Morgan Stanley closed-door meeting: MiniMax (00100) and Z.AI (02513) both experience explosive growth in ARR, and the competition in models will enter a tiered elimination round.

Morgan Stanley recently held a closed-door meeting to engage in in-depth discussions on the latest developments among domestic AI large model companies and major internet firms. Minutes from the meeting indicate that two leading companies, MiniMax and Z.AI, are both maintaining rapid revenue growth. By the end of August, MiniMax's annual recurring revenue (ARR) reached $800 million, and analysts have revised their year-end estimate upward to $1.3 billion. Z.AIs ARR exceeded $2 billion on a weekly basis in August, with its year-end guidance adjusted to $2.4 billion. Tencent's Harmony 4 Preview was released at the end of August, showing slightly better performance in internal blind tests compared to GLM-5.3 and Kimi K3, while the WorkBody open platform is evolving from an office assistant to an Agent operational layer. Analysts believe that current buyers are primarily focused on the capability level of models, followed by ARR growth and gross margin. Model competition is entering a phase of tiered elimination, where companies with state-of-the-art (SOTA) capabilities can launch high-cost-performance versions through distillation, while companies that only have cost advantages struggle to break through technical barriers and face pressure on gross margins. MiniMax: ARR Expected to Exceed Expectations by Year-End, Three New Models Awaiting Release As of the end of August, MiniMax's ARR stood at $800 million, maintaining guidance for year-end revenue of over $1 billion, though analysts have raised expectations to $1.3 billion. The company plans to release three new models, M3.1, H3.1, and M3 Pro, in the second half of the year, and currently has sufficient computing power reserves to support all training within the year and to prepare for next year's 10 trillion parameter model in advance. Z.AI: Rapid ARR Growth, Domestic Chip Cluster Supports Trillion Parameter Training Z.AI's ARR is rapidly increasing, surpassing $2 billion weekly in August, with its year-end guidance raised to $2.4 billion, contingent on the delivery timeline of domestic chips. The GLM-5.3 Flash model, released in August, utilizes a new architecture and performs inference on a domestic chip cluster. This architecture will be used for the GLM-6 model set to be released in October. Z.AI boasts a cluster of 100,000 domestic chips, with its top ten clients contributing over 40% of its revenue, and nine out of the top ten internet companies in China are its clients. Tencent Harmony: Accelerating Iteration, WorkBody Building an Agent Open Ecosystem Tencent's Harmony 4 Preview was released at the end of August, ahead of initial expectations; the parameter size and context length have been significantly expanded, with internal blind test performance slightly exceeding GLM-5.3 and Kimi K3. The WorkBody is positioned as an open Agent ecosystem that integrates hardware, applications, and developers, rapidly advancing its payment and distribution capabilities. If Harmony becomes a SOTA model, it can develop into a Model as Service; even if it is not SOTA, the WorkBody ecosystem still holds commercial value. Going forward, the focus will be on new products and strategic releases at the global digital ecosystem conference in October. Q&A Session Q: What is MiniMaxs ARR as of the end of August? What is the statistical method used? A: As of the end of August, MiniMax's ARR stands at $800 million, using a weekly statistical method consistent with the $400 million disclosed in May; this figure exceeds market expectations. Q: What is MiniMaxs year-end ARR guidance? How do market expectations compare to the companys explanation? A: The company maintains its year-end ARR guidance at above $1 billion, but this guidance is considered conservative; analysts have raised expectations to $1.3 billion. The company has not adjusted the guidance because it is waiting to see the actual performance following the release of new models in the second half of the year, and the potential for upward adjustment remains to be seen. Q: What new models does MiniMax plan to release in the second half of the year? What are their expected release dates and key features? A: MiniMax plans to release three new models: M3.1, H3.1, and M3 Pro. The company currently has ample computing power reserves to support the training of all models within the year and to prepare computing power for next year's 10 trillion parameter model in advance. Q: What were MiniMaxs R&D expenses in the first half of the year? What are the expectations for R&D expenses in the second half and next year? A: R&D expenses in the first half of the year were approximately $300 million; annualized R&D expenses in the second half are expected to exceed $1 billion; next year, R&D expenses will increase further, but the growth rate will be significantly lower than that of revenue and gross margin. Q: What were the main reasons for MiniMaxs decline in gross margin in the first half of the year, and what is the outlook for the second half? A: The decline in gross margin is mainly due to initial poor market response to the M3 model leading to price reductions and user subsidies, the optimization learning curve of new model inference, the lower gross margin associated with the Token plan and the prevalence of discounts, and an increase in the proportion of revenue from text models, whereas multimodal models have a gross margin above 50%. The company points out that most of these factors are one-time events, and it expects improvements in gross margin on a month-on-month basis in the second half of the year. Q: How does MiniMax view the relationship between high intelligence-level models and high cost-performance models? A: MiniMax believes that intelligence level and cost performance are two sides of the same coin: high inference efficiency and low cost can support larger scale post-training, thereby enhancing model intelligence. Thus, inference efficiency is a core capability that cannot be overlooked in the pursuit of high intelligence. Q: What is the planned timeline for MiniMaxs next financing round, and what are the company's intentions? A: The quiet period for the last financing round ends on September 12, allowing the company to initiate financing as early as September 13, though management has indicated that they are not considering financing in the short term, and they plan to proceed after releasing an industry SOTA model, with the market generally expecting the financing to occur after October. Q: What are the ARR figures and statistical methods for Z.AI at various points in time? How does the year-end guidance compare to market feedback? A: Z.AIs ARR figures are: March $250 million, June $530540 million, July $1 billion, August $1.6 billion; based on weekly statistics, ARR in August has surpassed $2 billion. The year-end guidance has been raised to $2.4 billion, and the company believes this guidance is conservative, with actual scale depending on the delivery timeline of computing power, and it is expected that the guidance will be updated in the fourth quarter based on domestic card delivery conditions. Q: What are the key updates in Z.AI's model pipeline for the second half of the year? What is the positioning and features of GLM-6? A: The GLM-5.3 Flash model released in August employs a new architecture and performs inference on a domestic chip cluster; this architecture will be used for the GLM-6 model to be released in October. GLM-5.3 may be updated to version 5.5 through post-training, but if GLM-6 is completed, it may skip version 5.5 and be released directly. Q: What are the reasons for the decline in Z.AI's gross margin in the first half of the year, and what are long-term gross margin targets? A: The decline in gross margin was primarily due to significant fluctuations in computing costs, frequent model iterations leading to user switching and inference learning curves, and initially low efficiency during the optimization phase of the domestic chip cluster. The company plans to enhance inference efficiency through collaborative optimizations across model, network, and chip levels, targeting to increase gross margin for the open platform to over 50% within the next 12-18 months. Q: What is Z.AI's computing power reserve and customer structure like? A: Z.AI has a cluster of 100,000 domestic chips and sufficient computing power to train models at the trillion parameter level. The top ten clients contribute over 40% of its revenue and ARR; nine of the top ten internet companies in China are Z.AI clients, four of which have chosen GLM models as their primary models nationwide. Q: How does Z.AI view the market space for AI and its future expansion directions? A: The company believes the global coding market is approximately $500 billion, with a still low penetration rate; the next step will be to expand from coding to coworking, representing a market opportunity ten times that of coding. In the long run, the TOMAS AI market could reach $30 trillion. The GLM-5.3 Flash has shown capabilities in cybersecurity, supporting expansion towards the Copilot direction. Q: What is the financing timeline for Z.AI, and what are market expectations? A: The quiet period for the last financing round ends on September 12, allowing for financing to start as early as September 13; the company has not specified a financing plan, but the market expects a possible financing round during the window period in mid-September. Q: What is the release timeline, performance, and iteration pace of Tencent's Harmony 4 Preview model? A: The Harmony 4 Preview was released at the end of August, ahead of the original expectations; the parameter size and context length have been significantly broadened, with internal blind test performance slightly higher than GLM-5.3 and Kimi K3. The iteration pace is accelerating: Harmony 3 Preview was released in April, and the official version followed in July, demonstrating improved model development efficiency. Q: What is the strategic positioning and key ecological construction focus of Tencent's WorkBody open platform? A: The WorkBody is positioned as an open Agent ecosystem that integrates hardware, applications, and developers: supporting continuous access of multiple hardware devices; the application side allows industry partners to build AI workbenches based on the platform; developers can contribute skills and participate in distribution. The company is rapidly enhancing payment and distribution capabilities, promoting its evolution from an office assistant to an Agent operational layer. Q: What are the investment implications of Tencents AI strategy and key observation points going forward? A: The Harmony 4 Preview has marginally alleviated market concerns regarding model capabilities and iteration speeds; the WorkBody and Harmony form a data flywheel. If Harmony becomes a SOTA model, it could develop into a Model as Service; even if not SOTA, the WorkBody ecosystem still holds commercial value. Future focus will be on the performance of Harmony models, WorkBody progress, WeChat AI, and new products and strategic releases at the global digital ecosystem conference in October. Q: What are the highlights of Meituans Q2 2026 performance and Q3 business guidance? A: Q2 performance exceeded expectations, primarily due to the recovery of in-store business gross margin to 30%; the improvement stems from competition with Douyin diverting low-ticket orders, and the company proactively deferring some marketing expenses to Q3. Q3 guidance appears weak: the expected drop in takeout user engagement to 0.1, with an in-store gross margin guidance of 25%, primarily due to seasonal impacts, increased rider subsidies, and lack of significant easing in the competitive environment. Q: How do analysts view Meituan's stock price, support levels, and recent catalyst judgments? A: Analysts have adjusted the target price from 120 to 110, noting that the stock price is oscillating between 70 and 90 yuan, with 70 yuan being a critical support level. Recent catalysts include the anticipated release of the Qianwen 4.0 update and CAPA-related guidance at the Silver Conference on September 22. Q: Regarding the performance of the Extra and F5.1 models, the domestic model competitive landscape, and payment scenario expansion, how do you view the risks of price wars and user payment willingness? A: Evaluation of the Extra and F5.1 models is still ongoing, making it difficult to conclude whether they underperform expectations. Model competition will stratify: companies with SOTA capabilities can efficiently distill high cost-performance versions, while companies relying solely on cost-performance models struggle to break through technical barriers and face gross margin pressure, making a purely price-driven competition unsustainable. After market misallocation is corrected, AI penetration rates will increase. Beyond AI coding, users in high knowledge-density industries such as finance, law, and healthcare express a clear willingness to pay, while monetization on the consumer side remains challenging, thus model companies strategic focus continues to shift towards productivity scenarios. Q: What are the core factors that buyers are currently most concerned about regarding AI companies? A: Buyers primarily focus on the capability level of models, followed by ARR growth and gross margin levels; as long as the model capabilities are recognized by the market, there is a high tolerance for valuations. Short-term stock price fluctuations are primarily influenced by expectations surrounding financing rhythms.