Morgan Stanley recently hosted a private meeting to conduct in-depth discussions on the latest developments among domestic AI large-model companies and major internet firms. According to the meeting minutes, both MINIMAX-W (00100) and Z.AI (02513) — two leading large-model companies — are sustaining rapid revenue growth. MINIMAX-W reached an ARR of $800 million by the end of August, and analysts have raised their year-end forecast to $1.3 billion. Z.AI surpassed $2 billion in monthly recurring revenue on a weekly basis in August, with the year-end guidance upgraded to $2.4 billion. Tencent's Hunyuan 4 Preview was released at the end of August, with internal blind-test performance slightly surpassing GLM-5.3 and Kimi K3, while the WorkBody open platform is evolving from an office assistant toward an Agent operations layer. Analysts believe buyers are currently most focused on model capability levels, followed by ARR growth and gross margins. Model competition will enter a tiered elimination phase, where companies with SOTA capabilities can distill cost-effective versions, while those with only cost-performance advantages will struggle to break through technical barriers and face margin pressure.
MINIMAX-W: Year-End ARR Expected to Exceed Guidance, Three New Models Await Release
MINIMAX-W recorded an ARR of $800 million by the end of August, maintaining its guidance of over $1 billion by year-end, though analysts have already 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. Its current computing power reserves are sufficient to support all training for the year and to pre-储备 capacity for next year's 10T-parameter model.
Z.AI: Rapid ARR Growth, Domestic Chip Cluster Supports Trillion-Parameter Training
Z.AI is experiencing rapid ARR growth, surpassing $2 billion on a weekly basis in August, with year-end guidance raised to $2.4 billion, though the actual figure depends on the delivery progress of domestic chips. The GLM-5.3 Flash released in August adopts a new architecture and runs inference on an all-domestic chip cluster; this architecture will be used for the GLM-6 scheduled for October release. Z.AI operates a 100,000-card domestic chip cluster, with its top 10 customers contributing over 40% of revenue, and 9 of China's top 10 internet companies are its clients.
Tencent Hunyuan: Faster Iteration, WorkBody Builds an Open Agent Ecosystem
Tencent Hunyuan 4 Preview was released at the end of August with significantly expanded parameter counts and context length, and internal blind-test performance slightly exceeds GLM-5.3 and Kimi K3. WorkBody is positioned as an open Agent ecosystem connecting hardware, applications, and developers, with payment and distribution capabilities being actively accelerated. If Hunyuan becomes a SOTA model, it can develop Model as a Service; even if it falls short of SOTA, the WorkBody ecosystem retains commercial value. Key focus points ahead include new product launches and strategic announcements at the October Global Digital Ecosystem Conference.
Q&A Session Highlights
Q: What is the ARR scale of MINIMAX-W by the end of August, and what is the statistical methodology? A: As of the end of August, MINIMAX-W's ARR reached $800 million, using a weekly statistical basis consistent with the $400+ million disclosed in May. This figure exceeded market expectations.
Q: What is MINIMAX-W's year-end ARR guidance, and what are market expectations versus company explanations? A: The company maintains its year-end ARR guidance of over $1 billion, though this is conservative; analysts have raised expectations to $1.3 billion. The company has not raised its guidance because it is awaiting actual performance following the release of new models in the second half, leaving room for upward adjustment.
Q: What new models does MINIMAX-W plan to release in the second half, and what are the timelines and core features? A: MINIMAX-W plans to release three new models: M3.1, H3.1, and M3 Pro. The company's current computing power reserves are sufficient to support all model training for the year and to pre-reserve capacity for next year's 10T-parameter model.
Q: What were MINIMAX-W's R&D expenses in H1, and what are the expectations for H2 and next year? A: R&D expenses were approximately $300 million in H1; annualized R&D expenses for H2 are expected to exceed $1 billion; next year's R&D spending will increase further but at a significantly slower pace than revenue and gross profit growth.
Q: What are the main reasons for the H1 gross margin decline at MINIMAX-W, and what is the H2 outlook? A: The gross margin decline was primarily due to: price cuts and user subsidies following the poor initial market reception of the M3 launch; the optimization ramp-up period for new model inference; low gross margins and generous discounts on Token plans; and an increased share of text model revenue, while multimodal models have gross margins above 50%. The company indicated most H1 factors are one-time in nature and expects sequential gross margin improvement in H2.
Q: How does MINIMAX-W view the relationship between high-intelligence models and cost-effective models? A: MINIMAX-W believes intelligence and cost-effectiveness are two sides of the same coin: high inference efficiency and low cost enable larger-scale post-training, which enhances model intelligence. Inference efficiency is therefore a core capability that cannot be ignored in the pursuit of high intelligence.
Q: What is the timing window for MINIMAX-W's next financing round and the company's plans? A: The cooling-off period for the last financing round ends on September 12, and the company could theoretically launch financing as early as September 13. However, management indicated it does not consider financing in the short term and plans to proceed after releasing an industry-leading SOTA model. Market consensus expects the financing timing to fall after October.
