摩根大通：中国AI大战，“百模大战”已结束，最大的利润池归属大厂，智谱和MiniMax如何突围？

Core Insights - The Chinese AI industry is transitioning from a "hundred model battle" to a phase where commercial viability, model innovation, and global layout are key determinants of success [3][4] - The largest profit pool in the AI sector is expected to flow to major platforms like Tencent and Alibaba, rather than model companies [6][10] - Independent firms like Zhipu and MiniMax are finding niches through "structural neutrality," with Zhipu focusing on localized deployment and MiniMax expanding into global markets [3][6] Group 1: Industry Overview - The number of capable and well-funded model developers in China has decreased from over 200 to less than 10 [3] - The industry is no longer rewarding the ability to create models but rather the ability to sustain operations long-term [3][6] - The competition is shifting from technical capabilities to the ability to build commercial systems [4] Group 2: Profit Distribution - The report emphasizes that the sustainable profit pool in generative AI will likely be concentrated among large internet platforms due to their control over distribution and monetization channels [6][10] - Major platforms have established mechanisms for monetization across various sectors, making AI a tool to enhance average revenue per user (ARPU) and conversion rates [6][10] Group 3: Independent Firms' Strategies - Zhipu's business model is divided into localized deployment and cloud-based services, with 85% of its revenue coming from localized deployment, which has a gross margin of 59.1% [14][16] - MiniMax has a global revenue structure, with 73% of its income coming from international markets, providing it with structural advantages [21][22] - Both companies are expected to achieve significant revenue growth, with Zhipu projected to have a compound annual growth rate (CAGR) of 127% from 2026 to 2030, and MiniMax at 138% [17][25] Group 4: Financial Projections - Zhipu is expected to achieve profitability by 2029, with a normalized net profit margin of 20% by 2030 [18] - MiniMax is also projected to become profitable by 2029, with a normalized net profit margin of 24% by 2030 [26] - Both companies are anticipated to require external financing in 2026 and 2027, with Zhipu needing 5 billion RMB annually and MiniMax needing 700 million USD [18][27] Group 5: Cost Structure Changes - The cost structure of AI companies is shifting from "training-driven" to "inference-driven," with inference costs expected to become the dominant expense [28][34] - Zhipu's inference-related computing cost is projected to rise from 7% in 2025 to 68% by 2030, while MiniMax's will increase from 20% to 72% [34][35] - This shift indicates that future competition will focus on inference efficiency and pricing power rather than merely the size of models [35]