Kimi K3 pushes Chinese AI model competition into a new stage from cost efficiency toward intelligence and pricing power

Overview

This report focuses on the competitive landscape of Chinese AI models, with the release of the Kimi K3 open-weight model as the central event. The report believes that Chinese AI models are transitioning from the cost-efficiency phase represented by DeepSeek, to the intelligence-improvement phase of models such as GLM, and further into the stage of pricing power for high-end models as reflected by Kimi K3. Kimi K3's parameter scale, coding capability, and API pricing together indicate that the intelligence level and commercialization potential of leading Chinese models are improving, although the competitive landscape remains highly dynamic.

Core views

The report's core views include: first, Kimi K3 has reached a new frontier level in coding and some agentic capabilities, bringing Chinese open-weight models close to the tipping point for global diffusion; second, competition in high-end coding models will intensify in 2H26, with model parameter size potentially expanding to 2-5 trillion; third, lower-end agentic text models still face pressure on API pricing and gross margins; fourth, video generation models have a better commercialization environment than basic text models due to global demand, feature breakthroughs, and tight computing supply; fifth, cloud and data centers are the most preferred sub-sector within China Internet, benefiting from increased capital expenditure by AI hyperscalers.

Analysis framework

The report assesses the competitive positioning of Chinese AI model vendors by combining model rankings, the Artificial Analysis Intelligence Index, API pricing, parameter scale, cost efficiency, cash reserves, commercialization capability, model release schedules, and industry competitive structure, and maps this to related stocks including MiniMax, Knowledge Atlas, Alibaba, Tencent, Kuaishou, GDS, VNET, and Kingsoft Cloud.

Methodology notes

• Competitive positioningCompetitive Positioning framework

  Assessment of competitive positioning of AI model companies

  Goldman Sachs uses this framework to compare independent AI model companies in model intelligence, pricing power, cost efficiency, financial strength, and commercialization capability, and the report believes independent AI model companies still stand out under this framework.

• Valuation methodsDCF valuation

  12-month target price methodology

  The target prices for both MiniMax and Knowledge Atlas are based on DCF valuation, and the report discloses the use of a 12% WACC and a 2% terminal growth rate.

• Stock attribute analysisGS Factor Profile

  Growth, Financial Returns, Multiple, and Integrated factors

  GS Factor Profile provides an investment context for stocks relative to the market and industry peers through growth, financial returns, valuation multiples, and integrated percentiles.

Asset mapping & comparison

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

• MiniMax Group (0100.HK)

  Beneficiary in AI models and video generation

  Strengths

  Buy-rated; financial strength strengthened after fundraising; Hailuo and the upcoming H3 video generation model are in a more favorable industry landscape; M3 updates and M3 Pro are expected to improve coding intelligence.

  Weaknesses

  Competition in basic text models is intense, short-term profit visibility is limited, and commercialization still needs validation.

  Comparison

  Compared with lower-end text models, the video generation track has healthier pricing and gross margins; compared with Kimi K3, the report focuses on the intelligence improvements of MiniMax's subsequent M3/M3 Pro.

  Risks

  Model performance below expectations, monetization path slower than expected, commercialization capability weaker than expected, IP and content generation risks, cash burn and self-funding capability, and geopolitical risks arising from US-China technology competition.

• Knowledge Atlas Technology (2513.HK)

  Representative company in enterprise and coding AI models

  Strengths

  The report says it is one of China's top enterprise and coding AI models; upside risk would emerge if model intelligence, profit visibility, non-inference take-rate revenue, and commercialization exceed expectations.

  Weaknesses

  The report assigns a Neutral rating; the stock price reaction on July 17 was -28%, reflecting market concerns about competition in Chinese AI models and uncertainty over long-term winners.

  Comparison

  After the launch of Kimi K3, competition in high-end coding models is more intense, and Knowledge Atlas needs to prove its ability to maintain leadership.

  Risks

  Competition in the global foundation model industry, limited near-term profit visibility due to high R&D expenses, cash burn and self-funding capability, and geopolitical risks under US-China technology competition.

• Alibaba Group (BABA / 9988.HK)

  Beneficiary of the cloud and large-model ecosystem

  Strengths

  Buy-rated; the report highlights Alibaba among preferred sub-sectors in cloud and data centers, and expects AI hyperscaler capex to increase in 2H26.

  Weaknesses

  E-commerce GMV, retail monetization, and cloud revenue growth are still affected by macro conditions and competition.

  Comparison

  Compared with pure model companies, Alibaba benefits from both cloud infrastructure and the Qwen model ecosystem.

  Risks

  GMV growth below expectations, slower-than-expected monetization in China retail, weaker-than-expected execution of strategic investments, and slower cloud revenue growth.

• Tencent Holdings (0700.HK)

  Internet platform and AI-related name

  Strengths

  Buy-rated; Workbuddy is listed as an example entry point for coding and agentic applications, and Tencent also has platform and cloud resources.

  Weaknesses

  If AI progress is slower than expected, it may affect the related valuation narrative.

