GLM-5.3 capability upgrade and DeepSeek price increases reinforce profitability expectations for China's large-model industry
AI summary card
GLM-5.3 capability upgrade and DeepSeek price increases reinforce profitability expectations for China's large-model industry
Morgan Stanley believes GLM-5.3 significantly improves coding and cybersecurity capabilities through scaled post-training, while DeepSeek V4 price increases could ease price competition; it remains positive on Z.AI.
- GLM-5.3 retains the GLM-5.2 base model and achieves capability improvements through longer and richer reinforcement-learning post-training.
- Internal evaluations indicate GLM-5.3's coding capability is 50% higher than GLM-5.2, with added advantages in white-box code auditing and vulnerability discovery.
- Both DeepSeek V4 Pro and V4 Flash raised prices; the report interprets this as a positive development for China's large-model industry.
- The report considers user feedback on V4 Pro relatively weak in long-horizon tasks, coding capability, and stability.
Report interpretation
Overview
This report tracks recent developments in China's frontier AI foundation models, focusing on Z.AI's launch of GLM-5.3 and DeepSeek's formal release of the V4 series alongside price increases. The core view is that model competition is increasingly emphasizing intelligence capabilities and product value rather than purely low-price competition; against this backdrop, Morgan Stanley maintains its positive view on Z.AI.
Core views
GLM-5.3's capability progress primarily comes from scaling post-training on the same base model, demonstrating that existing model foundations still have room to unlock further intelligence. Feedback on DeepSeek V4 Pro is relatively negative regarding long-horizon tasks, coding, and stability, while its price increases mean it is no longer China's cheapest flagship large model. The report believes the price increases will help improve the competitive environment and provide positive support for Z.AI.
Analysis framework
The report assesses the competitive landscape using product release information, internal evaluations, user feedback, and cross-sectional comparisons of China's frontier large models; it values Z.AI using discounted cash flow and cross-checks against the 2027 price-to-sales multiple implied by the target price.
Methodology notes
Estimates enterprise value by discounting future cash flows
The report uses DCF valuation for Z.AI, assuming a weighted average cost of capital of 15% and a terminal growth rate of 3%.
Measures valuation as market capitalization relative to sales revenue
The report states that the target price corresponds to 42x 2027 P/S.
Compares model capabilities, user feedback, and pricing
Compares the GLM and DeepSeek V4 series across coding, cybersecurity, long-horizon tasks, stability, and API price changes.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Z.AI CO., LTD. (2513.HK)Core covered company and direct beneficiary
- Strengths
- GLM-5.3 has upgraded coding and cybersecurity capabilities; scaled post-training indicates that the existing base model still has room for capability improvement; benefits from easing industry competition and potential global expansion.
- Weaknesses
- Model capabilities still require ongoing validation, and the company depends on compute resources and commercialization execution.
- Comparison
- Compared with DeepSeek V4 Pro, the report notes weaker feedback for the latter on long-horizon tasks, coding, and stability; after its price increases, DeepSeek is also no longer the cheapest Chinese flagship model.
- Risks
- Compute constraints, model capabilities lagging peers, geopolitical risks, and a renewed deterioration in the competitive landscape.
Key data
- GLM-5.3 parameter scaleMore than 700 billion total parametersUses the same base model as GLM-5.2.
- GLM-5.3 coding capability improvement50% higher than GLM-5.2Internal evaluation result cited in the report.
- DeepSeek V4 Pro price adjustmentNon-cached input +1.25x; cached input +8x; output +2.38xBlended off-peak and peak-period pricing.
- DeepSeek V4 Flash price adjustmentNon-cached input +1.25x; cached input +2.75x; output +2.38xBlended off-peak and peak-period pricing.
- DCF assumptionsWACC 15%; terminal growth rate 3%Used for Z.AI valuation.
- Valuation implied by target price42x 2027 P/SDisclosed in the report.
Impact & implications
If large-model vendors can support higher pricing through capability improvements, industry competition may shift from price wars toward performance, stability, and commercialization efficiency. For Z.AI, GLM-5.3's improved coding and cybersecurity capabilities, potential open-weight release, and easing industry competition are all positive catalysts; however, product performance and compute supply remain key constraints to realizing the valuation.
Risks
- Compute constraints may limit model training, inference, and commercialization expansion.
- If model performance lags peers, product competitiveness and customer adoption may come under pressure.
- Geopolitical risks may affect chips, cloud resources, and overseas expansion.
- If industry price competition intensifies again, it may compress commercialization returns for large models.
- Valuation is sensitive to assumptions of a 15% WACC, 3% terminal growth rate, and future revenue growth.
What to watch
- Whether GLM-5.3 open weights will be released as planned within two weeks after security hardening.
- GLM-5.3's actual performance in third-party benchmarks, enterprise customer use cases, and developer scenarios.
- Customer retention, demand elasticity, and competitors' responses following DeepSeek V4's price increases.
- Z.AI's partnerships with overseas cloud service providers and progress in globalization.
- Changes in compute supply, model iteration speed, and industry pricing discipline.