Domestic AI compute, DeepSeek cost advantage, and enterprise-grade LLM commercialization emerge as the core themes

Overview

This report summarizes the key points from a call between Nomura’s China Internet team and experts from an AI lab affiliated with a Chinese research institution. The lab has deployed its proprietary foundation model at more than 100 enterprise clients. The report focuses on hardware supply constraints in China’s LLM industry, the real cost barriers behind DeepSeek’s open-source models, price differentiation between foundation and advanced models, and the trade-offs enterprises face among security, compliance, and return on investment when adopting AI.

Core views

The core view is that China’s LLM industry is not entering broad-based price deflation. Foundation models remain commoditized in many tasks and are still easily transferable, so suppliers may continue using them as a low-cost customer acquisition tool. But for large enterprises, high-performance models, customized services, and workflow-embedded solutions command more robust pricing potential because clients place greater emphasis on reliability, completion quality, latency, and service stability. DeepSeek’s advantage is not only that model weights are open-sourced, but also system-level inference efficiency, operational experience, and hardware utilization; therefore, even if third-party cloud platforms deploy the same open-source model, they may not replicate its native deployment cost advantage.

Analysis framework

The report uses expert interviews and industry feedback synthesis, analyzing China’s LLM value chain across three dimensions: supply-side compute and model efficiency, demand-side enterprise adoption barriers, and commercialization scenarios. Discussion topics include Nvidia high-end chips, Huawei Ascend domestic accelerators, DeepSeek open-source models, cloud provider deployment efficiency, and vertical applications in finance, office, code, manufacturing, legal, and healthcare.

Methodology notes

• Industry expert interviewLLM expert conference call

  Expert interview method

  Validates LLM supply, pricing, hardware, and commercialization trends through industry experts with enterprise-level model deployment experience.

• Value chain analysisHigh-end Chip Constraint and Inference Unit Economics Analysis

  AI compute supply-demand and unit inference cost

  Assesses supplier cost structure from high-end chip supply, price changes, domestic accelerator substitution, and model inference efficiency.

• Business model analysisPrice tiering between foundation and advanced models

  Customer acquisition and workflow lock-in

  Treats low-end general models as the entry point for customer acquisition, while advanced models and customized services are more sticky and have greater pricing power.

Asset mapping & comparison

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

• DeepSeek

  Unlisted model provider and core participant in the open-source model ecosystem

  Strengths

  Lower inference costs with strong system-level optimization, cache efficiency, low latency, and hardware utilization, and more flexible official API pricing.

  Weaknesses

  Some key implementation details and operational know-how are not public, making direct mapping to listed assets difficult for external investors.

  Comparison

  Compared with cloud platforms deploying the same open-source model, native deployment efficiency is higher and average inference costs are lower.

  Risks

  The long-term value could be affected by open-source strategy, commercialization path, competitive dynamics, and changes in the regulatory environment.

• Huawei Ascend

  Representative of domestic AI accelerators

  Strengths

  As high-end Nvidia chip supply remains tight and prices rise, inference economics improve; localization preferences in government and state-owned enterprises support adoption.

  Weaknesses

  Ecosystem, software adaptation, and ultra-high-end training capability still need to compete with internationally leading hardware.

  Comparison

  Relative to high-end Nvidia chips, price increases have been more moderate, increasing the appeal of domestic substitution.

  Risks

  Supply chain, performance iteration, ecosystem adaptation, and customer migration costs remain key uncertainties.

• Tencent (700 HK)

  Cloud platform and model deployment participant

  Strengths

  Has cloud services, enterprise customers, and an application ecosystem, enabling it to support open-source model deployment and enterprise AI demand.

  Weaknesses

  If deploying the same open-source models with lower efficiency than the original vendor, it may face cost and pricing pressure.

  Comparison

  The report notes that cloud vendors may find it difficult to replicate DeepSeek’s native deployment cost advantage when deploying DeepSeek open-source models.

