DeepSeek Expert View: Opportunities and Constraints for China LLMs in Coding, Agents, and Chip Supply
AI summary card
DeepSeek Expert View: Opportunities and Constraints for China LLMs in Coding, Agents, and Chip Supply
HSBC, through an expert conversation with DeepSeek product manager Eric Chen, outlines the Chinese large-model competitive landscape, capability gaps versus the United States, chip supply bottlenecks, and AI commercialization pathways.
- In coding capability, the expert believes Zhipu (Knowledge Atlas), Kimi and DeepSeek are among the domestic leaders. Key drivers include post-training strategy, data quality, interdisciplinary talent, and chip resources.
- On Agent capability, the expert places Alibaba Qwen, Kimi and DeepSeek as leading domestic players; Tencent and Alibaba benefit from user base, data and ecosystem tools, while independent labs rely on base-model iteration and architectural differentiation.
- In the China-US capability comparison, the expert believes Chinese video generation is broadly on par with the U.S., while the U.S. still leads China by about 1-3 months in text/reasoning, 6 months in multimodal, and 8-12 months in Agent productization.
- Chip supply remains a constraint: domestic chips currently cover less than 40% of China’s inference demand, and could rise to 40-45% by the end of 2026; if a 7nm-level manufacturing breakthrough is achieved, it could rise to 50%.
- In commercialization, overseas markets rely more on higher-priced subscriptions and closed ecosystems, while Chinese internet platforms monetize more through traffic, cloud, e-commerce and super-app cross-entry points. Long-term declines in token prices will push toward value-based services and custom workflow fees.
Report interpretation
Overview
This report is an HSBC expert-view digest on China's AI and large-language-model industry, based primarily on an exchange on May 8, 2026 with DeepSeek product manager Eric Chen. It covers the relative positioning of domestic model vendors in coding, Agent, multimodal, and video generation, and discusses the China-U.S. capability gap, domestic chips’ coverage of inference demand, and AI product commercialization strategy.
Core views
The core view of the report is that China’s LLM ecosystem is competitive in coding, video generation and low-cost models, but still trails the U.S. in Agent maturity, multimodal data quality, advanced architectures, talent, and dependence on the CUDA ecosystem. In the near term, advanced chip shortages may constrain training efficiency and model iteration speed, but they are not the only determinant of capability gaps. Commercialization is expected to shift from pure token-consumption billing toward service-based offerings, workflow customization, and pricing based on delivered value.
Analysis framework
The report uses expert interviews and an industry cross-sectional comparison approach, breaking domestic and international model capabilities into dimensions such as coding, Agent, multimodal, video generation, chip supply and commercialization models, with supporting summaries of selected internet and AI company ratings, current prices and target prices.
Methodology notes
Judges the China LLM competitive landscape through the industry experience of DeepSeek product manager Eric Chen.
This approach is suitable for capturing industry trends in model productization, engineering deployment, and commercialization strategies that are difficult to fully quantify from public data, but conclusions depend on expert perspective rather than complete empirical statistics.
Decomposes large-model capabilities into coding, Agent, multimodal, video generation, data, talent and chip resources.
The report uses category-by-category comparisons to explain capability gaps between China and the U.S. and to identify the relative strengths of different domestic players.
HSBC uses target price versus current price upside/downside to classify Buy, Hold, and Reduce ratings.
The appendix states that the target price generally reflects a 6 to 12 month performance window; when the target price is more than 20% above the current price it is generally classified as Buy, while prices near the current range are typically Hold.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Alibaba / QwenOne of the leading domestic players in Agent capability and tool ecosystem
- Strengths
- It has a large ecosystem, user base, proprietary data, and ready-made products/tools, which benefit Agent integration and commercialization.
- Weaknesses
- Compared with top-tier U.S. Agent products, mature toolkits and deployment-ready product layers still lag.
- Comparison
- The expert groups Alibaba Qwen together with Kimi and DeepSeek as a leading domestic Player in Agent capability.
- Risks
- If the speed of Agent productization is slower than that of the U.S., it may reduce commercialization premium.
- TencentA Chinese AI platform benefiting from a large internet ecosystem
- Strengths
- It has a large user base, high-quality proprietary data, and strong ecosystem entry points.
- Weaknesses
- The report does not place it in the first tier of coding or Agent models, but it highlights ecosystem advantages.
- Comparison
- Like Alibaba, Tencent benefits more from ecosystem and product integration capabilities than independent labs.
- Risks
- If model capability and productization pace lag, ecosystem advantages may be insufficient to convert into AI revenue.
- DeepSeekRepresentative of leading domestic large models and low-cost models
- Strengths
- It is listed as among domestic leaders in coding ability, and price cuts help it gain share.
- Weaknesses
- As an independent lab, it needs to maintain competitiveness through base-model iteration and architectural differentiation.
- Comparison
- In coding, it is in the same leading group as Zhipu and Kimi; in Agent, it is in the same leading group as Alibaba Qwen and Kimi.
- Risks
- Long-term declines in token prices may compress pure API revenue, requiring a shift toward service and workflow-based value realization.
- KimiOne of the domestic leading models in coding and Agent capability
- Strengths
- The expert includes it among domestic leaders in both coding and Agent dimensions.
- Weaknesses
- The report does not provide specific commercialization or financial data.
