Chinese Open-Source AI Models Going Global: From Call Volume to Enterprise Profit

    This article examines why the fast-growing use of open-source and open-weight AI models — especially Chinese models — has not translated into proportional enterprise revenue. It analyzes the "high call volume, low revenue" gap, the engineering-efficiency advantages behind Chinese models' adoption, and the shift from selling APIs to selling end-to-end solutions. It also outlines practical evaluation frameworks for enterprises and three key thresholds model vendors must cross to build sustainable businesses.

    In recent years, the adoption of open-source and open-weight models among enterprises worldwide has accelerated noticeably. A growing number of companies no longer treat a model as just a conversational interface; they embed it directly into production workflows such as search, customer service, coding assistance, data extraction, process automation, and intelligent agents. Chinese models are entering the workflows of more and more overseas developers and enterprises, thanks to their strong inference cost-effectiveness, rapid iteration, and ability to handle batch tasks.

    One phenomenon deserves attention: as model call volume keeps rising, the amount enterprises actually pay does not necessarily grow at the same pace. In other words, models are getting more usage without necessarily earning profit proportional to that scale. This “high call volume, low revenue” scissors gap reflects both the commercial pressure facing open-source models and a broader shift in AI infrastructure from high-premium competition toward scale- and cost-based competition.

    1. By market share, open-source model competition has entered deep waters

    An enterprise AI market report from an industry organization shows that global enterprise AI spending is still growing rapidly. Among US enterprises’ spending on open-source models, Chinese model products account for a high proportion, with leading models capturing most of the budget.

    This result shows at least three things.

    First, enterprise customers no longer choose models purely by origin. For production environments, they care more about whether a model can reliably complete tasks, whether it offers continuous service capability, and whether the cost per task is controllable. As long as a model strikes a reasonable balance among accuracy, latency, availability, and price, it has a chance to enter real enterprise workflows.

    Second, the open-source or open-weight model lowers the barrier to trial. Enterprises can quickly validate through public APIs, and can also choose private, dedicated-network, or hybrid deployment based on compliance requirements. Compared with relying entirely on a single closed API, this flexibility better fits large enterprises’ requirements for data security, business continuity, and supply-chain controllability.

    Third, model competition has shifted from a pure parameter-scale race to a competition of overall engineering capability. The model itself is only the foundation. What truly determines whether an enterprise keeps using it long-term also includes supporting capabilities such as inference services, elastic scaling, logging and auditing, version management, failover, evaluation systems, and data governance.

    Therefore, growth in market share does not mean the business model is already mature. A model entering an enterprise workflow is only the first step; moving from “being used” to “being continuously purchased” still requires solving problems of value proof and revenue conversion.

    2. Why token consumption growth doesn’t necessarily bring revenue growth

    In large model services, a Token can generally be understood as the basic unit of computation a model uses when processing text or multimodal information. The larger an enterprise’s call volume, the more tasks the model typically participates in. But call volume does not equal revenue, for reasons that mainly come from the following.

    1. Basic tasks are naturally low-priced

    Tasks vary greatly in complexity when enterprises use models.

    For example, long-chain reasoning, complex planning, professional analysis, and high-risk decision support often require stronger model capability, longer context, and higher stability. These tasks are more sensitive to accuracy, and enterprises are willing to pay more.

    In contrast, tasks like web content organization, batch text classification, field extraction, summary generation, simple Q&A, and data cleaning typically feature large scale, high repeatability, and relatively large tolerance for error. Enterprises will prioritize lower-cost models and further reduce per-call cost through batching, caching, and concurrency scheduling.

    When Chinese models enter the latter category of tasks in large numbers, Token counts may grow rapidly, but the per-Token price will be markedly lower than for complex reasoning tasks. As a result, a natural gap emerges between call-volume growth and revenue growth.

