China LLM: Open source does not mean the cost advantage is replicable; enterprise applications and inference hardware are the core themes

Overview

This report focuses on LLM trends in China's internet and new media sector. Its core input came from a conference call with experts from Nomura's China Internet team and an AI laboratory under a Chinese research institution. That lab has deployed its own foundational models to more than 100 enterprise customers. The report discusses high-end AI chip supply, domestic accelerator economics, DeepSeek cost advantage, commercialization of open-source models, model pricing segmentation, and key constraints for enterprise adoption of LLMs.

Core views

The core thesis is that the China LLM industry is not entering full price-deflation across the board. Base models may continue to be cut in price to win customers because tasks are becoming more commoditized and customer switching costs are low; however, advanced models, customized services, and models deeply embedded in enterprise processes can still sustain a premium. Although DeepSeek open-sourced some model weights and software stack, its inference cost advantage mainly comes from system-level engineering and operating experience. Even when deployed on similar hardware, third-party cloud platforms are unlikely to fully replicate its efficiency. Tight supply and rising prices for high-end Nvidia chips have increased the appeal of domestic accelerators such as Huawei Ascend in inference scenarios.

Analysis framework

The report combines expert interviews with industry-chain observation: first using the practical deployment experience of experts from a Chinese AI lab to assess model deployment, hardware fit, client demand, and pricing strategy, then incorporating Nomura's sector feedback on China internet, cloud vendors, and the AI model ecosystem to form a view on LLM commercialization pathways and competitive dynamics.

Methodology notes

• Expert interviewChinese AI lab expert conference call

  Validating industry trends through frontline deployment experience

  The expert's lab has deployed its own base models to over 100 enterprise customers, so its feedback can be used to observe enterprise buying standards, model price segmentation, domestic hardware adaptation, and actual inference cost.

• Industry competitive analysisOpen-source model replicability analysis

  Open-source weights do not equal an open operational playbook

  The report distinguishes model weights, portions of software stack, and full production deployment capability, arguing that cache efficiency, latency control, hardware utilization, and engineering experience are the core sources of DeepSeek's cost advantage.

• Business model analysisDual-track LLM pricing

  Base models for acquisition, advanced models for monetization

  Base models face price competition for commoditized tasks, while advanced models and customized services command higher switching costs and stronger pricing power due to quality, stability, low latency, and workflow integration.

• Customer segmentationSecurity, compliance, and ROI two-axis

  Different enterprise customer segments use different LLM adoption criteria

  Government and state-owned entities prioritize data security, regulatory compliance, and local deployment; private firms are more focused on whether an AI initiative can recover initial investment within 12 to 18 months.

Asset mapping & comparison

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

• DeepSeek

  Representative of low-cost inference and open-source models

  Strengths

  System-level optimization, cache efficiency, low latency, and hardware utilization deliver lower average inference cost, enabling greater room for official API token price cuts.

  Weaknesses

  Key implementation details and operating experience are not fully public, so external platforms cannot fully replicate its native deployment efficiency using open weights alone.

  Comparison

  Compared with Tencent, Alibaba, and ByteDance cloud platforms that deploy the same open-source model, experts consider DeepSeek's native deployment to have higher operating efficiency and lower inference cost.

  Risks

  Whether open-source strategy remains sustainable, whether cost advantages persist, and commercialization scale and customer willingness to pay still require monitoring.

• Tencent (700 HK)

  Cloud platform and LLM deployment participant

  Strengths

  Has cloud infrastructure, enterprise customer channels, and model deployment capability.

  Weaknesses

  Even when deploying the same DeepSeek open-source model, it may be difficult to replicate DeepSeek's native deployment efficiency.

  Comparison

  The report groups it with Alibaba and ByteDance as third-party platforms deploying the open-source DeepSeek model for comparison.

  Risks

  Competition on base model token pricing, inference cost gaps, and the depth of enterprise workflow integration can affect monetization.

• Alibaba (BABA US)

  Cloud platform, model provider, and open-source ecosystem participant

  Strengths

  Early proponent of open-source strategy; Qwen ecosystem has developer influence. The report notes that Qwen is competitive in the code model lineup.

  Weaknesses

  Its latest flagship model, Qwen 3.6 Max, has moved to closed-source, highlighting trade-offs between open-source and commercialization.

  Comparison

  Consistent with Moonshot expert views, open-source is better for brand and developer ecosystem building, while closed-source remains the main path to large-scale monetization for proprietary models.

  Risks

  Base model pricing pressure, changes in open-source strategy, and uncertainty around advanced model commercialization remain.

• Huawei Ascend

  Domestic AI accelerator and inference hardware substitute

  Strengths

  In an environment of tight high-end Nvidia chip supply and rising prices, domestic accelerators become more economically attractive for inference workloads.

  Weaknesses

  Need to continue proving software ecosystem depth, model adaptation, stability, and large-scale deployment efficiency.

  Comparison

  Nvidia's high-end chip price increases are more pronounced; domestic accelerators have also risen in price but to a smaller extent.

  Risks

  Supply, ecosystem maturity, performance stability, and customer migration costs are key risks.

• Nvidia (NVDA US)

  Supplier of high-end training and inference chips

  Strengths

  High-end chips remain the core hardware foundation for AI training and high-performance inference.

  Weaknesses

  Tight supply and higher prices weaken economic attractiveness for some clients.

  Comparison

  Domestic accelerators gain relative appeal for inference workloads as Nvidia's high-end chip prices rise.

  Risks

  Demand-supply imbalance, price volatility, export restrictions, and penetration of substitute hardware may reshape demand dynamics in China.

• Financial services

  High-potential LLM commercialization vertical

  Strengths

  Has abundant proprietary data and relatively high willingness to pay for advanced AI tools.

  Weaknesses

  Requires high levels of security, compliance, and data governance.

  Comparison

  The expert lists financial services as one of the most attractive verticals.

  Risks

  Regulatory constraints, model reliability, and data-security requirements may lengthen implementation timelines.

• Office productivity and code

  Enterprise LLM use cases

  Strengths

  Office productivity can cover document generation, meeting management, and ERP/CRM integration; code capability is a foundational layer for agents, bots, and automated operations.

  Weaknesses

  General office tasks may face stronger price competition, while code models require continuous improvements in accuracy and reliability.

  Comparison

  The expert sees GLM-5.2 as a leading Chinese code model, followed by new releases from DeepSeek, Moonshot, and Alibaba Qwen.

  Risks

  Customer ROI validation, depth of workflow integration, and model error costs are major constraints.

Key data

• Report date2026-07-13The report cover page shows the date as 13 July 2026.
• Lab deployment scale100+ enterprise customersThe expert stated that its own foundation models have been deployed to more than 100 enterprise customers.
• Private enterprise ROI requirement12-18 monthsPrivate companies generally want to clearly see that initial AI investment is recovered within 12 to 18 months.
• Industrial manufacturing workflow penetrationbelow 10%The expert estimated that industrial manufacturing is still in an early adoption phase, with workflow penetration below 10%.
• Nomura global equity research rating distribution58% Buy / 39% Neutral / 3% ReduceThis data comes from the disclosure page and is the Nomura Group rating distribution disclosure, not a rating on China LLM names in this report.
• Mentioned listed targetsTencent (700 HK, Buy); Alibaba (BABA US, Buy); Nvidia (NVDA US, Not rated)The ratings are those appearing in the report body in parentheses; the report itself does not provide new target prices or rating revisions.

Impact & implications

For investment research, the focus of monitoring the LLM industry should shift from single Token price declines to commercialization quality, customer switching costs, and inference cost structure. Participants with advanced model capability, strong engineering-grade deployment efficiency, enterprise workflow integration, and domestic hardware adaptation are more likely to maintain margin resilience in a price-competitive environment. The opportunity for domestic AI accelerators mainly comes from tight high-end Nvidia chip supply and demand for inference cost optimization, but their long-term competitiveness still needs to be validated through stability, ecosystem strength, and real deployment efficiency.

Risks

• Continued tightness and price volatility in high-end AI chip supply may further affect training and inference costs.
• Token-competitive pressure may persist as base model tasks become more commoditized and switching costs for customers remain low.
• Open-source models may not remain perpetually open, and open-source weights do not guarantee replication of original developers' operating efficiency.
• Government and state-owned customers' strong requirements for data security, regulatory compliance, and localized deployment may slow penetration by some private-model providers.
• Private companies expect AI investments to be recovered within 12 to 18 months; if ROI is not clear, enterprise AI budget deployment may be lower than expected.
• The premium for advanced models depends on reliability, service stability, and depth of workflow embedding; if implementation quality is weak, renewal and repricing power will be pressured.

What to watch

• Changes in supply, pricing, and actual inference costs of high-end Nvidia chips and domestic AI accelerators.
• Whether base model token pricing continues to decline and whether advanced models and customized services maintain a premium.
• Whether the cost gap between DeepSeek native deployment and third-party deployment of the same open-source model remains persistent.
• After Alibaba Qwen 3.6 Max moves to closed source, whether other Chinese AI labs continue to open frontier models.
• Procurement pace across government, state-owned, and private enterprises regarding localized deployment, data security, and ROI thresholds.
• Actual commercialization revenue contribution from verticals such as financial services, office productivity, code, industrial manufacturing, legal, and healthcare.