China's large model competition is shifting from a price war to a flywheel of intelligence, compute, and commercialization

Overview

The report refutes the market view that China's open-weight models will inevitably lead to commoditization and a prolonged price war. Morgan Stanley believes China's large model industry is shifting toward a monetization phase driven by model intelligence: pricing is becoming more rational, open-weight licensing is gradually strengthening commercial constraints, and model scale is expanding rapidly. Continuous training of frontier models requires more capital, compute, infrastructure optimization, and engineering execution capabilities, so industry barriers will rise and well-resourced leading vendors will have greater advantages.

Core views

Model intelligence, rather than low prices, is the long-term competitive moat. Frontier vendors can launch lower-cost models downstream, but ordinary model vendors will find it difficult to move up to SOTA levels. Stronger models drive more adoption and monetization, which in turn secure more financing and compute investment and support the next round of model upgrades; commercialization and the pursuit of AGI are not in conflict, but mutually reinforcing. The report is positive on Z.ai's faster formation of this flywheel, remains constructive on MiniMax, and favors Alibaba's full-stack AI and cloud computing advantages.

Analysis framework

The report compares the industry by combining parameter scale, intelligence indices, API prices, licensing terms, product release plans, compute supply, and financing conditions of China's frontier models; at the company level, it evaluates commercialization progress using ARR, revenue, and R&D expense forecasts, and validates target prices using DCF and 2027 P/S multiples.

Methodology notes

• Competition & strategyIntelligence-commercialization flywheel

  Stronger models drive adoption and monetization, with revenue and financing then converted into training and compute investment.

  The report argues that the ability to sustain this cycle is more decisive for long-term competitive positioning than any single model leaderboard ranking.

• Industry structureHigh-end and low-end market segmentation

  The high-end market is differentiated by frontier intelligence, while model capabilities in the low-end market converge and price competition is more intense.

  Frontier vendors can more easily launch lower-cost, lower-capability versions to cover the mass market, while ordinary vendors need to cross higher funding and technical thresholds to upgrade to frontier levels.

• Demand analysisJevons paradox

  Declining unit inference costs may stimulate usage growth rather than reduce overall demand for compute.

  Cheaper and more efficient open-weight models are expected to accelerate AI adoption and increase enterprise demand for deployment, orchestration, governance, security, and infrastructure.

• Valuation methodsDiscounted cash flow model

  Assesses a company's equity value by discounting future free cash flows.

  Both Z.ai and MiniMax are valued using DCF, assuming a WACC of 15% and a perpetual growth rate of 3%; target price changes mainly come from adjustments to financial forecasts.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Z.AI CO., LTD. (2513.HK)

  Core favored stock, target price raised

  Strengths

  GLM-5.2 has reached a globally leading level, access to overseas compute has improved, financing scale is sufficient, and the flywheel of model capability and commercialization is forming relatively quickly.

  Weaknesses

  Inference compute, especially overseas compute, may still constrain ARR growth, and continued training of large-parameter models will significantly increase R&D expenses.

  Comparison

  Compared with MiniMax, its near-term ARR growth path is clearer and commercialization delivery is earlier, but its market valuation is also higher.

  Risks

  Global competitors launching models with stronger performance and lower prices, U.S. policy restrictions, a decline in model rankings, and compute expansion falling short of expectations.

• MiniMax (0100.HK)

  Constructive view maintained, but target price lowered

  Strengths

  The M3 architecture has advantages in long context and inference efficiency, M3 Pro has potential for a large-parameter upgrade, H3 performs strongly in video multimodality, and financing can support training and compute expansion.

  Weaknesses

  M3 initially encountered negative feedback on coding capability and pricing; the timing of model releases makes its 2026 ARR contribution more back-end loaded, and loss forecasts have widened.

  Comparison

  Compared with Z.ai, near-term commercialization growth is more back-end loaded, but it has differentiated potential in multimodal, video models, and vertical industry data.

  Risks

  M3 Pro failing to reach SOTA levels, insufficient developer adoption, competitor price cuts, shareholder lock-up expiry pressure, and continued high R&D investment.

• Alibaba Group Holding (BABA.N)

  Beneficiary of the AI industry chain

  Strengths

  Has full-stack AI capabilities spanning models, cloud computing, and compute, with infrastructure scale and commercial customer advantages.

