Open Source Securities: The sustainability of the continued strengthening of medium and long-term computing power demand through Sovereign AI and token consumption.

Open-source Securities released a research report stating that as of May 2025, the daily token call volume of Dou Bag's large model reached 16.4 trillion, accounting for 46.4% of China's public cloud large model call volume, a year-on-year increase of 137 times, and a 29% increase from March; Overseas Google disclosed in May that the monthly token consumption had reached 480 trillion (16 trillion per day), a 50-fold increase in one year. Previous concerns in the market regarding capital expenditure on AI are gradually fading away. With the clarity of the sovereign AI market and no deceleration in the token consumption of leading large model vendors, AI computing power capital investment still has strong sustainability. In investment focusing on generative AI, traditional AI targets that are more engineering-centric, repeatable, and emphasize accuracy of information are expected to benefit. Key points from Open-source Securities: Concerns about computing power needs are gradually diminishing, opening up long-term demand space for sovereign AI. On June 11th, Nvidia announced at the GTC Paris keynote speech that Europe will plan and build 20 AI factories, with 5 super factories exceeding 100,000 chips. Currently, Nvidia is collaborating with Mistral to prepare for the construction of the Nvidia GB system with 18,000 chips; Nvidia's AI factory in Germany will be equipped with 10,000 GPUs; Nebius also announced plans to deploy 4,000 Nvidia B300 chips in the UK in Q4 2025. Although Nvidia's actual degree of realization is affected by factors such as supply chain and product cost-effectiveness, compared to CSP, sovereign AI led by the government has relatively higher certainty, to some extent extending the outlook for computing power demand in the medium to long term. Evidence of the robustness of computing power demand can also be seen from the token consumption of NETDRAGON, a leading application at home and abroad. As of May 2025, Dou Bag's large model's daily token call volume reached 16.4 trillion, accounting for 46.4% of China's public cloud large model call volume, a year-on-year increase of 137 times, and a 29% increase from March; Overseas Google disclosed in May that the monthly token consumption had reached 480 trillion (16 trillion per day), a 50-fold increase in one year. Considering the momentum of subsequent token growth, for Dou Bag, there is still sufficient penetration space in the short term in content review, customer service, search recommendations, and intelligent marketing. In the medium to long term, it can also look forward to a large volume of multi-modal tokens. For Google, in addition to its leading Gemini2.5Pro in the market, AI Overview and AI Mode starting to land on the search engine have also become important drivers of token consumption (and have not caused a decrease in overall search volume). There is considerable room for improvement in the future, combined with Google's announcement of many projects at the 1/O conference, the demand for tokens at the terminal may not have felt a clear turning point. In the context of CoT and model illusions that need improvement, non-Gen-AI industry opportunities are also expected to benefit. Model illusions may not only be a technical issue but also a mathematical issue. In long thought chains or agent solutions, if a complex problem needs to be broken down into 10 steps, each step has an accuracy of 99%, then under compound probabilities, the final accuracy is only 90%, meaning that 1 out of 10 problems cannot be accurately answered. We can also see exploration by model vendors in technical directions (such as Google's Project Mariner). Therefore, in scenarios that are more manpower-engineering, repeatable, and prioritize information accuracy, traditional AI can complement Gen-AI well, have good industry vitality, and benefit targets: FOURTH PARADIGM (06682) (focus on decision-type AI), BAIRONG-W (06608) (focus on the financial industry). Risk Warning: Difficulty in the progress of large model technology, slowdown in AI capital expenditure, and increased market competition.