中国互联网：两大AI实验室的一小步……关于战略、竞争与盈利路径的思考

China Internet: One small step for two AI labs... thoughts onstrategy, competition, and the path to profits Minimax, Z.ai give us a first glimpse at AI lab economics.Since the beginning ofJanuary, the Z.ai (a.k.a. Zhipu, Knowledge Atlas) and Minimax IPOs (not covered) havecontributed to buoyant AI sentiment in China, and given public investors a first look atthe economics of AI model labs. In this note, we’ve included comparisons of their growthstrategies, our takeaways from discussions with the companies, some thoughts on thereported financials, and high-frequency data tracking user engagement for key properties. Robin Zhu+852 2123 2659robin.zhu@bernsteinsg.com Charles Gou+852 2123 2618charles.gou@bernsteinsg.com Fast followers in the model layer.Both the Minimax M2 and Z.ai GLM-4.5 modelsbenchmarked well when released in Q4 2025, and subsequent updates continue torank highly on publicly available AI leaderboards like Artificial Analysis. Z.ai was notablyidentified by OpenAI’s PR department last year as having made “notable progress” onmodel development. In our discussions, both companies pointed to next-generation (GPT-5grade?) models launching in Q1 2026. We expect the combination of rapid model iterationand cheaper token costs to remain a powerful combination as these companies develop,and expect revenue growth to remain rapid as they expand both domestically and overseas. Min-Joo Kang+852 2123 2644minjoo.kang@bernsteinsg.com Charlie Peng+81 3 6777 6993charlie.peng@bernsteinsg.com Minimax’s pivot to enterprise was inevitable.Whereas Z.ai has focused exclusively onthe enterprise and developers market, Minimax’s revenues have alternately been drivenby AI chatbots (Talkie/Xingye) and AI video generation (Hailuo)… before the more recentpivot to its Open Platform business. With AI increasingly shifting from chatbots to “doingstuff” (e.g. see Qwen), we expect the social and transactional ecosystems surroundingconsumer-facing AI to matter as much as the underlying models… which puts the largeInternet ecosystems we cover at an advantage. We suspect a similar dynamic will play outoverseas - see Google’s build-out of its Universal Commerce Protocol. The economics of China AI.In the foreseeable future, we expect the ability of theseAI labs to iterate model performance and application layer engagement to matter morethan losses and cash burn. But the combination of reasonably solid balance sheets and60-70% gross margins on key segment revenues feels reassuring as far as sustainabilityis concerned. Going forward, the extent to which these companies can drive leverage onmodel training costs will be key to watch. The fact these leading AI labs plan to increasetraining spend at 30%+ CAGR - while inference costs presumably grow faster - feelssupportive of the growth outlook for domestic sellers of GPU compute (e.g. Alicloud). Dance while the music still goes on.Minimax and Z.ai remind us of past waves of techstart-up listings in Hong Kong, like the EV start-ups of 2018-2020, and autonomousdriving companies in 2024. We’re AI bulls at a high level, and look forward to following theirprogress as newly listed entities. In our recent travels in Europe, it was clear the market’sappetite to fund AI narratives (both in China and globally) remained immense… and centralto investor interest in the China Internet sector. We also encountered pockets of investordiscomfort with valuations anchored on price-to-ARR multiples, and sum-of-the-partsgymnastics… but all-in it was clear the music still goes on. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS We’re generally bullish on the Chinese Internet companies’ ability to remain fast followers on AI, and ability to continue to winshare overseas. The recent Minimax and Z.ai (a.k.a. Zhipu, Knowledge Atlas) IPOs have provided rare insight into the innerworkings and financial state of AI model labs. The world is a big place, and we expect the combination of rapid model iterationand cheaper token costs to remain a powerful combination as these companies grow both domestically and overseas. While weexpect the large Internet platforms we cover to remain dominant in consumer-facing AI, leveraging entrenched user habits andlarge transactional ecosystems, we think there should be ample room for AI labs like Minimax and Z.ai to grow substantially. Within our China Internet stock coverage, Tencent and Alibaba remain top picks. On the margin, while we’re happy to seeinvestors reward Alibaba for its AI-related endeavours, some perspective is probably wise when it comes to recent excitementaround the potential T-head spin-off. We suspect group-level PE multiples will represent a valuation constraint at some point,in the context of ongoing negative revisions in the company’s e-commerce business. In contrast, we think Tencent’s recent de-rating - related to sentiment headwinds around the speed of the company’s AI chatbot development - presents an opportunityto accumulate. While the recent teet