CITIC SEC: Overseas AI applications accelerate, continued optimism for supply chain opportunities benefiting from overseas computing power demand.

CITIC SEC released a research report stating that overseas AI applications have been accelerating this year. From the demand perspective, token consumption has maintained rapid growth, large model calls and income levels are rapidly increasing; from the supply perspective, on one hand, generic applications such as AI search, AI Coding, and Agents based on the LLM model have taken off, while on the other hand, multimodal model capabilities continue to iterate, and image and video generation applications have explosive potential. In addition, applications in various verticals such as marketing, customer service, recruitment, education, healthcare, and law are constantly emerging. There is continued optimism for the supply chain opportunities benefiting from overseas computing power demand, and it is recommended to pay attention to domestic cloud and internet companies with AI infrastructure, model capabilities, and application scenarios, as well as investment opportunities brought by the landing of applications such as Coding, Agents, and image/video generation. Trend 1: Since 2025, overseas AI applications have accelerated landing, with strong demand for tokens. 1) Microsoft's Q1 token processing volume exceeded 10 trillion, a five-fold year-on-year increase, with a record-breaking processing of 5 trillion tokens in just the month of March; 2) Google's monthly token processing volume increased from 970 billion in May 2024 to 48 trillion in May 2025, an increase of about 50 times; 3) OpenAI's May ARR exceeded $10 billion, doubling in half a year; Anthropic's May ARR has reached $3 billion, tripling in 5 months. CITIC SEC believes that whether it is cloud providers rapidly consuming tokens or model companies significantly increasing their ARR, it can prove that B-side and C-side applications are accelerating landing, thereby driving the demand for model calls. In the next stage, with the landing of Multi-agent and multimodal applications, token consumption is expected to maintain rapid growth. Trend 2: From the perspective of income and financing, AI applications are highly centralized, and models are eating up the core value of the most common scenarios. From an income perspective, CITIC SEC compiled statistics on the top 100 overseas applications in terms of June ARR, with a total revenue scale of $33.37 billion. Two model companies, OpenAI ($10 billion) + Anthropic ($3 billion) account for about 39% of the revenue, and the top five applications account for 64%; from a financing perspective, in Q1 2025, the US first-level market AI track attracted $60.9 billion in financing, with OpenAI ($40 billion) + Anthropic ($3.5 billion) + Safe SuperIntelligence ($2 billion) accounting for about 75%. Trend 3: Applications based on LLM models are leading the way, with AI search and AI Coding achieving PMF, while Agents are poised for growth. 1) AI Search: From search engines to answer engines, AI is reshaping the internet entrance. OpenAI launched a connected search + Bing empowerment based on Chatbot, with a weekly activity exceeding 800 million; Google, as the traditional search leader, launched AI Preview for defense, with monthly active users reaching 1.5 billion in Q1, a significant increase of 50% compared to the end of last year; 2) AI Coding: The fastest-growing and first to achieve PMF in the segmented track, Cursor's May ARR has reached $500 million, a fourfold increase compared to the end of 2024, while Windsurf's May ARR reached $100 million, a 150% increase compared to February; 3) General Agents: Manus/Genspark and other agents have emerged, but due to task stability and single task pricing, C-side penetration has not yet been widespread. Attention to follow-up iterations and cost reduction progress. Trend 4: In multimodal AI applications, image generation is growing the fastest, while video generation and companion AI have great potential, and the future may see products for millions of C-side users. 1) Image Generation: Midjourney has amassed approximately 21.2 million members on Discord, with estimated revenue of $200 million in 2023 and projected ARR of $500 million in May 2025 (Contrary Research estimate); 2) Video Generation: Currently, the focus is primarily on the B-side, such as the video company RunwayML with revenue of only about $4.5 million in 2022, increasing to $49 million in 2023 and $122 million in 2024, with Sacra estimating revenue to reach $265 million in 2025. AI virtual human video generation platform Synthesia had revenues of approximately $42.8 million and $62 million in 2023 and 2024 respectively, and had surpassed $100 million in ARR by April 2025 (The Information estimate). Trend 5: In the commercialization process of AI vertical applications, customer support, marketing tools, and education tracks are leading the way, achieving large-scale breakthroughs based on the scene characteristics of "closer to the revenue side + higher tolerance for AI illusions". In the future, pay attention to the progress of agent landing in high-value areas such as law, healthcare, scientific research, and finance. 1) Customer Support: Leading the charge is Palantir, using "big models + data integration" technology to integrate multi-source data for government and business, building end-to-end self-service processes, binding high-stick customers through enterprise subscription services, with ARPU per customer reaching hundreds of thousands of dollars, Palantir seeing a 71% year-on-year increase in revenue from US commercial clients in 2025, affirming the commercial potential of government and enterprise service scenarios. 2) Marketing Tools: Companies like AppLovin are utilizing technologies such as the Axon engine to optimize advertising delivery through AI across the entire chain, expanding into non-gaming verticals like financial technology and CTV (Connected TV), building a "advertiser-developer" bilateral market, with the revenue scale of marketing tool AI applications reaching $4.1 billion as of June 2025, demonstrating the value of technology integration in marketing throughout the process for reducing costs and increasing efficiency. 3) Education: Platforms like Duolingo leverage the characteristics of the education scene, "low trial threshold + short iteration cycle + broad market demand", to rebuild the entire teaching process through personalized learning engines and generative AI, with Duolingo seeing a 40.8% year-on-year revenue growth in 2024. These three types of applications account for over 95% of revenue in vertical applications, serving as a leading benchmark for the commercialization of AI applications from technical verification to scale profitability.