LLM Valuation in China: From Narrative to Operating Validation

    This article explains why valuations of Chinese LLM companies are shifting from model scale, benchmark scores, and funding narratives toward operating validation. It examines usage-based revenue as a signal of demand quality, the gap between revenue growth and profit or cash flow, compute commitments as a debt-like risk, unit economics, and a five-dimension framework for investors and enterprise buyers.

    1. Valuation Logic Is Undergoing a Fundamental Shift

    In the early stage of the generative AI industry, the market typically judged a company’s potential by model size, training data volume, context length, public benchmark scores, and funding amounts. That approach was appropriate when technical routes had not converged and business models were not yet mature. As model capabilities become more widespread, however, technical narratives alone are no longer enough to support long-term valuations.

    The reason is that LLM capabilities are converging rapidly. Basic language understanding, text generation, coding assistance, knowledge Q&A, and multimodal interaction have gradually moved from being distinctive advantages of a few companies to common capabilities across the industry. The real differences between companies are now beginning to appear in the following areas:

    1. Whether model capabilities can be converted into stable revenue;
    2. Whether revenue comes from real, ongoing customer usage;
    3. Whether each invocation has a reasonable gross margin;
    4. Whether compute and infrastructure investment is controllable;
    5. Whether customer migration costs and business stickiness are high enough;
    6. Whether operating cash flow can cover ongoing R&D and expansion needs.

    This means that valuations of model companies are shifting from “what might be achieved in the future” to “what has already been validated today.”

    2. Usage-Based Pricing Becomes an Important Window into Revenue Quality

    From a business model perspective, revenue for model companies can be broadly divided into subscription revenue, project delivery revenue, platform service revenue, and usage-based revenue.

    Subscriptions help create stable recurring revenue, but if users do not actually use the product frequently after purchase, the subscription amount does not fully represent real demand. Project delivery can bring larger orders, but it often involves long recognition cycles, high delivery costs, and significant revenue volatility. By contrast, usage-based pricing more directly reflects whether customers are actually embedding models into business processes.

    Usage-based pricing is usually tied to input data volume, output data volume, number of invocations, inference duration, or resource consumption. Its value lies in creating a clearer correspondence between revenue growth and customer business activity. For example, an increase in customer service system invocations may indicate growth in service requests; an increase in code generation invocations may show that a development team has incorporated models into daily development; an increase in document review invocations may reflect that the model has entered a company’s compliance or operations workflow.

    However, usage revenue is not inherently high-quality revenue. Further observation is needed to assess:

    • Whether usage is concentrated among a small number of large customers;
    • Whether customers are conducting short-term tests or one-off batch invocations;
    • Whether invocation growth depends on substantial price cuts;
    • Whether model output requires extensive manual review;
    • Whether customers have long-term renewal and expansion intentions;
    • Whether revenue per invocation can cover inference, storage, network, and operations costs.

    Therefore, companies should disclose total revenue, active customer count, customer retention rate, repeat usage rate, average contract value, and unit inference cost together. Only by combining revenue with usage behavior can the market determine whether growth comes from real demand or from temporary subsidies and marketing spending.

    3. Revenue Growth Cannot Replace Profit and Cash Flow Validation

    LLM companies generally face a particular problem: revenue may grow quickly, but profit improvement may lag significantly.

    This is because the cost structure of the model business differs from traditional software. A single model invocation involves not only server resources but potentially also depreciation of high-performance computing equipment, data center energy consumption, network transmission, storage, model scheduling, engineering maintenance, and security auditing. If companies continue to cut prices to expand market share, revenue growth may even be accompanied by declining gross margin.

    In financial analysis, at least three profit measures should be distinguished:

    1. Gross Profit

    Gross profit mainly reflects what remains after direct delivery costs are deducted from model service revenue. For model companies, direct costs usually include inference compute, storage, network, and resource consumption directly related to business delivery.

    Gross margin can answer a basic question: for each additional unit of model invocation, does the company have an opportunity to earn a positive contribution?

    2. Adjusted Operating Profit

    Adjusted operating profit usually excludes certain non-cash items, equity compensation, or one-time expenses, which helps observers assess trends in the core business. However, this measure cannot fully replace standard accounting measures, because excluded items may still affect shareholder value and cash spending.

    3. Operating Profit Under Standard Accounting Principles

    This measure is closer to the company’s actual financial responsibilities and can reflect depreciation, R&D investment, personnel costs, and other operating expenses. If a company is already profitable on an adjusted basis but still loss-making under standard accounting principles, it usually indicates that the company remains in a high-investment stage, or that its cost recognition has not yet fully reflected the long-term burden of expansion.

    Therefore, the market should not focus only on whether a company is profitable. It should also examine where profit improvement comes from: higher prices, improved invocation efficiency, higher hardware utilization, or cuts to R&D and marketing. The first two are usually more sustainable, while the latter two may damage long-term competitiveness.

    4. Compute Commitments Are an Easily Overlooked Debt-Like Factor in Valuation

    Contractual obligations and infrastructure commitments of model companies often reflect future operating pressure more clearly than current profit.

    To secure training and inference resources, companies may lock in computing resources, data center capacity, network bandwidth, and related equipment in advance. Such commitments can reduce business risks caused by resource shortages, but they also create strong fixed-cost constraints. If customer growth falls short of expectations, the company may still need to pay for the resources it has already committed to.

    When evaluating such commitments, it is not enough to look only at the amount. Their structure also needs to be analyzed:

    • Whether the commitment period is short-term or multi-year;
    • Whether there are flexible adjustment or exit mechanisms;
    • Whether resources can be transferred, subleased, or used across businesses;
    • Whether payment obligations are tied to actual invocation volume;
    • Whether the scale of commitments matches future revenue forecasts;
    • Whether resource utilization has obvious seasonal fluctuations.

    If a company’s compute commitments are far higher than its validated revenue capacity, the corresponding execution risk must be incorporated into the valuation. Conversely, if resource procurement is highly flexible and can expand in step with customer usage, fixed-cost pressure will be relatively manageable.

    From a technical management perspective, more compute is not inherently better. What matters is how much effective revenue each unit of compute generates and whether unit inference cost can continue to fall at the same service quality.

    5. Model Companies Need to Build Unit Economics

    Traditional internet businesses often use customer acquisition cost, customer lifetime value, and retention rate to measure business quality. LLM companies need to go further and build unit economics centered on invocations.

    A simplified unit contribution can be expressed as:

    1. Unit contribution = Unit invocation revenue
    2. - Inference compute cost
    3. - Storage and network cost
    4. - Scheduling and operations cost
    5. - Security and manual review cost

    If unit contribution remains negative for a long time, the company can only expand scale through financing subsidies. If unit contribution is positive but very low, the company needs scale growth and infrastructure optimization to improve overall profit. Only when unit contribution can expand steadily can the company potentially generate healthy cash flow while growing revenue.

    Technical teams can improve unit economics in the following directions:

    1. Model tiering
      Assign complex tasks to high-capability models and simple tasks such as classification, extraction, and rewriting to lower-cost models, avoiding the use of the highest-spec resources for every request.

    2. Caching and result reuse
      Cache repeated questions, fixed templates, and stable knowledge content to reduce repeated inference consumption.

    3. Context compression
      Summarize, chunk, and filter historical conversations, retrieved materials, and long documents by relevance to reduce ineffective input.

    4. Batch processing
      Schedule tasks with low latency requirements in batches to improve resource utilization.

    5. Dynamic routing
      Select different inference paths based on task complexity, response latency, and cost budget.

    6. Quality monitoring
      Monitor not only API availability and response speed but also answer adoption rate, manual modification rate, and task completion rate.

    These optimizations should ultimately be reflected in financial metrics, not remain only in technical reports.

    6. How Chinese Model Companies Can Move Beyond a Single Narrative

    For Chinese model companies, market competition has shifted from “whether they can train a model” to “whether they can continuously create value in real industries.”

    On one hand, companies need to reduce one-sided promotion of parameter scale, benchmark rankings, and funding amounts, and instead disclose more verifiable operating metrics, such as effective invocation volume, paid customer retention rate, revenue share from key industries, unit inference cost, and customer usage depth.

    On the other hand, companies need to avoid packaging all revenue as platform revenue. Model services, industry projects, technical consulting, and manual delivery should be clearly distinguished, because different revenue types differ greatly in replicability, gross margin, and cash collection cycle.

    For enterprise customers, procurement of model services should also not compare unit price alone. A more reasonable evaluation approach includes:

    • Whether business task completion rates improve;
    • Whether manual review workload decreases;
    • Whether system response meets production requirements;
    • Whether data isolation and access control are reliable;
    • Whether model services can connect to existing systems;
    • Whether migration costs and long-term lock-in risks are acceptable;
    • Whether total usage costs are lower than traditional processes.

    Only when models are truly embedded in business processes and generate stable productivity gains will model companies’ revenue have stronger persistence.

    7. Building a Comprehensive Scoring Framework for Investment and Procurement

    To avoid being misled by a single metric, a model company can be evaluated from five dimensions:

    Commercialization Capability

    Focus on revenue scale, revenue growth, number of paying customers, customer retention, and the share of usage-based revenue. The key is not growth speed itself, but whether growth is repeatable and predictable.

    Technical Efficiency

    Focus on cost per invocation, resource utilization, service stability, inference latency, and model quality. Technical capability should ultimately translate into higher gross margin or stronger customer stickiness.

    Financial Health

    Focus on gross profit, operating profit, cash flow, R&D expense ratio, and the difference between adjusted measures and standard accounting measures.

    Infrastructure Risk

    Focus on compute commitments, resource procurement flexibility, equipment depreciation pressure, and future capacity requirements. Infrastructure investment should match the pace of revenue validation.

    Customer Value

    Focus on whether the model solves high-frequency, essential, and quantifiable problems. Applications that directly reduce costs, increase revenue, or lower risk usually have more commercial resilience than purely demonstrative applications.

    Conclusion: From Storytelling to Delivering Results

    The LLM industry will not lose growth space simply because valuation logic becomes more rational. On the contrary, as metrics such as revenue, profit, cash flow, and compute efficiency receive more attention, the market is moving from the concept validation stage to the industry validation stage.

    In the future, model companies with true long-term value may not be those with the loudest promotion, the largest parameter scale, or the highest funding amounts. They are more likely to be companies that can consistently answer the following questions:

    • Why are customers willing to pay over the long term?
    • Does each invocation create positive economic value?
    • Can compute investment be absorbed by real demand?
    • Can technical upgrades improve the profit structure?
    • Can the business grow independently without continuous subsidies?

    When these questions are supported by data and business results, the valuation of model companies will no longer rely on a single narrative. It will gradually be built on an operating foundation that is verifiable, replicable, and sustainable.