  Comparison

  Compared with independent model companies, Tencent is more exposed to platform-based and application-based AI opportunities.

  Risks

  Intensifying performance advertising competition, delays in game launches or license approvals, slower growth in FinTech and Cloud, slower-than-expected AI development, and reinvestment risk.

• Kuaishou Technology

  Beneficiary in video generation models

  Strengths

  Buy-rated; Kling is listed as a representative video generation model, with industry pricing and gross margins better than basic text models.

  Weaknesses

  The commercialization pace of combining video generation with LLMs still requires continued validation.

  Comparison

  Compared with lower-end text models, video generation has tighter supply-demand dynamics and healthier pricing.

  Risks

  New feature breakthroughs slower than expected, computing supply constraints, and weaker-than-expected global adoption.

• GDS Holdings、VNET Group、Kingsoft Cloud

  Beneficiaries of AI computing demand in cloud and data centers

  Strengths

  The report lists Cloud & Data Centers as the most preferred sub-sector in China Internet, and specifically mentions these names as beneficiaries of increased AI hyperscaler capex.

  Weaknesses

  Earnings depend on the realization of capital expenditure, customer demand, and delivery of computing infrastructure.

  Comparison

  Compared with model companies, cloud and data centers are more directly exposed to AI computing demand.

  Risks

  AI capex below expectations, data center supply and cost pressure, and changes in customer concentration and the financing environment.

Key data

• Kimi K3 parameter scale2.8 trillion parametersThe report states that this open-weight model was released on 2026-07-17.
• Kimi K3 blended API priceUS$2.3 / 1 million tokensHigher than Qwen3.7 Max at US$1.4, Zhipu GLM5.2 at US$0.9, MiniMax M3 at US$0.22, and DeepSeek V4 Pro at US$0.18, but still below global SOTA models.
• Expected new model parameter scale in 2H262-5 trillion parametersThe report expects Chinese AI models such as Zhipu GLM, Alibaba Qwen, and MiniMax M3 Pro to continue upgrading in 2H26.
• Lower-end agentic model price rangearound US$0.1-0.2 / 1 million tokensThe report expects API pricing and gross margins in this range to remain under pressure in the second half of the year.
• ByteDance SeeDanceReportedly around 70% gross margin, latest US$2bn ARR run-rateThe report views video generation as a track with healthier pricing and gross margins than basic text models.
• MiniMax rating and target priceBuy, 12-month target price HK$860Valuation is based on DCF, with 12% WACC and 2% terminal growth rate.
• Knowledge Atlas Technology rating and target priceNeutral, 12-month target price HK$1,880The report states coverage was recently initiated at a US$110bn valuation.
• Tencent Holdings target priceBuy, 12-month SOTP target price HK$700Risks include advertising competition, delays in game approvals, slower FinTech and Cloud growth, slower-than-expected AI progress, and reinvestment risk.
• Alibaba Group target priceBABA US$186 / 9988.HK HK$18012-month SOTP target price, based on FY27E.

Impact & implications

From an investment perspective, if high-end models can improve intelligence while maintaining pricing power, they may improve the long-term revenue quality of Chinese AI model companies; however, price wars in lower-end text models will pressure gross margins, making financial strength critical to sustaining investment. The report prefers the cloud and data center chain that benefits from AI computing demand, while also believing that MiniMax's H3 video generation, M3 updates, and M3 Pro progress may become stock catalysts.

Risks

• Competition in high-end coding models among Chinese AI model companies intensifies further, and long-term winners remain uncertain.
• API pricing for lower-end agentic text models remains low, putting pressure on gross margins.
• Model performance or intelligence improvement may fall short of expectations, affecting pricing power and commercialization.
• The path to profitability may be slower than expected, with high R&D investment and cash burn weakening sustainability.
• Risks related to IP, content generation, and human oversight may affect the expansion of AI applications.
• US-China technology competition, policy restrictions on Chinese models in overseas markets, and tighter overseas access to frontier Chinese models in the future could all alter the revenue growth trajectory.
• Access to high-end training compute and domestic ASIC supply are key swing factors.

What to watch

• The launch pace in 2H26 of 2-5 trillion parameter models such as Zhipu GLM, Alibaba Qwen, and MiniMax M3 Pro.
• Pricing, usage volume, and developer adoption of high-end coding models after the launch of Kimi K3.
• Whether coding and agentic entry points such as Zhipu ZCode, Tencent Workbuddy, and Alibaba Qoder can form real data closed loops.
• Whether multimodality and visual understanding become the next key upgrade focus for Chinese foundation models.
• Whether API pricing for lower-end agentic models continues to stay around US$0.1-0.2 / 1 million tokens.
• The launch of MiniMax H3 video generation model, Hailuo adoption, M3 updates, and the timing of M3 Pro release.
• The transmission of AI hyperscaler capex in 2H26 to Alibaba, GDS, VNET, and Kingsoft Cloud.
• Domestic ASIC supply, overseas policy, overseas access restrictions on frontier models, and the availability of high-end training compute.