LLM expert call: Improving economics of domestic AI accelerators, stronger commercialization resilience for advanced models
AI summary card
LLM expert call: Improving economics of domestic AI accelerators, stronger commercialization resilience for advanced models
Nomura's call expert believes that tight supply and rising prices for Nvidia high-end chips are increasing the attractiveness of domestic AI accelerators, DeepSeek's low inference costs are not easily replicable, and finance, office productivity, coding, manufacturing, legal, and healthcare will be the key LLM monetization scenarios.
- Both high-end Nvidia chips and domestic accelerators are seeing price increases due to strong demand and tight supply, but the smaller increase in domestic hardware makes it more economically attractive for inference workloads.
- DeepSeek's cost advantage comes from system-level optimization, cache efficiency, low latency, and hardware utilization, not just open-source model weights, so cloud vendors deploying the same open-source model also struggle to replicate its native efficiency.
- Base-model prices may continue to decline and serve customer acquisition purposes, but advanced models deeply embedded in enterprise production workflows, customized services, and highly reliable services have stronger pricing power.
- Government and state-owned enterprises place greater emphasis on data security, compliance, and localized deployment, while private enterprises focus more on visible ROI from AI investments within 12 to 18 months.
Report interpretation
Overview
This report summarizes a call between Nomura's China Internet team and an expert from a Chinese AI laboratory. The lab is affiliated with a research institution and has already deployed its self-developed foundation model to more than 100 enterprise clients. The discussion focused on high-end AI chip supply and demand, the substitution economics of domestic accelerators, the true moat of DeepSeek's open-source models, LLM price differentiation, barriers to enterprise adoption, and the industries with the strongest commercialization potential.
Core views
The core views are: first, supply of high-end training hardware remains tight, Nvidia high-end chip prices have risen sharply, and domestic accelerators have also become more expensive but to a lesser extent, making them more cost-attractive in inference scenarios. Second, open source does not mean commercial secrets are fully disclosed; DeepSeek's low inference costs depend on system-level operating efficiency, caching, latency, and hardware utilization, which third-party platforms cannot fully replicate even if they deploy the same open-source model. Third, the LLM industry is not seeing across-the-board deflation; declining base-model prices are mainly used for customer acquisition, while advanced models and customized services can still maintain premiums due to reliability, quality, latency, stability, and high switching costs. Fourth, financial services, office productivity, coding, manufacturing, legal, and healthcare are considered the most important verticals for future revenue contribution.
Analysis framework
The report uses a combination of expert interviews and industry observation, breaking the AI value chain into hardware supply, model deployment efficiency, model pricing strategy, enterprise customer segmentation, and vertical-industry commercialization opportunities, with cases such as DeepSeek, Nvidia, Huawei Ascend, Tencent, Alibaba, ByteDance, and Moonshot used to support the analysis.
Methodology notes
Use feedback from front-line AI lab experts to assess industry supply-demand, costs, and commercialization trends.
The expert's institution has already deployed its self-developed foundation model to more than 100 enterprise clients, so its views can reflect the practical constraints around enterprise adoption, model pricing, and localized deployment.
Break down LLM costs into chip supply, inference efficiency, cache efficiency, latency, and hardware utilization.
The report emphasizes that DeepSeek's advantage is not just that the model is open source, but also its system-level optimization and operational experience, enabling it to achieve lower average inference costs under similar hardware conditions.
Differentiate between customer-acquisition pricing for base models and premium pricing for advanced models/customized services.
Base models handle homogenized tasks and customers have low migration costs, leading to intense price competition; once advanced models are embedded into production workflows, switching costs rise and vendors gain stronger bargaining power.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- NVIDIA CORP (NVDA.US)Upstream supplier of high-end AI chips; the report mentions tight supply and rising prices for its high-end chips.
- Strengths
- High-end training and inference chips still enjoy strong demand, and some high-end products have seen significant price increases over the past year.
- Weaknesses
- In the China market, it faces improving economics of domestic accelerators and stronger appeal of localized substitution.
- Comparison
- Compared with domestic accelerators, Nvidia's high-end chip price increases have been more pronounced; domestic hardware is becoming more cost-attractive for inference workloads.