  Weaknesses

  Training and inference for large open-weight models require continued capital investment, and licensing commercialization is still evolving.

  Comparison

  Compared with pure model vendors, Alibaba can monetize through synergy between cloud infrastructure and model services, and has broader funding and customer resources.

  Risks

  Intensifying global model competition, capex returns below expectations, regulatory and export restrictions, and pressure on cloud business margins.

Key data

• Parameter scale of China's frontier modelsKimi K3: 2.8T; Qwen3.8-Max: 2.4TThe report expects the parameter scale baseline for China's frontier models to gradually rise to 2T to 3T in 2H26.
• Enterprise adoption rate of open models63%Surveyed enterprises use both open models and closed-source models, supporting the view that multiple model types will coexist.
• Z.ai 2026 ARR forecastUS$2bnRaised from US$1bn; bull and bear case scenarios are US$3bn and US$1.5bn, respectively.
• Z.ai target priceHK$1,700Raised by about 72% from HK$990, corresponding to approximately 42x 2027 P/S.
• Z.ai financingHK$31.4bnH-share placement with 4.25% dilution; plans to use 55% of proceeds for model training and compute expansion.
• MiniMax 2026 ARR forecastUS$1bnOriginal forecast maintained; bull and bear case scenarios are US$1.5bn and US$500mn, respectively.
• MiniMax target priceHK$900Lowered 18% from HK$1,100, corresponding to approximately 32x 2027 P/S.
• MiniMax financing planHK$16.0bnIncludes HK$9.5bn equity placement and HK$6.5bn convertible bonds; plans to use 80% of proceeds for model training and inference compute expansion.
• MiniMax M3 Pro scale2.7T parametersExpected to be released from September to October 2026, and is an important catalyst for model performance and pricing power.
• Core DCF assumptionsWACC 15%; perpetual growth rate 3%DCF assumptions for both companies remain unchanged.

Impact & implications

Improved industry pricing discipline and licensing commercialization will expand the ARR pool and reduce constraints from relying solely on proprietary compute to sell APIs. Rising model scale will increase demand for training, inference optimization, financing, and compute procurement, driving market share concentration toward leading vendors. For investors, greater attention should be paid to model iteration capabilities, compute access, financing channels, and monetization efficiency, rather than single leaderboard rankings or the lowest API quotes. Cloud service providers, compute infrastructure, and platforms for model routing, orchestration, observability, governance, and security may also benefit from multi-model adoption.

Risks

• Global competitors launch models with stronger performance and lower prices, reigniting price competition and compressing valuation multiples.
• Training, inference, and infrastructure costs for large-parameter models exceed expectations, leading to expanded R&D expenses and operating losses.
• Delayed model releases, performance failing to reach SOTA levels, or poor developer feedback could weaken adoption and monetization.
• Restricted overseas compute supply or tighter U.S. policies could affect the global expansion of Chinese model vendors.
• Stricter open-weight licensing could slow ecosystem adoption, and the effectiveness of enforcing commercial terms is also uncertain.
• Equity financing, convertible bonds, and potential new listings may bring shareholder dilution and episodic selling pressure.
• Excessive market reactions to changes in single model rankings could cause sharp volatility in valuations and prices of related stocks.

What to watch

• API pricing adjustments and gross margin changes by other Chinese model vendors after DeepSeek's price increase.
• Performance and adoption of Z.ai's next-generation large-parameter GLM model after its expected release in October 2026.
• Progress in Z.ai's cooperation with AWS and other overseas cloud service providers, and expansion of overseas inference compute.
• MiniMax M3's coding capability upgrades and M3 Pro's release performance from September to October 2026.
• The effect of MiniMax H3 in driving ARR growth from multimodal and video models.
• Whether Alibaba's future Qwen open-weight models introduce large-scale user revenue-sharing terms.
• Whether China's frontier models generally move into the 2T to 3T parameter range, and the progress of ByteDance's potential 5T-plus model.
• Whether the two companies' 2027 R&D expenses exceed US$1bn and the impact on cash flow and financing needs.
• Policy changes between China and the U.S. regarding model exports, overseas access, and advanced compute supply.