TL;DR
· On July 13, a panel discussion by Nomura showed that China’s LLM market is shifting from pure price-cutting to a mix of low-pricing for base models and premiums for advanced models.
· DeepSeek’s cost advantage comes from system optimizations such as caching, scheduling, latency, and hardware utilization; releasing open model weights does not equate to replicating operational efficiency.
· Domestic accelerators are gaining more opportunities in inference and localized deployments, but business projects still need to pass ROI validation within 12 to 18 months.

After Nomura’s China internet team exchanged views with experts from a Chinese AI lab on July 13, it offered a judgment that is closer to commercial reality: China’s large-model market is not heading toward lower prices across the board. Instead, it is split into two tiers—base models keep lowering prices to acquire customers, while advanced models, private deployments, and enterprise customization services preserve a premium.

The lab’s proprietary foundation model has already been deployed to more than 100 enterprise customers, and the team is also an early adopter of domestic accelerators such as Huawei Ascend. The key signal released at the expert panel is that while model capabilities are becoming easier to compare, what truly determines platform profits and customer stickiness has expanded from leaderboard rankings to inference costs, deployment efficiency, and enterprise workflows.

This is not a public research report by Nomura, nor does it represent industry-wide statistics. However, it provides an observation lens that is more aligned with enterprise procurement: customers are not buying just a model—they also need to factor in chip prices, per-call costs, system integration, data security, and how long the project takes to break even.

DeepSeek’s low cost is hard because of 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 barrier to use—they do not automatically replicate the operational efficiency of the native platform. What really determines the inference bill is also things like cache hit rate, request scheduling, batching strategies, latency control, and hardware utilization.

The DeepSeek-V3 technical report discloses architectures such as MLA and DeepSeekMoE, and the infrastructure documents cover load balancing and throughput optimization—pointing to the same thing: accomplishing more calls with less hardware usage.

This means that even if platforms like Tencent, Alibaba, and ByteDance can deploy the same open-source weights, they may not achieve the same costs in real business environments. When enterprises make long-term calls—differences of a few milliseconds in latency, a few percentage points in caching efficiency, and in hardware utilization can ultimately translate into substantial bill differences.

Therefore, the competitive pressure brought by DeepSeek is not simply “models that are cheaper.” It is forcing the entire industry to recalculate the actual cost of each token, each call, and each business workflow.

Base models handle customer acquisition; deep deployment makes money

A price war in China’s large-model market is beginning to show segmentation.

For foundation models aimed at developers and lighter needs, commercialization is becoming increasingly high. Prices still face pressure to go further down. Platforms can expand call volume through low pricing—even subsidies—turning the model into an entry point for cloud services and the AI ecosystem.

But when the model enters customer service, financial risk control, code repositories, ERP, CRM, or production scheduling systems, customers are no longer buying just a single API—they are buying a business system that needs to run stably. The deeper the deployment, the higher the switching cost: replacing vendors requires migrating data again, reworking workflows, testing security, and training employees.

This allows model vendors to simultaneously adopt two pricing strategies: cutting prices on base capabilities to acquire customers, while advanced models, industry solutions, private deployments, and tailored delivery shoulder the monetization task.

Open-source and closed-source do not necessarily have to be an either/or choice. Open-source models can attract developers and expand the ecosystem, while closed-source flagship models and API services are better suited as paid entry points. While Alibaba continues to maintain the Qwen open-source ecosystem, it also captures higher-tier demand through API forms such as Plus and Max Preview—reflecting exactly this segmented business model.

Domestic accelerators find opportunities first in the inference market

Hardware supply is reinforcing this shift.

Public reports show that some constrained NVIDIA chips and servers face price pressure due to reduced supply and increased customer demand. More precisely, not all NVIDIA products are rising in price, but the procurement cost and availability of certain high-end or constrained products are affecting Chinese enterprises’ deployment choices.

Training determines the upper limit of model capability, while inference determines daily operating bills. High-end training still relies on a mature software and hardware ecosystem, but for inference, private deployments, and specific industry scenarios, customers are more willing to balance performance, costs, 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 state-owned enterprise customers especially value data security, compliance, localized deployment, and controllable supply chains, giving domestic compute power providers such as Huawei Ascend clearer use cases.

However, rising cost appeal does not mean domestic hardware has fully replaced high-end GPUs. Model migration involves underlying operators, frameworks, caching, scheduling, and deployment tools; the long-term accumulated developer ecosystem remains the key gap. Domestic accelerators are more likely to first enter from inference and industry deployments, and then gradually expand application scope.

For government and enterprises: safety; for private firms: recoup in 12 to 18 months

The payment logic of enterprise customers is also diverging.

Government and state-owned enterprises place more emphasis on data security, compliant auditability, local deployment, and long-term supply stability. These requirements expand opportunities for domestic software and hardware, but they also mean projects must go through longer procurement, testing, and acceptance cycles.

Private companies calculate investment returns more directly. The experts’ phrasing indicates that many private customers hope to see a clear ROI within 12 to 18 months, including reducing customer service headcount, improving sales conversion rates, shortening development cycles, or lowering operating costs.

Scenarios such as financial services, office productivity, and coding are more likely to be commercialized first, because data is intensive, labor costs are higher, and results are relatively easier to quantify. Manufacturing, healthcare, and legal areas also have demand, but they additionally need to handle workflow transformation, accuracy, compliance, and responsibility boundaries—so pilot projects taking the step toward large-scale deployments usually take longer.

This also means that model leaderboard rankings are difficult to directly translate into enterprise revenue. What customers ultimately pay for depends on whether the model can be reliably connected to real business and whether, within limited time, it delivers quantifiable benefits.

China’s large-model price war has not ended, but the way competition happens has changed. Base models will continue to cut prices, while advanced models, private deployments, and industry services must bear profit pressure; domestic accelerators are gaining more opportunities in the inference market, and DeepSeek has also raised industry-wide standards for cost efficiency.

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

Click to learn about Lydon BlockBeats hiring positions

Welcome to join Lydon BlockBeats’ official community:

Telegram subscription group: https://t.me/theblockbeats

Telegram discussion group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia