Chinese AI models have reached the intelligence inflection point for global diffusion

Overview

The report focuses on the global diffusion of Chinese AI models, with the central thesis that Chinese open-source/open-weight models are reaching a critical intelligence point to compete with globally leading proprietary models. It highlights how Chinese models achieve high performance under compute constraints through smaller parameter scales, MoE, sparse attention, post-training, and real code-data feedback; why the open-source or open-weight path is being chosen; how domestic and offshore addressable markets are evolving; and which companies are most likely to become long-term winners.

Core views

Goldman believes the Chinese AI model market is forming a two-layer structure: high-performance models gain pricing power through model intelligence and listing speed, while low-cost agent models expand price-sensitive enterprise and SMB demand through very low token pricing. The report favors players with larger ARR scale, gross margin advantage, and financial strength; in core text models it highlights Zhipu and DeepSeek, in multimodal it highlights ByteDance, and it sees MiniMax as having a favorable cost-efficiency and risk-return profile.

Analysis framework

The report uses a competitive positioning framework, breaking AI model companies into three quantifiable dimensions: pricing power, cost advantage, and financial strength, and layering this with token scale, market share, ARR, gross margin, cash position, and valuation multiples. It also compares major Chinese LLM players on model performance, pricing tiers, API efficiency, commercialization progress, capital expenditure, and overseas diffusion potential.

Methodology notes

• Competitive positioningChinese AI model competitive positioning framework

  Pricing power, cost advantage, financial strength

  The report uses model release timing, Arena performance in real use cases, price level, token scale, throughput/cache hit rate, parameter-to-activated ratio, inference gross margin, and cash and valuation multiples to judge long-term winners.

• Commercialization analysisARR maximization quadrants

  Combination of token scale and pricing level

  The report argues that AI model revenue upside comes from the combination of token usage and price, with both high-performance premium models and low-cost high-volume agent models potentially forming attractive ARR quadrants.

• Technical efficiencyMoE and sparse activation

  Improving inference efficiency with lower activated parameters

  MoE dynamically selects a small number of expert networks to process input, reducing activated parameters and inference cost while maintaining high overall model capacity.

• Demand measurementShifting from token-maxxing to ROI-first

  Shifting from pure token consumption to task cost and actual output

  The report expects enterprises to focus more on overall cost per task, daily active agents, backend process automation, and realized output, rather than pursuing token burn alone.

Asset mapping & comparison

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

• Zhipu / Knowledge Atlas

  Leading text model player

  Strengths

  Strong model capability in coding and agent scenarios, and listed as one of the strongest positioned players in text foundation models; GLM5.2 is associated with high intelligence levels and stronger pricing power.

  Weaknesses

  Needs to continue demonstrating commercialization scale, financial strength, and durable long-term gross margin advantage.

  Comparison

  Has higher pricing power than lower-end models; positioned alongside DeepSeek as a stronger text model player.

  Risks

  Overseas market access, advanced model access restrictions, training compute constraints, and fast follower competition.

• DeepSeek

  Leading player in base models and cost efficiency

  Strengths

  Supports scalable token usage with very high cost efficiency; V4 series and DSpark improve inference speed; listed as one of the strongest positioned players in text foundation models.

  Weaknesses

  Peak and off-peak pricing indicates compute tightness, and margins in lower pricing tiers may still come under pressure.

  Comparison

  Co-leads text foundation models with Zhipu; more text/foundation-model focused than ByteDance.

  Risks

  Advanced model access restrictions, compute supply, price competition, and policy uncertainty.

• ByteDance

  Multimodal leader

  Strengths

  Seedance performs strongly in multimodal/video generation models; the report says Seedance's ARR run-rate is above US$2bn with healthy gross margins; by 2030E, it is expected to contribute with DeepSeek to a majority of Chinese AI model token share.

  Weaknesses

  Seed models follow a closed-source route, and international expansion may face stronger regulatory and data-security scrutiny.

  Comparison

  Leads in multimodal capability, distinct from the text foundation-model strengths of Zhipu and DeepSeek.

  Risks

  Western market regulation, data localization, compute jurisdictional restrictions, and geopolitical policy risk.

• MiniMax

  Beneficiary in multimodal and low-cost models

  Strengths

  The M3 model sits in an attractive ARR quadrant of low price and high token volume; the report highlights its cost efficiency and valuation discount, and references a Buy rating.

  Weaknesses

  Its overall competitive positioning still depends on improving pricing power and financial strength.

  Comparison

  Lower-priced than top-tier text models, and may scale through overseas revenue and multimodal product expansion.

  Risks

  Low-end API price wars, pace of future M3 upgrades, performance of H3 video models, and capital strength.