  Risks

  Model commoditization, API price declines, compute costs, and enterprise ROI requirements.

• Alibaba (BABA US)

  Cloud platform, Qwen model, and open-source/proprietary strategy participant

  Strengths

  Has cloud infrastructure, the Qwen model ecosystem, and enterprise service capabilities.

  Weaknesses

  Open-source helps with branding and developer ecosystem building, but large-scale monetization of proprietary models still depends on closed approaches and enterprise services.

  Comparison

  The report notes that while Alibaba previously supported open-source, its latest flagship model Qwen3.6Max was introduced in closed-source form.

  Risks

  Balancing open-source and closed-source strategy, price competition, compute costs, and changing customer demand.

• Nvidia (NVDA US)

  High-end AI training chip supplier

  Strengths

  Demand for high-end training hardware remains strong, and tight supply supports pricing.

  Weaknesses

  Rising prices and constrained supply increase the attractiveness of domestic alternatives for Chinese inference scenarios.

  Comparison

  The report says some high-end Nvidia chips had significantly higher price increases over the past year than domestic accelerators.

  Risks

  China market supply constraints, customer shifts to domestic hardware, and cost pressure in inferencing economics.

Key data

• Number of enterprise clients in expert’s labOver 100The lab has deployed its proprietary foundation model across more than 100 enterprise clients.
• Private enterprise AI payback horizon12 to 18 monthsThe expert said private-enterprise clients typically want clear evidence that initial AI investment can be recovered within 12 to 18 months.
• Estimated industrial manufacturing workflow penetrationBelow 10%The expert views AI adoption in industrial manufacturing as still early stage, with expected acceleration as multimodal capability improves.
• Nomura Group Buy rating share58%Disclosed materials show a Buy rating share of 58% in Nomura Group global equity research.
• Nomura Group Neutral rating share39%Disclosed materials show a Neutral rating share of 39%.
• Nomura Group Reduce rating share3%Disclosed materials show a Reduce rating share of 3%.

Impact & implications

For investment research, the key variables in the LLM sector are shifting from pure model capability competition to computing cost, deployment efficiency, workflow integration, and monetization in vertical use cases. Improved inferencing cost competitiveness of domestic AI accelerators may increase the importance of local hardware and software adaptation ecosystems. At the same time, open-source models do not necessarily weaken vendor advantages; model providers with system-level optimization and commercialization channels may still maintain differentiation. If cloud vendors and internet platforms rely only on open-source weights and price competition, they may face pressure on costs and pricing; those that provide stable, low-latency, compliant, and business-process-embedded enterprise services are more likely to achieve sustainable revenue.

Risks

• Persistent tightness in high-end AI chip supply may constrain the pace of model training and enterprise deployment.
• Continued declines in foundation model API pricing may compress margins on generic model services.
• Open-source model brand and ecosystem advantages may not directly translate into monetization, and closed-source plus customization paths still carry execution risk.
• Different security, compliance, and ROI priorities across government, state-owned, and private enterprises may lengthen sales cycles.
• If enterprises cannot see clear returns within 12 to 18 months, they may delay or reduce AI spending.
• High-potential sectors such as legal, healthcare, and manufacturing face regulatory, data-security, system integration, and reliability barriers.

What to watch

• Changes in supply, pricing, and performance gaps between high-end Nvidia chips and domestic AI accelerators.
• Relative performance of Chinese models such as DeepSeek, Qwen, GLM, and Moonshot in inference cost, coding capability, and enterprise deployment.
• Whether foundation model API prices continue to fall and whether advanced models plus customized services can maintain premiums.
• Actual unit inference cost, latency, and service stability of cloud vendors after deploying open-source models.
• Order conversion from localization requirements of government and state-owned enterprises versus private-enterprise ROI-oriented demand.
• Paid penetration in financial, office productivity, code, manufacturing, legal, and healthcare scenarios.