploration of model capability boundaries in the industry and academia jointly supports the demand for training computing power. Additionally, in the context of intensive model iteration releases, if the capabilities of models like Grok-5, GPT-6, and Veo-4 exceed expectations, the corresponding Nvidia and Google chains are likely to see opportunities for a widespread increase. Particularly, Nvidia's chain, due to Gemini's ability to surpass OpenAI in the previous round, has been suppressed by the narrative of "TPU seizing GPU market share," but this round of model iteration may lead to a greater reassessment of flexibility. Financial report catalysis: The upcoming quarterly earnings reports of U.S. stocks will be a key confirmation point for computing power demand. There are current market concerns about the performance of U.S. CSPs and the sustainability of their capital expenditures, with the quarterly reports of the big four CSPs becoming essential observation windows. Microsoft and Meta's financial reports are set to be released on January 29, providing the first validation of the market's judgment on the demand for computing power and the continuity of capital expenditures. Subsequently, the financial reports of Google on February 5 and Amazon on February 6 will further refine the annual capital expenditure guidance. Nvidia's financial report on February 26 and the GTC conference on March 16-19 are also expected to kickstart the market's assessment of a new round of rising investment cycle in computing power for the year, serving as a significant catalyst for a shift toward an optimistic outlook for overseas computing power narratives. Risk factors: Models not iterating as expected; Technical lock-ins by major companies; Unexpected progress in open-source software; Decreased willingness to pay by users; Increased cost of model usage.