China’s Open-Source Models Go Global: From High Usage to Higher Value

    Chinese open-source and open-weight models are gaining significant adoption in global enterprise workflows, but high token usage does not necessarily translate into strong revenue. This article examines the forces driving adoption, the limits of API-only monetization, the shift toward workflow and industry solutions, and the metrics enterprises should use to evaluate model value beyond token pricing.

    Global enterprises are changing how they use open-source and open-weight models. These models are no longer limited to research experiments. They are increasingly embedded in customer service, marketing, software development, knowledge management, data analysis, and office automation.

    As usage has grown rapidly, a new industry tension has become increasingly visible: models are being called more frequently, while the revenue generated by each token continues to decline.

    An industry study of the enterprise AI market found that global enterprise AI spending is still expanding. Among US enterprise spending on open-source models, nearly 90% is directed toward Chinese model products. Based on the study’s breakdown, one leading Chinese model accounts for approximately 60% of the share, while two other Chinese model categories account for roughly 10% and slightly more than 10%, respectively.

    These figures indicate that Chinese models are no longer serving only domestic developers. They have achieved meaningful adoption among global developers and enterprise application users.

    However, leading in usage does not necessarily mean leading in commercial returns. From the perspective of enterprise procurement, model services are undergoing a major pricing reset.

    1. Model Usage Is Growing, but Revenue Share Is Not

    One industry trend is particularly notable: the share of open-source models in enterprise token consumption increased from approximately 10% to nearly 20%. At the same time, the spending share associated with these models fell from roughly 6% to about 4%.

    This means enterprises are relying more heavily on open-source models, while the average amount they pay per token is declining. In simple terms, companies are making more calls but are willing to pay less for each one.

    The relationship can be expressed through a basic formula:

    1. Model revenue = Token consumption × Price per token

    If token consumption grows faster than unit prices decline, model revenue may still increase. But if prices fall too sharply, revenue may fail to grow even when usage doubles.

    For model providers, the real challenge is not generating a single API call. It is turning frequent usage into stable and sustainable profit.

    Three main factors are driving the decline in prices.

    1. Inference costs continue to fall

    As quantization, caching, batching, parallel inference, and hardware optimization mature, models of similar size can perform inference at lower cost. Falling costs naturally put downward pressure on market prices.

    2. Differences in general-purpose model capabilities are narrowing

    For tasks such as general question answering, text rewriting, summarization, and basic coding assistance, the user experience gap between models may be smaller than it was in the early stages of the market.

    Enterprises can therefore switch more easily between models and evaluate them based on price, latency, and reliability as well as output quality.

    3. Enterprise procurement is becoming more operationally focused

    Early adopters primarily wanted to know whether a model worked. More mature users continuously calculate the cost of each request, each business workflow, and each employee’s automated output.

    Model providers must demonstrate measurable business value rather than relying only on parameter counts or benchmark scores.

    2. Why Chinese Models Are Achieving Strong Adoption Overseas

    From a technology and product perspective, the growing use of Chinese models in overseas enterprise markets is generally associated with several factors.

    1. Competitive pricing

    For many text generation, classification, summarization, and information extraction tasks, enterprises do not necessarily need the largest available model. If a model meets requirements for accuracy, context length, response speed, and reliability, a lower-cost option can be highly attractive.

    This is especially true for high-volume workloads such as document archiving, ticket classification, first-draft marketing content, and internal knowledge assistants. In these scenarios, a small difference in the price per token can become a significant annual cost difference.

    2. Open ecosystems reduce migration barriers

    When a model offers clear invocation methods, compatibility with common software development kits, and standardized input and output formats, enterprises can migrate at relatively low cost.

    For organizations that already use a unified model gateway, replacing the underlying model often does not require rewriting the entire business system. Teams may only need to adjust routing, authentication, and model configuration.

    As a result, how quickly a model can be integrated into existing systems can matter more than its score on an individual benchmark.

    3. Different models fit different tasks

    Enterprise applications do not need to use one model for every task. Complex reasoning, long-document processing, real-time conversation, batch classification, and code generation all require different capabilities.

    With model routing, enterprises can select models according to task complexity and balance cost against performance.

    A common routing strategy looks like this:

    1. Incoming request
    2. ├─ Simple classification, extraction, rewriting → Lightweight model
    3. ├─ General Q&A and summarization → General-purpose model
    4. ├─ Multi-step reasoning and complex planning → High-capability model
    5. └─ Sensitive data and internal workloads → Private or on-premises model

    This architecture gives low-cost models a large number of usage opportunities. It also explains why some models can achieve very high call volumes while generating relatively little revenue per call.

    3. API-Only Monetization Has a Clear Ceiling

    APIs are the most direct way to commercialize models, but they are not the only option. A business based solely on API sales typically faces four challenges.

    First, pricing is transparent and easy to compare. Enterprises can evaluate input and output prices, concurrency, and response latency across providers. It is difficult for a model service provider to maintain a high premium over the long term.

