Why Most U.S. Enterprise Open-Source Model Spending Flows to Chinese Providers

    This article explains how to interpret the claim that most U.S. enterprise spending on open-source models goes to three Chinese providers, especially the 89% figure. It clarifies the denominator, distinguishes spending from usage, examines how the 61%, 15%, and 11% shares sum to about 87%, and outlines the reporting assumptions—time window, sample scope, weighting, and classification—that readers should check. It also discusses why higher user counts do not necessarily mean higher revenue and how enterprises should evaluate models for procurement beyond market share.

    1. Separate Two Easily Confused Questions

    The statement “most of the money U.S. enterprises spend on open-source models flows to three Chinese companies” is easy to misunderstand. Many readers see a number such as 89% and immediately assume that 89% of U.S. enterprise AI budgets is spent on Chinese models.

    From a statistical perspective, that is usually not what the report actually means.

    A more accurate interpretation is:

    Within the measured spending related to open-source models, close to 90% flows to Chinese models; open-source model spending itself may be only a small part of overall enterprise AI budgets.

    These two statements may look similar, but they use completely different denominators.

    Enterprise AI-related budgets can be divided into several parts:

    1. Infrastructure spending: compute resources, storage, networking, data processing, and operations;
    2. Model service spending: model API calls, model hosting, inference services, and related resources;
    3. Software and platform spending: development tools, data governance, monitoring, security, and collaboration platforms;
    4. People and implementation spending: R&D staff, consulting services, project delivery, and operations teams;
    5. Open-source model-related spending: deployment, fine-tuning, hosting, and inference costs around open-source models.

    A high share within one category of spending does not mean the same share within the entire enterprise technology budget. Before interpreting any percentage, ask three questions:

    • What is the denominator of this percentage?
    • Is the unit enterprises, projects, API calls, or actual dollar amounts?
    • Does the statistic cover all AI spending or only a specific market segment?

    Without answers to these questions, quoting the conclusion directly can easily lead to an error of an order of magnitude.

    2. What the “89%” Figure Actually Says

    According to the statistical scope in the original material, the report concludes that within the target sample and target spending range, about 89% of open-source model-related funds flow to Chinese models.

    This includes at least three constraints.

    1. It is a share within a market segment

    It reflects a defined spending set, not all enterprise technology budgets.

    For example, an enterprise may purchase several things at the same time:

    • general model services;
    • private deployment capabilities;
    • database and storage resources;
    • compute clusters;
    • security audit systems;
    • business software and development tools.

    Even if a high proportion of that enterprise’s open-source model spending flows to Chinese models, it does not follow that all of its AI budget flows in the same direction.

    2. It measures money, not necessarily usage

    “Used more” is not equivalent to “spends more.”

    One model may have many users but a very low price per call. Another model may have fewer users but handle more complex tasks, long-context tasks, or high-compute tasks, generating higher revenue in the end.

    A simple representation is:

    Total spending = number of calls × cost per call

    If resource consumption across different tasks is considered, this can also be written as:

    Total spending = number of users × calls per user × resource cost per call

    User scale, call volume, and actual dollar amounts must therefore be measured separately. “Used by more people” cannot be directly equated with “generates more revenue,” and “more popular among developers” cannot be directly equated with “captures most enterprise budget.”

    3. Spending scope may cover only observable parts

    When enterprises use open-source models, not all spending necessarily appears as model service fees. Some costs may be recorded under infrastructure, container platforms, compute resources, storage services, or internal R&D budgets.

    This means that if a report only counts directly identifiable model calls and hosting spending, it may reflect the “observable model service market” rather than the full economic cost of enterprise open-source model adoption.

    3. How the Three Vendors’ Shares Are Derived

    The original material provides three specific shares:

    • the first model accounts for 61%;
    • the second model accounts for 15%;
    • the third model accounts for 11%.

    Adding the three together:

    61% + 15% + 11% = 87%

    If the report also states that “Chinese models as a whole account for 89%,” the remaining part can be calculated as:

    89% - 87% = 2%

    On the surface, the three models already cover the vast majority of spending related to Chinese models.

    However, “87%” cannot simply be described as “87% of all U.S. enterprise AI budgets,” because its denominator is still the relevant spending within the target market segment.

    A more precise statement is:

    In the open-source model spending sample defined by the report, the three Chinese models together account for about 87%, close to the overall 89% result for Chinese models.

    The word “about” is important here. The reported figures of 61%, 15%, and 11% may have been rounded. The actual values could be 60.6%, 14.7%, and 10.8%, and there may also be differences in sample selection, data cleaning, or time windows.

    The 2 percentage points therefore do not necessarily represent a real and stable market category. They may simply be the result of several factors:

    1. individual percentages have been rounded;
    2. the statistical scope of the three models and that of Chinese models overall are not exactly the same;
    3. data collection times differ;
    4. different models are classified differently;
    5. the report uses different weighting methods;
    6. some spending in the sample cannot be clearly classified.

    4. Why More Users Can Sometimes Mean Less Spending

    The open-source model market differs from the traditional software market. Traditional software is often charged by account, seat, or fixed subscription, while model service costs are often related to call volume, input and output length, and computational complexity.

    When a model is adopted by many developers and enterprises, three phenomena may occur.

    1. Price per call declines

    As model inference efficiency improves, hardware utilization increases, and service scale expands, the cost per call may fall. Even if call volume grows rapidly, total spending may not grow proportionally.

    For example:

    • user numbers grow;
    • call volume increases;
    • but the price per call drops to one quarter of the earlier level.

    The final total spending may be only three quarters of the original.