- Comparison
- Along with DeepSeek, Zhipu, and Alibaba Qwen, it is part of the domestic first-tier model comparison set.
- Risks
- Sustained investment in data, talent, architecture, and productization capability is needed to narrow the gap with the U.S.
- Zhipu / Knowledge AtlasA domestic model company with relatively strong coding and multimodal understanding
- Strengths
- The expert considers it strong in coding and notes its multimodal understanding capability is relatively strong.
- Weaknesses
- The stock table shows Knowledge Atlas rated Hold, with a target price of 920.00 HKD below the current price of 923.00 HKD.
- Comparison
- Its coding capability is listed in the same leading group as Kimi and DeepSeek; its multimodal understanding is highlighted alongside Minimax.
- Risks
- Valuation and earnings realization pressure, as well as multimodal data quality gaps.
- MinimaxAI company linked to multimodal understanding and video generation
- Strengths
- The expert highlights its strong multimodal understanding and notes it together with Kuaishou in video generation.
- Weaknesses
- The stock table rates it Hold, even though the target price is above the current price.
- Comparison
- It is mentioned alongside Zhipu in multimodal understanding and alongside Kuaishou in video generation.
- Risks
- Competition in video generation is intense, and commercialization and cost control still need to be proven.
- KuaishouInternet platform with video generation-related capabilities
- Strengths
- The expert highlights its advantage in video generation, and the stock table rates it Buy.
- Weaknesses
- The report does not detail its overall large-model capability.
- Comparison
- It is highlighted together with Minimax in video generation; China and the U.S. are viewed as broadly comparable in video generation ability.
- Risks
- It remains to be seen whether video-generation strength can translate into advertising, content ecosystem, and commercial revenue.
- Domestic AI chipsA variable in domestic replacement of China’s AI inference demand and training efficiency constraints
- Strengths
- Domestic chip coverage of inference demand is expected to rise from below 40% to 40-45% by end-2026, with room to reach 50% if there is a technology breakthrough.
- Weaknesses
- Chipsets are only used in small-scale experiments for training scenarios, while manufacturing capacity and dependence on CUDA software ecosystems are key bottlenecks.
- Comparison
- Still at a disadvantage versus the U.S. in advanced chips and software ecosystem, but with substantial substitution potential on the inference side.
- Risks
- Advanced chip shortages may limit training efficiency and model iteration cycles.
Key data
- Meeting date2026-05-08Hosted by HSBC with DeepSeek product manager Eric Chen.
- Report issuance time2026-05-11 07:30 GMTThe report states that the authors prepared and issued it at this time.
- Current domestic chip coverage of inference demandunder 40%The expert estimates domestic chips currently cover less than 40% of China’s inference demand.
- Expected domestic chip inference-demand coverage by year-end 202640-45%If there is a breakthrough in 7nm-class manufacturing technology, the coverage could reach up to 50%.
- US lead over China in text/reasoning1-3 monthsThe expert believes the U.S. leads China by about 1 to 3 months in text and reasoning capabilities.
- US lead over China in multimodal6 monthsThe expert believes the gap mainly comes from data quality and availability, as well as more advanced model architectures in the U.S.
- US lead over China in Agent capability8-12 monthsThe gap mainly appears in mature toolkits, skill systems, and deployable Agent product layers.
- Related stock summary700 HK Buy; BABA US Buy; 2513 HK Hold; 100 HK Hold; 1024 HK BuyThe table lists the ratings, current prices, and target prices for lencent, Alibaba, Knowledge Atlas, Minimax, and Kuaishou.
Impact & implications
For investors, the report indicates that opportunities in China’s AI value chain are not only about model parameters or token prices, but more about ecosystem entry points, proprietary data, product and tool integration, enterprise workflows, and domestic inference-chip substitution. Internet platforms may monetize AI indirectly through traffic, cloud, e-commerce, and super-app cross-selling, while independent labs depend more on memberships, API, private deployment, licensing, fine-tuning, and maintenance services.
Risks
- Advanced chip supply tightness may limit China's model training efficiency and iteration speed.
- Insufficient domestic chip manufacturing capacity and continued dependency on the CUDA software ecosystem could slow the substitution process.
- China lags the U.S. by about 8-12 months in mature Agent toolkits, skill stacks, and deployment-ready product layers.
- The multimodal capability gap is mainly constrained by data quality and accessibility, and is not easy to fully close in the short term.
- Long-term declines in token prices may compress base API revenue, forcing providers toward more complex service-based and workflow-based pricing.
- The report is an expert-opinion digest, and some judgments lack independently verifiable quantitative samples.
What to watch
- Whether domestic chip coverage of China’s inference demand can rise to 40-45% by the end of 2026.
- Whether a 7nm-class manufacturing breakthrough can further raise domestic chip inference coverage to 50%.
- The speed of iteration in Agent toolchains and productization layers for domestic models such as Alibaba Qwen, Kimi, and DeepSeek.
- Commercial execution in multimodal understanding and video generation by Zhipu, Minimax, and Kuaishou.
- Whether Chinese internet platforms can monetize AI indirectly through cloud, e-commerce, super-app entry points and ecosystem cross-selling.
- Whether independent labs can shift from price competition to revenue from memberships, API, private deployment, licensing, fine-tuning, and maintenance services.