CITIC Securities: Overseas AI narrative may return to an optimistic outlook

CITIC Securities Research Report states that recently, overseas inference and training computing power demand has been strong, with Amazon Web Services and Google Cloud both raising prices. On the inference side, products like MoltBot (formerly ClawdBot), Claude Code, and other Agent products are accelerating deployment, leading to a significant increase in demand for cloud computing resources. Since early 2026, token call volume has been growing rapidly for 2-3 consecutive weeks. On the training side, models such as Grok-5 and Veo4 are continuously iterating, with industry players exploring the limits of scaling to support training computing power demand. Although the current visibility of large-scale commercialization of AI applications remains limited, looking ahead 3-6 months, with the intensive deployment of inference AI applications and the continuous iteration of training models, CITIC Securities forecasts that demand for computing power may further increase. Once the demand during the US earnings season is confirmed, concerns about a “computing power bubble” that previously suppressed sentiment and valuations in the computing industry could see a phased easing, potentially leading to a new wave of growth in the computing industry chain.

Full Text Below

Technology | Overseas AI Narrative May Return to Optimism

Recently, overseas inference and training computing power demand has been strong, with Amazon Web Services and Google Cloud both raising prices. On the inference side, products like MoltBot (formerly ClawdBot), Claude Code, and other Agent products are accelerating deployment, leading to a significant increase in demand for cloud computing resources. Since early 2026, token call volume has been growing rapidly for 2-3 consecutive weeks. On the training side, models such as Grok-5 and Veo4 are continuously iterating, with industry players exploring the limits of scaling to support training computing power demand. Although the current visibility of large-scale commercialization of AI applications remains limited, looking ahead 3-6 months, with the intensive deployment of inference AI applications and the continuous iteration of training models, we forecast that demand for computing power may further increase. Once the demand during the US earnings season is confirmed, concerns about a “computing power bubble” that previously suppressed sentiment and valuations in the industry could see a phased easing, potentially leading to a new wave of growth in the computing industry chain.

▍ Origin of the Report: Overseas inference and training computing power demand is strong, with AWS and Google Cloud both raising prices.

After AWS implemented approximately a 15% price increase on EC2 machine learning capacity on January 23, 2026, Google Cloud announced on January 27 that its cloud network transmission services would undergo a price hike, with North America experiencing doubling of prices. We believe this round of cloud service price hikes is driven by continued growth in GPU inference and training demand. Looking ahead 3-6 months, inference AI application deployment is expected to continue, combined with intensive model iteration on the training side. We forecast that short-term demand for computing power may further increase, and concerns about a “computing power bubble” that previously suppressed sentiment and valuations could see a phased easing.

▍ Inference Side: Agent Deployment Accelerates, Potentially Supporting Inference Computing Power Demand

Recently, AI agent application MoltBot (formerly ClawdBot) has gained rapid popularity. AI Agents continue to advance, capable of obtaining higher permissions for computer operations and handling more complex scenarios, which also increases computing power consumption. Additionally, products like Claude Code, Agent Skills, Claude Cowork, and Claude in Excel released by Anthropic have expanded the deployment scenarios for Agents. Based on actual usage data, according to OpenRouter, since early January 2026, token call volume has maintained high-speed growth for 2-3 weeks, with AI application usage continuously rising. According to The Information, Anthropic has raised its revenue forecasts for 2026 and 2027 to $18 billion and $55 billion respectively. Similar to the previous Gemini boom that brought prosperity to Google Cloud, Anthropic’s cloud service providers are also expected to see a rapid increase in inference demand.

▍ Training Side: Industry Continues to Explore Scaling Limits to Support Training Computing Power Demand

On the model front, a new round of models is expected to be densely released in Q1 2026. Regarding language models, Grok-5, GPT-6, and others are exploring scaling effects, likely trained on more data and larger parameter scales on the new generation of computing platforms GB300, to support training demand. For multimodal models, such as Veo-4 and other video generation models, rapid iteration continues, demanding higher computing power. Industry and academia’s joint exploration of model capability boundaries supports the demand for training computing power. Furthermore, in the context of dense model releases, if models like Grok-5, GPT-6, and Veo-4 exceed expectations in capabilities, the corresponding Nvidia and Google chains could see a broad rally. Nvidia’s chain, previously suppressed by narratives of “TPU overtaking GPU” due to Gemini 3’s capabilities surpassing OpenAI, may see greater revaluation potential during this model iteration window.

▍ Earnings Catalysts: US Stock Quarterly Reports as Key Moments for Confirming Computing Power Demand

Currently, the market is concerned about the performance of US CSPs and whether Capex can continue. The quarterly reports of the four major CSPs will be important indicators. On January 29, 2026, Microsoft and Meta released their earnings, which may verify market expectations regarding sustained demand for computing power and capital expenditure. Subsequently, Google’s report on February 5 and Amazon’s on February 6 will further refine the annual capital expenditure guidance. The Nvidia earnings report on February 26 and the GTC conference from March 16-19 are also expected to trigger market judgment of a new upward cycle in annual computing power investment, serving as important catalysts for the shift towards optimism in overseas AI narratives.

▍ Risk Factors:

• Underperformance of foundational model iterations
• Major vendors engaging in technological blockade
• Macroeconomic deterioration leading to tighter risk investment
• Open-source software advances beyond expectations
• Macroeconomic deterioration reducing user willingness to pay
• Increased costs of using foundational models

▍ Investment Strategy:

Although the visibility of large-scale commercialization of AI applications remains limited, recent intensive catalysis of overseas AI models and applications suggests that overseas demand for computing power is likely to maintain an upward trend, with phased recovery of previous concerns. We judge that in the next 3-6 months, the overseas computing power chain may return to a relatively optimistic scenario. Based on current industry evolution and model iteration pace, we recommend focusing on three types of investment opportunities: first, cloud providers benefiting from rising computing power demand, with the Anthropic-related industry chain potentially being a core beneficiary of Agent; second, overseas computing power chain investments driven by demand, where Nvidia’s chain, previously suppressed by narratives of “TPU overtaking GPU,” may see greater revaluation potential during this model iteration window; third, potential overperformance of models themselves could lead to revaluation of model companies.

(Source: People’s Financial News)