    2. Many users are in low-frequency usage states

    Developers may only run tests, validate prototypes, or integrate a small number of features. These users may account for a large share of “active users” but generate limited actual revenue.

    By contrast, a small number of enterprise customers may run large numbers of automated tasks every day, continuously calling high-complexity capabilities, and contribute higher revenue.

    Market analysis therefore cannot look only at user numbers. It should also observe:

    • number of active users;
    • average daily calls;
    • average input and output length;
    • share of high-complexity tasks;
    • retention rate of enterprise customers;
    • average revenue per user;
    • share of production-environment calls.

    3. Open-source deployment shifts costs to infrastructure

    If an enterprise deploys a model in its own compute environment, the model vendor may not directly receive all inference fees. The enterprise may purchase model adaptation, technical support, hosting capabilities, or related platform services, while compute resource costs are counted under internal infrastructure budgets.

    As a result, model usage may be high, but model revenue observable by the external market may not grow at the same rate.

    This is why “model influence” and “model commercial revenue” require two different sets of metrics.

    5. Four Categories of Scope to Check When Reading Industry Reports

    1. Check the time window

    Does the report cover one month, one quarter, or cumulative data over the past year? Short-term data may be affected by product launches, concentrated enterprise procurement, or test projects going live.

    If a model happens to receive a large number of trials during the statistical period, the result may be temporarily high. If some enterprises have not completed procurement settlement, the data may be temporarily low.

    2. Check the sample scope

    A report may cover only specific regions, customer types, or service channels.

    For example, consumer-grade usage by individual developers differs greatly from production deployment by large enterprises in payment model and call scale. Covering only one type of sample cannot directly represent the entire market.

    3. Check the weighting method

    Shares are usually calculated in two main ways:

    Weighted by customer count:

    customers of a model ÷ total customers

    Weighted by spending amount:

    spending generated by a model ÷ total model spending

    Other weightings may also be used, such as call volume, compute resource consumption, or enterprise size.

    Different weighting methods can produce completely different results. A model may have a high share of customers but an average share of dollar amounts; another may have few customers but a high dollar share because of large enterprise projects.

    4. Check classification rules

    What standard is used to classify a “model from a certain country”?

    Classification may be based on the R&D entity, model ownership, deployment location, service provider, or data collection channel. If the report does not clearly define its rules, readers cannot judge whether the results are comparable.

    This is especially important when models are retrained, redeployed, or provided through third-party platforms, because the model source and the service source may not be the same.

    6. How to Avoid Exaggerating the Conclusion

    When facing a conclusion such as “nearly 90% of funds flow to a certain type of model,” a three-step verification method can be used.

    First, restore the denominator. Confirm whether it is total AI budget, open-source model budget, model call fees, or transaction volume within a certain service channel.

    Second, confirm the metric. Determine whether the number describes dollar amounts, customer counts, call volume, compute volume, or market exposure.

    Third, evaluate the boundaries of extrapolation. Even if the data holds within the original sample, it cannot automatically be extended to the global market, all enterprises, or future trends.

    A more appropriate statement is usually:

    Within a specific statistical period, specific sample, and specific open-source model spending scope, Chinese models received a high share of funds, with three models together accounting for the main portion. This result cannot be directly equated with the structure of overall U.S. enterprise AI budgets, nor can it alone prove that the long-term competitive landscape has been determined.

    7. Enterprise Procurement Decisions Cannot Rely Only on Market Share

    For enterprises, market share is only a reference indicator and should not be the sole basis for procurement. Actual model selection also requires evaluating:

    • whether model output quality meets business goals;
    • whether private or isolated deployment is supported;
    • whether data can be effectively protected;
    • service stability and failure recovery capability;
    • ease of integration with existing systems;
    • whether inference costs are predictable;
    • whether monitoring, auditing, and permission controls are complete;
    • whether the model license and commercial use boundaries are clear;
    • whether the model can be continuously maintained and upgraded.

    In engineering practice, enterprises should establish a repeatable evaluation process, including a unified test set, fixed task scenarios, cost estimation, latency monitoring, and security review.

    A composite evaluation can be abstracted as:

    Composite score = quality weight × task performance + cost weight × cost per unit of output + stability weight × service availability + compliance weight × risk control capability

    Specific weights should be adjusted according to the business type. A system for internal knowledge Q&A may place more emphasis on accuracy and data isolation; a real-time interactive system may place more emphasis on latency and stability; a batch-processing system may focus more on cost per task and throughput.

    8. Conclusion: The Real Issue Is Statistical Method

    If the data source is reliable, the phenomenon that “most U.S. enterprise spending on open-source models flows to three Chinese companies” is indeed worth attention. But its meaning should be built on a correct statistical interpretation.

    The most important conclusion is not that “89% equals 89% of U.S. enterprise AI budgets,” but rather:

    1. 89% refers to a specific open-source model spending scope;
    2. the three models’ shares of 61%, 15%, and 11% add up to about 87%;
    3. the remaining 2 percentage points may be affected by rounding and differences in statistical scope;
    4. user scale, call volume, and revenue amount cannot substitute for one another;
    5. a single report cannot represent the entire market, let alone directly predict the future landscape;
    6. enterprise decisions should combine quality, cost, security, stability, and compliance.

    Only when the denominator, metrics, time window, and sample boundaries are placed back into the analytical framework do the numbers become truly meaningful. For developers, procurement staff, and technical managers, the most valuable capability is not remembering a particular percentage, but judging what that percentage actually measures, what it omits, and whether it is sufficient to support a larger conclusion.