    2. Open weights bring stronger price competition

    Once model weights are open, enterprises can choose different inference providers, and can also deploy on their own compute or general cloud resources. The same model may be hosted across multiple service channels, lowering enterprises’ switching costs.

    When model capabilities converge, competition among providers shifts to:

    • Cost per unit of inference;
    • Concurrency handling capacity;
    • Time to first token and overall response time;
    • Stability during peak periods;
    • Deployment and operations complexity;
    • Data isolation and compliance capability.

    In this environment, basic model calls gradually resemble infrastructure services. Providers can attract large volumes of traffic but find it hard to maintain high per-unit gross margins.

    3. Enterprises prefer hybrid model architectures

    Many enterprises do not assign all tasks to a single model; instead, they adopt a tiered routing strategy:

    • Simple tasks use low-cost models;
    • Medium-complexity tasks use general-purpose models;
    • High-value tasks use stronger reasoning models;
    • Sensitive data uses models in local or dedicated environments.

    This architecture can significantly reduce total cost, but it also means low-price models handle a large share of Token consumption. Call-volume figures in the statistics may therefore look impressive, yet high-value tasks that actually generate high revenue may not grow at the same rate.

    3. Chinese models’ core advantage lies first in engineering efficiency

    Low price alone cannot fully explain Chinese models’ expansion among overseas enterprises. Production environments rarely adopt an unstable model long-term just because it is cheap. A model that keeps entering real workflows usually also needs strong engineering adaptability.

    1. Inference cost optimization

    The cost of model services is made up mainly of compute, GPU memory, network, storage, and operations. Providers usually reduce per-unit inference cost through methods such as model compression, quantization, caching, batching, and scheduling optimization.

    Among these, batching combines multiple requests for processing to improve hardware utilization; caching avoids repeated computation; quantization reduces the GPU memory needed to run the model; and dynamic scheduling allocates resources based on request priority and resource status.

    These techniques do not necessarily directly improve a model’s theoretical capability, but they significantly affect whether an enterprise is willing to scale up usage. For enterprises processing millions or more requests per day, even a small reduction in per-unit cost can add up to substantial cumulative savings.

    2. Task-oriented model adaptation

    A general-purpose model does not mean it fits every task. What enterprises truly care about is the effective completion rate in specific business contexts.

    For example, data extraction tasks require stable structured output; customer-service tasks require controlling hallucinations and tone; coding-assistance tasks require correct formatting; and enterprise search requires strong citation and context-understanding capabilities.

    Therefore, model providers need to design prompt templates, output constraints, retrieval augmentation, tool calling, and result validation around tasks. Only by turning model capability into executable business processes can Token calls be converted into stable revenue.

    3. Adaptation to batch-automation scenarios

    Intelligent agents and automation workflows typically generate large numbers of parallel requests. For example, a task may include multiple steps such as web page reading, content filtering, field extraction, result comparison, and final aggregation. The value of a single step is low, but the whole can replace a large amount of manual work.

    Chinese models’ advantage in these scenarios often comes from lower overall cost and faster delivery. It is worth noting that batch automation places high demands on model stability; any low-probability error, once amplified at scale, can cause serious impact. So providers also need to offer retry, timeout control, exception fallback, and human review mechanisms.

    4. From “selling APIs” to “selling solutions”

    Simply providing a model API easily leads to price competition. A more sustainable path to commercialization in the future may be to provide complete capabilities around enterprise tasks rather than charging only by call volume.

    1. Building a model service layer

    A model service layer typically includes the following modules:

    • A unified calling gateway;
    • Identity authentication and permission management;
    • Request routing and model selection;
    • Rate limiting, retry, and circuit breaking;
    • Token statistics and cost accounting;
    • Logging and auditing;
    • Quality evaluation and version management.

    With a service layer, enterprises can shield themselves from differences among underlying models and switch between and compare multiple models. Providers can also gain a better grasp of customers’ real usage scenarios and offer more valuable optimization services.