• Alibaba Qwen

  Full-stack AI and cloud-ecosystem player

  Strengths

  Has cloud infrastructure, a model family, and commercialization scenarios; some high-performance Qwen models have relatively high pricing.

  Weaknesses

  Some largest and highest-performing models are shifting to closed-source to strengthen monetization, which may reduce open-source diffusion effects.

  Comparison

  Has cloud and ecosystem advantages as a large platform-integrated model player; belongs to the mega-cap cohort alongside Tencent and other large ecosystems.

  Risks

  Capital expenditure payback, cloud-to-model revenue conversion, model competition, and pricing pressure.

• Tencent Hunyuan / WorkBuddy

  Platform AI and agent application player

  Strengths

  Has WeChat ecosystem and WorkBuddy enterprise agent entry points, with potential for diffusion from consumer to enterprise use.

  Weaknesses

  The report summary gives weaker coverage of its text model leadership versus Zhipu, DeepSeek, and ByteDance.

  Comparison

  Stronger platform and traffic entry compared with standalone model companies; less evidence of multimodal superiority than ByteDance.

  Risks

  Model iteration speed, conversion of applications, intensifying competition, and regulatory environment.

Key data

• China AI model token growth outlook25X by 2030EThe report estimates that token output from Chinese AI model companies will be 25 times current levels by 2030.
• Overseas token share outlook5% of global tokens by 2030EThe report estimates Chinese AI model companies will generate 5% of global tokens overseas by 2030.
• China AI API and subscription revenue poolUS$125bn in 2030EThe report estimates the China AI API plus subscription revenue pool reaches US$125bn in 2030E.
• 2030E token concentrationByteDance and DeepSeek 80%The report expects ByteDance and DeepSeek to account for 80% of token output from Chinese AI model companies.
• Top-end Chinese model pricingaround US$1 per blended 1M tokensHigh-performance models such as Zhipu GLM5.2 and Alibaba Qwen3.7 Max are around US$1 per blended million tokens.
• Low-end agent model pricingUS$0.06-0.2 per blended 1M tokensLow-end agent model prices are significantly below high-end models, used to expand price-sensitive demand.
• DeepSeek V4 Pro/Flash peak priceUS$0.35 / US$0.12 per 1M tokensThe report notes that after DeepSeek introduced peak-offpeak pricing, peak time is double off-peak, with an implied blended price at this level.
• Chinese model parameter scale200bn to 1.6T parametersThe report says Chinese models mostly range from 200 billion to 1.6 trillion parameters, below some global SOTA models.
• Activated parameter ratio3-5% activated parametersArchitectures such as MoE and sparse attention keep the activated parameter share far below total parameters.
• ByteDance Seedance ARR run-rateUS$2bn+ ARR run-rateThe report cites reporting that Seedance's latest ARR run-rate is above US$2bn, with gross margins around 70%.

Impact & implications

If the report's thesis is correct, Chinese AI models could expand domestic and global adoption through open-source/open-weight models, low-cost inference, and localized compute stacks, potentially reshaping global AI model pricing curves and enterprise adoption patterns. For investment implications, the market may increasingly focus on AI model company ARR scale, quality of gross margins, inference efficiency, closed-loop real-world usage data, and balance sheet strength rather than model leaderboard standings alone.

Risks

• Tighter control over foreign access to China’s frontier AI models.
• Western market policy, data-security, and market-entry restrictions.
• Limited access to top-tier training compute and leased compute capacity.
• Price wars in low-end agent-model APIs compressing gross margins.
• Anti-distillation, entity-list rules, or further designated regulatory restrictions affecting international diffusion.
• SLM or new AI architectures could weaken the current commercial value of the large-model path.
• Shifts toward open-weight or commercial licensing terms could slow adoption.
• Compute tightness could affect service stability, inference costs, and customer experience.

What to watch

• Whether Chinese model players launch 2-5 trillion parameter models in 2H26.
• Code, agent, and multimodal benchmark performance of Zhipu GLM, DeepSeek V4, MiniMax M3/H3 and other models.
• Demand elasticity, service stability, and margin changes after DeepSeek's peak-offpeak pricing.
• ARR growth and gross margin evolution of video-generation models such as ByteDance Seedance, Kuaishou Kling, and MiniMax Hailuo/H3.
• Whether domestic ASIC supply and local training clusters can support larger model training.
• Whether enterprises shift from token burn metrics to daily active agents, task cost, and realized output.
• Adoption pace and regulatory boundaries for Chinese models in overseas, especially non-US, markets.
• Whether open-source models further shift toward open-weight and commercial licensing models.