    Second, customer switching costs are falling. As model gateways, unified invocation layers, and standardized interfaces become more common, the engineering cost of changing models declines. Customer retention increasingly depends on consistent performance and deeper integration with business operations.

    Third, lower inference costs are passed through to selling prices. Improvements in hardware efficiency and model compression will continue to reduce the cost of inference. If a product’s value is limited to returning a model response, its revenue can be heavily affected by price competition.

    Fourth, customer budgets are moving from dedicated model teams toward business units. Enterprises no longer pay only for the model itself. They ask whether it reduces manual review, shortens processing cycles, improves conversion rates, or lowers operating costs.

    An API that cannot be connected to business outcomes may eventually be treated as ordinary infrastructure.

    For this reason, Chinese model providers are exploring broader commercial paths, including managed model services, enterprise deployment, industry applications, AI agent workflows, model fine-tuning, knowledge base development, inference optimization, and security governance.

    4. From Selling Model Capability to Delivering Business Results

    The key change in model commercialization is the gradual shift from charging for individual calls toward charging for capabilities and results.

    1. Model service layer

    This is the foundational layer. It includes online inference, batch inference, model routing, context management, and usage monitoring.

    This layer is suitable for standardized requirements and is typically billed according to input and output tokens, request volume, or resource consumption.

    2. Enterprise platform layer

    When enterprises use models in production, they also need access control, data isolation, audit logs, key management, prompt version control, evaluation systems, and cost analysis.

    Without these capabilities, it is difficult to deploy models reliably in production environments.

    The value of an enterprise platform therefore lies not only in calling models, but also in helping organizations manage the full model lifecycle.

    3. Workflow and agent layer

    In real business processes, models often need to interact with retrieval systems, databases, compute services, and business APIs to complete multi-step tasks.

    For example, a system may first identify a user’s intent, then query an enterprise knowledge base, generate an answer, and finally write the result to a ticketing system. In this process, the model is only the decision-making and generation component within a larger workflow.

    When providers combine model capabilities with business processes, they can move beyond token-based billing and charge by workflow, seat, task volume, or business module.

    4. Industry solution layer

    Industries such as finance, manufacturing, retail, healthcare, education, and government have demanding requirements for data security, compliance, specialized terminology, and process integration.

    A general-purpose model alone cannot solve these problems. Enterprises also need data governance, access controls, knowledge updates, human review, and ongoing quality evaluation.

    This creates room for higher-value services, but it also requires industry expertise, delivery capabilities, and long-term operational support.

    5. How Enterprises Should Evaluate Models Beyond Token Price

    When selecting a model, enterprises should not compare only input and output prices. They should build a complete cost model.

    One useful framework is:

    1. Cost per task =
    2. Input token cost
    3. + Output token cost
    4. + Retrieval and storage cost
    5. + Workflow orchestration cost
    6. + Failed retry cost
    7. + Human review cost

    Enterprises should also evaluate the following dimensions:

    • Effective accuracy: How much of the generated content can be used directly, rather than how much content the model produces.
    • Task completion rate: Whether the model actually achieves the business objective instead of merely returning plausible text.
    • Response latency: Real-time interactions and batch workloads have different latency requirements.
    • Reliability: Whether the system experiences timeouts, throttling, or output quality fluctuations during peak periods.
    • Observability: Whether teams can trace the call chain, identify failure causes, and determine where costs originate.
    • Data security: Whether sensitive data can be isolated and whether logs meet audit requirements.
    • Migration flexibility: Whether changing models or deployment methods would require major system redevelopment.

    In production environments, enterprises should consider using a unified model gateway to decouple the application layer from specific models.

    Applications call an internal standardized interface, while the gateway handles model routing, fallback, rate limiting, caching, and cost tracking. This reduces dependence on a single model and makes it easier to select models dynamically according to task type.

    6. The Next Competitive Frontier: Cost Control and Value Demonstration

    The high call volume achieved by open-source models shows that the market has passed its first validation stage: enterprises are willing to include Chinese models in real business workflows.

    The next stage of competition will shift from “which model is called more often” to “which model creates more value per unit of usage.”

    Model providers need to address three priorities at the same time.

    First, they must continue reducing inference costs so that high-frequency use remains economically viable.

    Second, they must improve reliability on specialized tasks, reducing incorrect outputs, repeated calls, and manual rework.

    Third, they must build products around enterprise workflows, turning models from replaceable interfaces into foundational capabilities that generate continuing business value.

    For enterprises, the most rational strategy is not to pursue the largest model for every workload. Instead, they should build a multi-model architecture: use low-cost models for simple tasks, high-capability models for complex tasks, and private deployments for sensitive scenarios. Continuous monitoring should be used to evaluate both quality and cost.

    Overall, Chinese open-source models have made clear progress in global enterprise adoption. However, market share is only the starting point for commercialization.

    As the price per token continues to decline, business models based solely on call volume and API revenue will face increasing pressure. Model services with lasting competitiveness must move deeper into enterprise data, business workflows, and delivered outcomes.

    The industry is moving from “providing a model” to “helping enterprises complete tasks and achieve measurable returns.”