TL;DR
· Nomura's expert meeting on July 13th showed that the Chinese Large Language Model (LLM) market is transitioning from a simple price reduction to a dual-tier model where basic models are offered at a low price while advanced models come with a premium.
· DeepSeek's cost advantage comes from system optimizations such as caching, scheduling, latency, and hardware utilization, indicating that open-source model weights do not necessarily lead to replicated operational efficiency.
· Domestic accelerators are gaining more opportunities in inference and on-premises deployment, but enterprise projects still need to pass a 12 to 18-month ROI validation.

The Nomura China Internet team, after a discussion with a Chinese AI lab expert on July 13th, provided a more practical assessment: the Chinese large model market is not simply heading towards lower prices but is instead dividing into two tiers – with basic models continuing to lower prices to attract customers, while advanced models, private deployments, and enterprise customization retain a premium.

This proprietary AI lab's basic model has been deployed to over 100 enterprise customers, and the team is also an early adopter of domestic accelerators like Huawei Ascend. The key signal released during the expert meeting is that while model capabilities are becoming easier to compare, what truly determines platform profitability and customer stickiness has shifted from leaderboard rankings to inference costs, deployment efficiency, and enterprise workflows.

This is not an official Nomura research report and does not represent industry-wide statistics. However, it offers a perspective closer to enterprise procurement: customers are not just buying a model but also calculating chip prices, cost per inference, system integration, data security, and the project's payback period.

DeepSeek's Low Cost, Hard in System Optimization

DeepSeek is the most typical case in this logic.

The market often attributes DeepSeek's low cost to open-source models, but open weights only lower the usage threshold and do not automatically replicate the native platform's operational efficiency. What truly impacts the inference bill are cache hit rates, request scheduling, batch processing strategies, latency control, and hardware utilization.

The MLA and DeepSeekMoE architectures disclosed in the DeepSeek-V3 technical report, along with the load balancing and throughput optimization mentioned in its infrastructure documentation, all point to the same goal: accomplishing more calls with less hardware footprint.

This implies that even if platforms like Tencent, Alibaba, ByteDance, and others can deploy the same open weights, they may not necessarily achieve the same costs in a real business environment. Over the long term, differences in milliseconds of latency, cache efficiency, and hardware utilization efficiency could eventually translate into significant billing discrepancies for enterprise customers.

Therefore, the competitive pressure brought by DeepSeek is not just "cheaper models," but is forcing the entire industry to recalculate the actual cost of each token, each call, and each business process.

The Base Model Drives User Acquisition, while Deep Deployment Drives Revenue

A price war in the Chinese large-scale model market is becoming stratified.

The base models targeting developers and lightweight use cases are becoming increasingly commoditized, with continued downward pressure on prices. Platforms can expand their call volume through low prices or even subsidies, turning models into the entry point for cloud services and the AI ecosystem.

However, when models enter customer service, financial risk control, code repositories, ERPs, CRMs, or production scheduling systems, customers are no longer purchasing just an API but a set of business systems that need to run reliably. The deeper the deployment, switching vendors requires data migration, process reengineering, security testing, and staff training, all leading to increased switching costs.

This allows model vendors to adopt two pricing strategies simultaneously: lowering prices for basic capabilities to acquire customers and using advanced models, industry solutions, private deployments, and customized deliveries to drive monetization.

Open source and closed source are not necessarily mutually exclusive. Open-source models can attract developers and expand the ecosystem, while closed-source flagship models and API services are more suitable as paid entry points. While Alibaba continues to maintain the Qwen open-source ecosystem, it is also meeting higher-level demands through APIs like Plus and Max Preview, embodying this tiered business model.

Domestic Accelerators First Seek Opportunities in the Inference Market

Hardware supply is reinforcing this shift.

Public reports indicate that some restricted NVIDIA chips and servers are facing price pressures due to supply constraints and increased customer demand. More accurately, not all NVIDIA products are experiencing price hikes, but the purchase cost and availability of some high-end or restricted products are influencing the deployment choices of Chinese companies.

Training determines the model's capability ceiling, while inference determines the daily operational costs. High-end training still relies on a mature software and hardware ecosystem, but in inference, private deployments, and specific industry scenarios, customers are more willing to balance performance, cost, and supply security.

If domestic accelerators can provide acceptable stability and inference efficiency, local and hybrid deployments are more likely to make it onto procurement lists. Government and SOE customers, in particular, value data security, compliance, local deployment, and supply chain controllability, providing clearer use cases for domestic computing power such as Huawei Ascend.

However, the increasing cost attractiveness does not mean domestic hardware has completely replaced high-end GPUs. Model migration involves underlying operators, frameworks, caching, scheduling, and deployment tools, and the long-term accumulated developer ecosystem remains a key gap. Domestic accelerators are more likely to initially focus on inference and industry deployment, then gradually expand their applications.

Government and Enterprise Prioritize Security, Private Enterprises Seek a 12 to 18-Month Payback

The payment logic of enterprise customers is also diverging.

Government and state-owned enterprises prioritize data security, compliance audits, on-premises deployment, and long-term supply stability. These requirements will expand the opportunities for domestic software and hardware, but also mean that projects need to go through longer procurement, testing, and acceptance cycles.

Private enterprises, on the other hand, more directly calculate return on investment. Experts point out that many private customers hope to see a clear ROI within 12 to 18 months, including reducing customer service manpower, increasing sales conversion rates, shortening R&D cycles, or lowering operating costs.

Scenarios such as financial services, office productivity, and coding are more likely to be commercialized first because they are data-intensive, have high labor costs, and the effects are relatively easy to quantify. Manufacturing, healthcare, and legal sectors also have demand but need to deal with process reengineering, accuracy, compliance, and responsibility boundaries. Pilot projects moving towards scaled deployment typically take a longer time.

This also means that model ranking lists are difficult to directly translate into enterprise revenue. What customers are ultimately willing to pay for depends on whether the model can reliably integrate into real business operations and deliver quantifiable returns within a limited time frame.

The price war of large models in China is not over, but the competitive landscape has changed. Basic models will continue to decrease in price, while advanced models, privatized deployment, and industry services will face profit pressures; domestic accelerators are gaining more opportunities in the inference market, and DeepSeek has also raised the industry's cost-efficiency standards.

What is truly difficult to replicate is not open-source weights but the system engineering hidden behind the model. Whoever can connect chips, inference efficiency, and enterprise delivery capabilities, and help customers see returns within 12 to 18 months is more likely to convert low-cost traffic into long-term revenue.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia