Domestic Foundation Model Competition Enters a New Phase of Layered Capabilities and Real-World Deployment

    As the number of AI models continues to grow, competition in China’s generative AI market is shifting from model launches and parameter counts toward capability tiers, industry adaptation, engineering efficiency, and measurable business value. This article explains how foundation models and industry-specific systems fit together, why model capability alone does not guarantee business results, how to control deployment costs, and how enterprises can build a practical path from pilot projects to production use.

    In the early stages of generative AI, market attention focused on the number of available models, parameter size, and multimodal capabilities. As model offerings multiplied, enterprises and users began to recognize that more model names do not necessarily represent equivalent improvements in technical capability.

    The value of a large model is usually determined less by the number of promotional concepts surrounding it than by whether its core capabilities, industry adaptation, engineering costs, and commercial delivery can form a complete loop.

    From this perspective, China’s large-model market is entering an important transition. Competition is moving away from the speed of model releases and toward capability tiers, application validation, and the efficiency of resource allocation.

    1. From “Hundreds of Models” to Layered Capabilities

    AI models in the current market can broadly be divided into two categories based on the source of their capabilities and their application scope.

    The first category is the foundation model, which provides general-purpose capabilities. These models typically support text understanding, content generation, reasoning, code processing, or multimodal analysis. Through continued training, prompt optimization, knowledge augmentation, and tool calling, they can be adapted to different business scenarios.

    The second category is the industry-specific or task-specific model. These models are optimized for fields such as finance, manufacturing, energy, healthcare, government services, and education. Their goal is not to handle every general-purpose task, but to improve accuracy, stability, and controllability in a specific business domain.

    These two categories are not substitutes for each other. Industry models are often built on top of foundation models and adapted through domain data, task instructions, retrieval systems, and business rules. If a foundation model is viewed as a general-purpose computing base, an industry model is closer to the solution layer designed for a particular business.

    A simplified capability hierarchy looks like this:

    1. Application layer: Intelligent customer service, knowledge Q&A, office assistance, coding assistants
    2. Task layer: Contract review, equipment diagnosis, report generation, risk identification
    3. Industry layer: Models for finance, manufacturing, healthcare, government services, and other domains
    4. Foundation layer: General language, multimodal, reasoning, and code models
    5. Infrastructure layer: Compute, storage, networking, data processing, and monitoring

    The challenge is that market participants do not use consistent definitions for these layers. Some products add an industry knowledge base to a general-purpose model and then market the result as an “industry large model.” Other systems add workflow orchestration and automation capabilities and present them as a new model.

    Unclear boundaries make it difficult for buyers to determine whether they are purchasing model capability, data services, or an application system.

    When evaluating a solution, enterprises should look beyond the product name and confirm the following:

    1. Whether the underlying model provides stable general-purpose capabilities;
    2. Whether industry data has been cleaned, de-identified, and evaluated for quality;
    3. Whether the model genuinely participates in reasoning or mainly performs text generation;
    4. Whether the system supports retrieval, tool calling, and access control;
    5. Whether results can be validated through evaluation datasets and business metrics.

    2. Model Capability Does Not Equal Business Value

    Technical benchmarks can describe how a model performs on certain public tasks, but they cannot directly represent the benefits of an enterprise project. Business users are more concerned with whether a system can reduce manual work, lower error rates, shorten processing time, and provide traceability and correction mechanisms when problems occur.

    For example, a knowledge Q&A system may generate fluent responses, but it remains risky in production if it cannot correctly cite internal documents or identify version differences within the knowledge base. Similarly, a code-generation system that produces plausible code but cannot pass tests, follow internal standards, or understand an existing architecture is unlikely to deliver meaningful improvements in developer productivity.

    Large-model applications therefore need to be evaluated at the system level rather than only at the model level. Overall capability can be divided into four areas:

    • Model capability: Understanding, generation, reasoning, classification, and multimodal processing;
    • Data capability: Data quality, knowledge updates, access isolation, and traceability;
    • Engineering capability: Service stability, concurrent processing, caching, monitoring, and failure recovery;
    • Business capability: Workflow integration, role collaboration, result review, and value measurement.

    Only when all four areas are in place can a model move from a demonstration system into formal business operations.

    3. High Costs Are an Essential Deployment Consideration

    The cost of generative AI does not come only from model training. Real-world projects usually also involve data preparation, training and fine-tuning, inference, storage, network transmission, system development, monitoring and auditing, and manual review.

    During training, the main costs include compute resources, data processing, and experimental iteration. Larger models, larger training datasets, and more frequent experiments generally require more resources.

    During inference, costs are closely related to request volume, input and output length, concurrency, and response-time requirements. For enterprise applications, inference costs may deserve more attention than one-time training costs because they continue to grow with user and business volume.

    A practical cost estimate can be expressed as follows:

    1. Total cost =
    2. Fixed infrastructure costs
    3. + Model training and fine-tuning costs
    4. + Inference costs that grow with usage
    5. + Data governance and security costs
    6. + Application development and operations costs
    7. + Manual review and exception-handling costs

    Engineering teams can reduce costs through the following measures:

    1. Differentiate task complexity
      Simple classification, keyword extraction, and format conversion do not require the most capable model. Higher-tier models should be reserved for complex reasoning and high-risk tasks.

    2. Control context length
      Use retrieval, summarization, and segmented processing to remove irrelevant content and avoid sending large amounts of historical information with every request.

    3. Introduce caching
      Cache repeated questions, stable knowledge, and common templates to reduce unnecessary repeated inference.

    4. Use asynchronous processing
      For report generation, batch analysis, and other tasks that do not require immediate responses, queues and background jobs can reduce peak pressure.

    5. Establish model routing
      Route requests to models with different capability levels based on task type, data sensitivity, and latency requirements.

    6. Set up quality safeguards
      Add rule-based validation, human review, or secondary verification for high-risk results instead of relying solely on model output.

    4. User Growth Ultimately Depends on Actual Usage Frequency

    Large-model products can attract many trial users through novelty in their early stages. However, trial volume does not equal long-term activity or enterprise revenue. Continued use depends on whether the model can be embedded into stable workflows and consistently solve specific problems.

    Factors that influence retention commonly include:

    • Output stability;
    • Integration with existing enterprise data;
    • Support for access control and auditing;
    • Compatibility with existing systems;
    • Whether results can be acted upon directly;
    • How easily errors can be corrected;
    • Whether the cost of use is proportional to the benefits.

    This means that large-model applications should not be designed only as chat windows. More effective products generally embed models into business workflows. Examples include automatically categorizing issues in a ticketing system, generating retrieval-based answers in a knowledge management system, assisting with test creation in the development process, and producing structured analysis in an operations platform.

    From a product design perspective, models should gradually shift from being merely “conversation entry points” to becoming “task components.” Users may not need to know which model is operating behind the scenes, but they should clearly experience simpler workflows, faster access to information, and less repetitive work.

    5. How Enterprises Should Choose Their Investment Priorities

    In a complex market, enterprises should not pursue every emerging concept at the same time. A more practical approach is to identify the layer they occupy and then determine where to invest.

    Organizations with large volumes of high-quality data, long-term compute budgets, and dedicated model development teams can consider building foundation models or general-purpose capabilities. This route requires a long development cycle and substantial funding, as well as continuous evaluation and long-term iteration.

    Organizations with strong industry resources and clearly defined business workflows may be better positioned to start with industry models or scenario-based applications. In such cases, the priority is not necessarily to retrain an extremely large model, but to improve data governance, knowledge retrieval, workflow orchestration, and outcome evaluation.

    Organizations that mainly consume models should focus on the following questions:

    • Does the model meet data security requirements?
    • Does it support private or isolated deployment?
    • Can it integrate with existing business systems?
    • Does it provide observability and auditing?
    • Can returns be evaluated using business metrics?
    • Are human and rule-based safeguards available when model performance is unstable?

    A practical deployment path usually has four stages.

    Stage 1: Identify High-Value Scenarios

    Start with tasks that are repetitive, relatively well-defined, supported by available data, and easy to measure. Examples include document summarization, knowledge retrieval, ticket classification, and information extraction.

    High-risk scenarios that require fully automated decision-making should not be the starting point.

    Stage 2: Build an Evaluation Framework

    Prepare representative test datasets covering normal inputs, edge cases, incorrect information, and malicious inputs. Evaluation should not rely on accuracy alone. It should also include response time, citation completeness, refusal quality, stability, and the percentage of outputs requiring manual modification.

    Stage 3: Build the Engineering System

    Connect model services with data permissions, logs, monitoring, caching, message queues, and business interfaces to create a complete operational chain.

    The system should also record inputs, outputs, model versions, and human corrections. This information provides the basis for future improvements.

    Stage 4: Expand Gradually

    Only after a small-scale application has demonstrated value and reached acceptable levels of security, cost, and stability should the enterprise expand its user base and business scope.

    During expansion, teams should continue monitoring inference costs, failure rates, and actual usage frequency. Blindly scaling a system can create unnecessary resource waste.

    6. Competition Will Ultimately Return to Engineering and Value

    The concepts used in the large-model market may continue to change, but the core questions of technical competition will remain largely the same:

    • Is the model reliable?
    • Is the data usable?
    • Are costs controllable?
    • Is the system stable?
    • Does the business actually benefit?

    For developers, future capabilities will involve more than prompt writing. They will also include data processing, retrieval-augmented generation, model evaluation, service deployment, access management, monitoring and alerting, and troubleshooting.

    For enterprises, a large model should not be treated as a black box that creates value independently. It should be integrated into the broader digital architecture.

    The general capabilities covered by technology service providers such as Baidu and Baidu AI Cloud—including compute, storage, container platforms, message queues, logging, and monitoring and alerting—represent infrastructure categories that enterprises may need to evaluate when building AI applications. Specific choices should still be based on data sensitivity, business scale, deployment model, and compliance requirements.

    Overall, generative AI is moving from competition based on model quantity toward competition based on layered capabilities and measurable value. The participants most likely to remain successful over the long term will not necessarily be those with the most concepts or the fastest model release cycles. They will be the teams that can consistently solve user problems, control operating costs, and embed model capabilities reliably into business workflows.