Q: What are Z.AI's ARR figures across different time points, the statistical methodology, and the year-end guidance with market feedback? A: Z.AI's ARR data: $250 million in March, $530–540 million in June, $1 billion in July, and $1.6 billion in August; on a weekly basis, August ARR exceeded $2 billion. Year-end guidance has been raised to $2.4 billion, which the company considers conservative, with the actual figure depending on computing power delivery. The company expects to update guidance in Q4 based on domestic chip delivery status.
Q: What are the key model pipeline updates for Z.AI in H2, and what is the positioning of GLM-6? A: The GLM-5.3 Flash released in August adopts a new architecture and runs inference on an all-domestic chip cluster; this architecture will be used for the GLM-6 slated for October. GLM-5.3 may be updated to version 5.5 through post-training, but if GLM-6 training is completed, the company may skip 5.5 and release directly.
Q: What caused the H1 gross margin decline at Z.AI, and what is the long-term gross margin target? A: The gross margin decline was primarily due to: significant fluctuations in computing power costs; user churn and inference ramp-up from frequent model iterations; and low efficiency during the initial optimization phase of the domestically-produced chip cluster. The company plans to improve inference efficiency through coordinated optimization across the model layer, network layer, and chip layer, targeting an open platform gross margin above 50% within the next 12–18 months.
Q: What is Z.AI's computing power reserve and customer structure? A: Z.AI operates a 100,000-card domestic chip cluster and has reserved advanced computing power sufficient to train trillion-parameter-level models. The top 10 customers contribute over 40% of revenue and ARR; 9 of China's top 10 internet companies are Z.AI customers, with 4 of them using GLM as their nationwide preferred primary model.
Q: How does Z.AI view the AI market space and future expansion directions? A: The company believes the global coding market is approximately $500 billion, with current penetration still low; the next step is expanding from coding to cowork, a market 10 times the size of coding; in the long term, the TOMAS AI market reaches $30 trillion. GLM-5.3 Flash has already demonstrated cybersecurity capabilities, supporting expansion toward the Copilot direction.
Q: What is Z.AI's financing timing window and market expectations? A: The cooling-off period for the last financing round ends on September 12, with the earliest possible financing start on September 13; the company has not specified its financing plans, and market expectations suggest a potential financing window in mid-September.
Q: What is the release timing, performance, and iteration pace of Tencent Hunyuan 4 Preview? A: Hunyuan 4 Preview was released at the end of August, earlier than originally expected; parameter counts and context length have significantly expanded, with internal blind-test performance slightly exceeding GLM-5.3 and Kimi K3. Iteration speed has accelerated: Hunyuan 3 Preview was released in April with the official version in July, demonstrating improved model R&D efficiency.
Q: What is the strategic positioning of Tencent's WorkBody open platform and the key focus of ecosystem building? A: WorkBody is positioned as an open Agent ecosystem connecting hardware, applications, and developers: supporting continuous multi-device hardware access; allowing industry partners to build AI workbenches on the application side; and enabling developers to contribute skills and participate in distribution. The company is accelerating the development of payment and distribution capabilities to drive its evolution from an office assistant toward an Agent operations layer.
Q: What are the investment implications of Tencent's AI strategy and key observation points going forward? A: Hunyuan 4 Preview marginally alleviates market concerns about model capability and iteration speed; WorkBody and Hunyuan form a data flywheel. If Hunyuan becomes a SOTA model, it can develop Model as a Service; if not, the WorkBody ecosystem still holds commercial value. Key focus points ahead include Hunyuan model performance, WorkBody progress, WeChat AI, and new product launches and strategic announcements at the October Global Digital Ecosystem Conference.
Q: What are the H1 2026 performance highlights for Meituan and the Q3 business guidance? A: Q2 results beat expectations, primarily due to the recovery of in-store business gross margin to 30%; the improvement stems from Douyin competition taking away low-ticket orders, and the company proactively deferring some marketing expenses to Q3. Q3 guidance is relatively weak: food delivery UE is expected to decline sequentially to 0.1, and in-store gross margin guidance is 25%, mainly due to seasonality, increased rider subsidies, and no significant easing in the competitive environment.
Q: What are analysts' views on Meituan's share price, support levels, and near-term catalysts? A: Analysts have lowered the target price from 120 to 110, viewing the stock as trading in a 70–90 yuan range, with 70 yuan as a key support level. Near-term catalysts include the September 22 banking conference, with expectations of a Qwen 4.0 update and CAPA-related guidance.
Q: Regarding the performance of Extra and F5.1 models, the domestic model competitive landscape, and paid scenario expansion — how do you view the risk of price wars and user willingness to pay? A: Evaluations of Extra and F5.1 are still underway, making it premature to determine whether they underperformed. Model competition will become tiered: companies with SOTA capabilities can efficiently distill cost-effective versions, while those with only cost-performance models struggle to break through technical barriers and face margin pressure — pure price wars are unsustainable. After market misallocations are corrected, AI penetration will increase. Beyond AI coding, users in high-knowledge-density industries such as finance, law, and healthcare show clear willingness to pay, while consumer-side monetization is more difficult, so model companies continue to shift strategic focus toward productivity scenarios.
Q: What are the core factors buyers currently prioritize for AI companies? A: Buyers prioritize model capability levels first, followed by ARR growth and gross margin levels; as long as model capability gains market recognition, valuation tolerance is relatively high. Short-term share price volatility is mainly driven by financing pace expectations.
Comments