- Risks
- Tight supply, elevated prices, customer migration to domestic hardware, and geopolitical/regulatory restrictions may affect related demand in China.
- DeepSeek (unlisted)Chinese LLM provider; the report focuses on its low inference costs and open-source strategy.
- Strengths
- System-level optimization, cache efficiency, low latency, and hardware utilization deliver lower average inference costs and support more flexible pricing for its official API.
- Weaknesses
- Some implementation details and operational experience are difficult to verify externally, and it is uncertain whether the open-source strategy will continue in the future.
- Comparison
- Even if platforms such as Tencent, Alibaba, and ByteDance deploy the same open-source DeepSeek model, the expert believes DeepSeek's native deployment efficiency remains higher.
- Risks
- Base-model price wars, a trend toward closed-source commercialization, competing model upgrades, and ROI scrutiny from enterprise clients.
- Tencent (700 HK)Cloud and internet platform; the report mentions its deployment of the open-source DeepSeek model.
- Strengths
- Has cloud infrastructure and enterprise customer reach; rated Buy by Nomura.
- Weaknesses
- Deploying the same open-source model may not replicate the original developer's operating efficiency.
- Comparison
- Compared with DeepSeek's native deployment, third-party platforms may be at a disadvantage in average inference cost and efficiency.
- Risks
- Model price competition, enterprise customer switching, and insufficient cost efficiency.
- Alibaba (BABA US)Cloud and AI model platform; the report mentions its Qwen model and changes in its open-source/closed-source strategy.
- Strengths
- Owns the Qwen model ecosystem and cloud platform capabilities; rated Buy by Nomura.
- Weaknesses
- Its latest flagship model, Qwen 3.6 Max, is described as closed source, indicating that open source is not a fixed strategic path.
- Comparison
- It promoted the open-source ecosystem in the early stage, but closed source is still viewed as the main path for large-scale monetization of proprietary models.
- Risks
- Strategic trade-offs between the open-source ecosystem and closed-source monetization, declining base-model prices, and iteration of competing models.
Key data
- Number of enterprise clientsMore than 100The self-developed foundation model of the expert's AI lab has been deployed to more than 100 enterprise clients.
- NVDA ratingNot ratedThe report mentions Nvidia (NVDA US, Not rated) facing tight chip supply and rising prices.
- Required payback period for enterprise AI investment12 to 18 monthsPrivate enterprises are more concerned with whether the initial AI investment can be recovered within 12 to 18 months.
- AI penetration rate in industrial manufacturingBelow 10%The expert estimates that AI adoption in industrial manufacturing workflows is still at an early stage, with penetration below 10%.
- Nomura Global Equity Research rating distributionBuy 58%; Neutral 39%; Reduce 3%The rating distribution of Nomura Group Global Equity Research shown on the disclosure page.
Impact & implications
From an investment perspective, the shortage of high-end AI chips strengthens the substitution logic for domestic accelerators in China's inference scenarios; at the model layer, simply open-sourcing weights is not enough to eliminate the operational moat of leading vendors; at the application layer, industries with high-quality proprietary data, strong ROI, and deep workflow integration are more likely to contribute LLM revenue first. For internet and cloud platforms, short-term price wars in base APIs may depress headline unit prices, but advanced models, customized services, and enterprise-grade stability can still support long-term monetization.
Risks
- Continued tight supply of high-end AI chips may push up training and inference infrastructure costs.
- Intensifying price competition among base models may compress margins on homogenized API workloads.
- Enterprise clients have high requirements for data security, compliance, localized deployment, and ROI visibility, which may slow adoption.
- Open-source models are not guaranteed to remain open, and leading model vendors may shift to closed source to strengthen monetization.
- When third-party cloud platforms deploy open-source models, they may lack sufficient cost advantage if they cannot replicate the original developer's operating efficiency.
What to watch
- Changes in supply, pricing, and delivery cycles for Nvidia high-end chips and domestic AI accelerators.
- Actual costs, stability, and ecosystem adaptation progress of domestic accelerators in inference workloads.
- Whether DeepSeek and other Chinese AI laboratories continue to open-source frontier models.
- Whether pricing for base-model APIs and advanced models/customized services continues to diverge.
- Enterprise paid conversion and ROI validation in finance, office productivity, coding, manufacturing, legal, and healthcare.