Open-source securities: Global multimodal AI acceleration advance, domestic model commercialization speeding up.

Open Source Securities released a research report stating that global multimodal technology continues to iterate, domestic models have achieved partial overtaking in the fields of video and audio generation, and the commercialization process has accelerated significantly. In 2026, the number of token calls for head models has surged, empowering downstream industries such as video, gaming, and marketing. The explosive development of multimodal applications further exacerbates the shortage of computing power, driving high demand for computing power leasing and the AIDC industry chain, and the media and internet industry is presented with vast market opportunities. The main points of Open Source Securities are as follows: - Global multimodal technology continues to iterate, with domestic models moving from catching up to partial overtaking, and overall commercialization accelerating - Since Open AIDALL-E first introduced large language models into the field of image generation in 2021, overseas technology giants and universities have focused on continually iterating multimodal technology models. The quality, efficiency, and cost of content generated by large multimodal models continue to improve. Domestic technology giants have made rapid progress and have achieved partial overtaking in the fields of video generation, audio/music generation. Increase in the maturity of technology and application promotion drive rapid growth in AI native applications' ARR, with Midjourney, Kuaishou Kelin, ElevenLabs ARR reaching hundreds of millions of dollars. Gemini's introduction of Nano Banana has increased its MAU by 200 million in three months since March. The concentration of global talent and capital may continue to drive rapid development of multimodal technology, and the commercialization of models has a promising market and is expected to accelerate. - Large multimodal models have a wide range of downstream applications, which may accelerate the growth of domestic model token calls/ARR. - Large multimodal models can deeply empower content production, marketing, industrial manufacturing, etc., as long as AI creates value higher than token costs in the workflow, the demand is immense. By 2026, the token calling volume of head models will soar, Douba achieved a thousand-fold growth in two years in March, KNOWLEDGE ATLAS's ARR in March increased by 60 times year-on-year, and increased by 6.4 times since the beginning of the year. Compared with Open AI's ARR exceeding 25 billion US dollars, and Anthropic's ARR exceeding 44 billion US dollars (an increase of 389% compared to the end of 2025), the company believes that there is still a vast space for the commercialization of domestic models. Referring to mainstream large multimodal models, compared to a single text, the input token consumption of a single second of audio/single image/single second of video is 1-2/3/2-3 orders of magnitude higher, and the output token consumption is 1/2-3/4 orders of magnitude higher. The company estimates that the daily token required for the domestic video consumption scenario alone is expected to reach 350 trillion, and is optimistic about the application expansion driving the continued high growth of various model revenues. Key recommendations: TENCENT, Kuaishou, beneficiaries: Minimax, Alibaba, KNOWLEDGE ATLAS, Kunlun Tech. - Large multimodal models deeply empower the content industry chain, actively layout the "AI + video/game/marketing" track (1) Video: In the past, high costs of animation/overseas live shooting, under AI production, the cost of ordering a short drama has dropped to tens of thousands of yuan. AI human dramas topped the total broadcast list of Hongguo, selected for the Cannes screening, reflecting that AI content has been recognized by the audience. "Content + Cost" closed-loop may open up incremental markets. Key recommendations: CHINA LIT, beneficiaries: COL Global Co., Ltd., Decai Decoration, Bona Film Group. (2) Games: AI empowers the entire process of game development, lowering the technical threshold to help distribution platforms/UGC ecological games/open world games promote the conversion of players to creators, creating a "content-user" positive cycle. AI reshapes the interactive experience, with narrative and social competitive games likely to benefit first. Key recommendations: Perfect World, NetEase, Giant Network Group, Kingnet Network, XD INC, Bilibili, beneficiaries: Zhejiang Century Huatong Group. (3) Marketing: Large multimodal models help advertising systems achieve deep personalization, with high CPMAI applications commercializing rapidly, increased content under AI efficiency will enhance the importance of advertising, or jointly drive the expansion of the advertising market. Leading programmatic advertising technology and marketing companies closely cooperating with top AI application companies are likely to benefit first. Key recommendations: MOBVISTA, Inly Media Co., Ltd, beneficiaries: BlueFocus Intelligent Communications Group, Easy Click Worldwide Network Technology. - Large multimodal models accelerate penetration or further widen the gap in computing power, computing power leasing/AIDC may benefit significantly - Supply-demand imbalance raises computing power prices, with domestic top cloud companies experiencing a 5%-34% price increase in Q1 2026. The increase in multimodal inputs increases the unit token computing power consumption, and application development may further widen the computing power gap. Bytedance/Alibaba/Tencent's capital expenditure increased by 88%/71%/3% in 2025, and with increased investment in computing power in 2026, companies in the computing power chain are optimistic. Key recommendations: Hangzhou Shunwang Technology, beneficiaries: Zhejiang Huace Film & TV, Zhejiang Daily Digital Culture Group. Risk warning: Iteration speed of large models, commercialization progress of AI applications, unexpected investment intensity in computing power, etc.