Chinese Open-Source Models Going Global: From Cheap Token Calls to Enterprise Monetization

    摘要

    Chinese open models are gaining share in overseas developer and enterprise workflows, yet rising token consumption has not produced proportional revenue growth, creating a visible “usage-revenue scissors gap.” This article explains the engineering and commercial forces behind that gap: inference turning into infrastructure, open weights expanding supply, high-volume tasks being low-value ones, and enterprises actively cutting token spend through semantic caching, retrieval narrowing, small-model routing, batching, and context compression. It then covers the engineering metrics that actually decide enterprise deployments (service stability, output consistency, inference efficiency, and data security), and outlines how model vendors can move beyond per-token API sales toward task solutions, model routing, private and hybrid deployment, and outcome-based evaluation. Finally, it presents a practical five-step selection framework based on task boundaries, real-world datasets, total cost per task, replaceability and migration cost, and security and governance requirements. The conclusion: the real competitive moat is not who serves the most tokens, but who delivers more high-value work at lower total cost.

    Introduction

    In recent years, a structural shift has appeared in the global enterprise AI market: a growing share of usage among overseas developers and enterprise workflows is going to certain Chinese models, particularly in open-source model, low-cost inference, and batch task processing scenarios, where they are notably competitive. At the same time, growth in model invocation volume has not produced proportional revenue growth. While the share of tokens consumed keeps rising, the share of spend captured by model providers may be falling. The result is a clear “usage-revenue scissors gap” across the model industry.

    This is not merely the outcome of price competition. It also signals that the enterprise AI market is becoming more segmented. Model capability, inference cost, deployment model, data governance, application integration, and delivery efficiency now jointly determine whether a model vendor can convert traffic into durable commercial value.

    1. Open-Source Models Have Entered Enterprise Workflows, but Revenue Structure Remains Unstable

    From an enterprise procurement perspective, the value of an open-source model is rarely limited to a single question-answer exchange. These models can be embedded in knowledge base retrieval, intelligent customer service, code assistance, document review, data extraction, web content summarization, operational analytics, and automation agents.

    Such applications share three characteristics.

    First, high invocation frequency. Enterprise systems may generate large volumes of short-text processing, classification, summarization, and structuring tasks every day. Individual requests are not complex, but total call counts are substantial.

    Second, strong concurrency requirements. Batch processing usually has to be completed within a short window, which places demands on inference throughput, queueing mechanisms, and resource scheduling.

    Third, high price sensitivity. These tasks typically have a defined cost ceiling. Enterprises care about cost per task, overall budget, and stability far more than about owning the strongest general-purpose model.

    Consequently, rising invocation volume does not translate into equivalent revenue for vendors. The price of many basic tasks falls as inference efficiency improves, and enterprises push costs down further through caching, batching, prompt compression, model routing, and local deployment.

    Commercially, invocation volume is better understood as a coverage metric for how deeply a model has penetrated enterprise processes, not as a final profit metric. What actually determines revenue quality is whether the model participates in critical business processes, and whether the vendor can deliver added value beyond basic inference.

    2. Why Token Volume Grows While Revenue per Unit Keeps Falling

    Model inference is gradually becoming infrastructure

    Once multiple models reach comparable quality on general text generation, summarization, classification, and information extraction, procurement decisions shift from “whose model is strongest” to “whose total cost is lowest.”

    In that environment, tokens stop being merely a unit of model capability and increasingly become a basic compute resource. Pricing is shaped by compute cost, GPU memory utilization, request concurrency, service scale, and scheduling efficiency, and it tends to converge toward actual inference cost.

    Open weights have expanded supply

    When model weights are openly available, enterprises can choose how to host them: a public inference service, their own data center, a general-purpose container environment, or elastic scheduling across multiple providers.

    That flexibility increases enterprise bargaining power. The same model can be served by several providers, basic invocation services commoditize easily, and margins that rely purely on per-token billing get compressed.

    High-growth tasks are often not high-value tasks

    Batch information processing, web content structuring, and basic automation agents generate heavy traffic, but the business value of a single task is not necessarily high. Enterprises are generally willing to pay more for complex reasoning, risk control, specialist decisions, and automation of critical processes, while applying strict cost control to low-value, high-frequency work.

    Vendors can therefore face two outcomes at once: fast-growing invocation volume alongside a steady decline in average revenue per million tokens.

    Enterprises are actively optimizing token consumption

    As enterprise applications mature, teams typically adopt the following measures to cut cost:

    • Use semantic caching for repeated questions, eliminating redundant inference on identical requests;
    • Use retrieval systems to pass only relevant fragments to the model, reducing context length;
    • Use small models for classification, routing, and format conversion, reserving complex models for critical tasks;
    • Apply batch processing to offline workloads to improve hardware utilization;
    • Use structured output to reduce wasted text generation;
    • Compress state across long workflows to avoid re-sending full conversation history at every step.

    These engineering practices improve system efficiency, but they also directly suppress revenue per token.

    3. Low Cost Does Not Mean Low Quality: Engineering Efficiency Becomes the Key Differentiator

    In real enterprise deployments, price is only the first threshold. Whether a model works reliably depends on a whole set of engineering metrics.

    Service stability

    Enterprise workflows usually have explicit latency and availability requirements. An occasional failure can cause customer service backlogs, delayed reviews, or interrupted automation. Providers need capabilities such as rate limiting, retries, failover, traffic isolation, and capacity forecasting.

    Output consistency

    Enterprise systems generally cannot tolerate frequently changing output formats. For tasks like data extraction, contract parsing, and process approval, output must conform to predefined fields, enum values, and validation rules. Model serving therefore needs to be combined with structured parsing, result validation, and fallback mechanisms; natural language prompting alone is not enough.

    Inference efficiency

    At equal task quality, models with lower GPU memory consumption, stronger concurrency, and more stable response latency are easier to deploy at scale. Techniques such as quantization, batching, caching, request merging, and dynamic routing all affect final cost per task.

    Data security

    Enterprises typically need clear data retention periods, access permissions, audit logging, and handling rules for sensitive information. For scenarios involving personal data, trade secrets, or internal knowledge, the deployment model and data boundaries often matter more than the model’s parameter count.

    Low-price competition therefore does not mean simply cutting quotes. Real cost advantage comes from coordinated optimization of model architecture, the inference system, and application engineering.

    4. Model Vendors Should Deliver Complete Capabilities, Not Just APIs

    If revenue depends entirely on token calls, vendors easily fall into three traps: continuous price erosion, low switching costs for customers, and revenue tightly coupled to resource consumption. A more sustainable direction is to offer tiered capabilities built around actual enterprise business needs.

    Move from model invocation to task solutions

    Enterprises buy outcomes, not tokens. For example, a customer wants invoice field extraction, after-sales ticket classification, or internal knowledge Q&A, rather than a block of text.

    Vendors can package model capability into deployable business modules by providing prompt templates, retrieval components, permission control, result validation, workflow orchestration, and runtime monitoring around specific tasks. Billing can then extend beyond call counts to tasks completed, workflow steps, or application subscriptions.

    Move from a single model to model routing

    Different tasks should not be handled by the same model. A system can dynamically select models of different capability tiers based on problem type, latency requirements, data sensitivity level, and budget policy.

    For instance, simple classification can use a lightweight model, complex reasoning a higher-capability model, sensitive data an isolated dedicated instance, and offline batch jobs a higher-throughput execution mode. Model routing balances quality and cost, and reduces customer dependence on any single model.

    Move from public services to private and hybrid deployment

    For some enterprises, data cannot leave the internal environment, or the business requires long-term dedicated resources. In those cases, private deployment, dedicated inference clusters, and hybrid cloud architectures offer greater value.

    However, private delivery is not just handing over model files. It also requires version management, resource monitoring, permission systems, audit logging, canary upgrades, fault diagnosis, and security policies. Only with a complete operations framework does a deployment model become commercially repeatable.

    Move from parameter counts to outcome evaluation

    What enterprises really care about is task success rate, not parameter count. Vendors need business-oriented evaluation covering metrics such as accuracy, recall, factual consistency, format compliance rate, human takeover rate, average response time, and cost per task.

    Cost per task deserves particular attention. A model with a low unit price may still end up more expensive than a pricier but more stable option if it frequently produces errors that require human rework.

    5. Enterprise Selection: A Three-Dimensional Framework of Quality, Cost, and Risk

    Enterprises can evaluate models in the following steps.

    Step 1: Define task boundaries

    Break requirements down into types such as generation, classification, extraction, retrieval, reasoning, and execution, and avoid applying one standard to every task.

    Step 2: Build a realistic dataset

    Test data should come as much as possible from real business traffic and should cover normal samples, boundary cases, long documents, noisy data, and abnormal inputs. Relying only on public examples tends to overstate model performance.

    Step 3: Calculate cost per task

    Beyond token fees, include retrieval, storage, network, monitoring, human review, and failure retries. A general formula can be used:

    Total cost per task = model invocation cost + infrastructure cost + data processing cost + human review cost + retry cost

    This metric is closer to real enterprise spending than comparing token unit prices alone.

    Step 4: Assess replaceability and migration cost

    Enterprises should confirm whether a model supports standardized invocation, structured output, and multi-model switching, so that business logic is not tightly coupled to a single service. A well-designed abstraction layer lowers the future cost of swapping models.

    Step 5: Set security and governance requirements

    Define whether data is used for training, how logs are retained, who can access call records, how sensitive information is masked, and how regression testing is performed after model upgrades.

    6. The Future: From Model Rankings to Value Density

    As the supply of foundation models grows, invocation volume and low prices alone will hardly form a lasting moat. The more important metric going forward may be how much business value a model creates per call.

    On one hand, model vendors must continue improving inference efficiency and lowering the adoption barrier for enterprises. On the other, they need to move into deeper layers such as data governance, knowledge retrieval, workflow orchestration, intelligent agents, and application operations.

    For enterprises, open-source models bring greater choice but also greater complexity in architecture design and governance. The optimal approach is usually not selecting a single model, but building observable, switchable, evaluable model infrastructure.

    Overall, the relatively high usage share Chinese models have gained in overseas enterprise markets demonstrates competitiveness in cost control, engineering deployment, and large-scale concurrent workloads. Moving from “widely invoked” to “sustainably profitable” still requires a business model upgrade: from selling basic APIs to delivering verifiable task outcomes, complete deployment capabilities, and stable enterprise-grade service.

    The endpoint of this competition is not which model serves more tokens, but which one helps enterprises continuously complete more high-value work at lower total cost.