    2. Building a knowledge and data layer

    The effectiveness of an enterprise’s model application often depends not only on the model itself but also on whether internal knowledge can be accurately retrieved and cited.

    So enterprises need to build capabilities such as document parsing, vector retrieval, permission filtering, knowledge updates, and citation tracking. For highly sensitive business, they also need to ensure that different departments and roles can only access information within their authorized scope.

    At this layer, model providers can offer general technical services around knowledge-base building, retrieval augmentation, data cleaning, and effectiveness evaluation, rather than being limited to basic API calls.

    3. Building a business process layer

    Only when a model is embedded into business processes can it produce measurable value. For example, in contract review, the system must not only identify risky clauses but also write the results into the approval workflow; in customer service, the system must not only generate answers but also link tickets, escalate to humans, and record outcomes.

    So what enterprises purchase in the future may no longer be a single model, but an integrated solution of “model plus data plus process plus operations.” Its billing may also gradually expand from pure per-Token charging to combined charging based on task volume, processing results, service levels, or overall subscription.

    5. How enterprises evaluate a model’s real value

    When choosing a model, enterprises should not look only at call volume, parameter scale, or public leaderboards, but should build a complete evaluation system.

    1. Evaluate task success rate

    Test sets should be built based on actual business, focusing on:

    • Accuracy rate;
    • Completion rate;
    • Format compliance rate;
    • Citation accuracy rate;
    • Tool-calling success rate;
    • Human-review pass rate.

    Different businesses weight these metrics differently, and a single score cannot replace the entire evaluation.

    2. Evaluate cost per unit task

    Token price is only part of the cost. Enterprises should also calculate:

    • Average input and output length;
    • Retry rate;
    • Concurrency resource consumption;
    • Data preprocessing cost;
    • Human review cost;
    • System operations cost;
    • Losses from failed tasks.

    What ultimately matters is the “overall cost of completing one effective task,” not the quoted price of a single Token.

    3. Evaluate stability and portability

    Models in production environments need to run stably over the long term. Enterprises should observe service availability, latency fluctuations, rate-limiting policies, version changes, and failure recovery capability.

    At the same time, enterprises should avoid binding business logic entirely to the special output format of a particular model. A unified gateway, an abstract calling protocol, and standardized evaluation can lower the cost of switching models in the future.

    6. Three thresholds model vendors need to cross

    The first threshold is moving from “cheap and usable” to “stable and usable.” Low prices can bring trial traffic, but only stability, accuracy, and service guarantees can drive customers to expand their purchases.

    The second threshold is moving from “basic calls” to “business results.” What enterprises ultimately buy is not Tokens, but faster information processing, lower labor costs, and higher process efficiency. Model vendors must understand customer businesses and establish quantifiable value metrics.

    The third threshold is moving from “model capability” to “platform capability.” Future competition will span the full chain of data governance, model routing, deployment management, monitoring and alerting, permission control, and security auditing. Providers with platform-level capability are more likely to break free from pure price competition.

    For enterprises, general infrastructure offered by cloud platforms such as Baidu AI Cloud — including object storage, container platforms, message queues, logging services, and monitoring and alerting — can serve as building blocks when constructing a model application platform. The actual architecture should still be chosen based on data sensitivity, business latency, compliance requirements, and cost goals, avoiding the mistake of overlooking overall system governance in the pursuit of model capability.

    Conclusion

    The high usage rate of Chinese models among overseas enterprises shows that they already possess strong engineering competitiveness and scenario adaptability. But call scale is only the starting point of commercialization, not the endpoint. As basic model prices keep falling, enterprises will pay more attention to task completion effectiveness, system stability, and overall input-output ratio.

    Truly competitive model services in the future will not stop at providing an API; they will help enterprises complete the full loop from data ingestion and model selection to process orchestration and result delivery. Only by converting low-cost calls into verifiable business value can model vendors break out of the “high usage, low revenue” predicament and build a more solid long-term business model in the global enterprise AI market.