The Statistical Logic Behind U.S. Enterprise Spending on Open Models

    Claims that “most U.S. enterprise spending on open models flows to Chinese models” can be misleading without clear definitions of the sample, spending category, denominator, and payment path. This article explains how to interpret the reported 89% figure, distinguish adoption from usage and spending, avoid combining incompatible market shares, and evaluate models based on total cost per successful business task rather than headline market share alone.

    1. Define the Denominator Before Discussing Where the Money Goes

    The statement that “most of the money U.S. companies spend on open models flows to Chinese models” is easy to misinterpret. It may involve several different dimensions: who is purchasing, where the model originated, who receives the payment, and what the statistics actually cover. If any one of these conditions is omitted, the conclusion may change substantially.

    Companies generally do not pay for the “open” attribute itself. They pay for model execution and related services, such as hosted API access, inference compute, dedicated deployment instances, operations, and technical support. Different purchasing methods produce different recipients of revenue, and a single report may not fully cover all of them.

    The source material provides one important figure: under the relevant statistical definition, 89% of spending goes to Chinese models. However, it does not provide the complete report, sample description, or statistical period. This percentage should therefore be treated as a statistical claim that requires verification, not immediately elevated into a statement about the entire market.

    In particular, it should not be interpreted as meaning that “89% of U.S. enterprise AI budgets are spent on Chinese models.” Enterprise AI budgets may also include compute, data governance, application development, security and compliance, software procurement, and personnel. Even if open-model-related purchases are highly concentrated in a particular sample, they may represent only a small portion of total AI spending.

    A share describes how a defined set is allocated internally. It does not describe how large that set is relative to the entire market.

    2. Expand the Budget Hierarchy to Understand the 89%

    A simplified model can be built with three variables:

    • T: total AI spending by the sampled companies;
    • O: spending categorized by the report as open-model-related purchases;
    • C: spending within O attributed to Chinese models.

    If the reported 89% represents C ÷ O, then the relevant relationship is:

    C ÷ T = (O ÷ T) × (C ÷ O)

    To determine the share of total enterprise AI budgets associated with Chinese models, we must at least know O ÷ T. The 89% figure alone is not enough.

    For example, purely for illustration, assume that open-model-related purchases account for 2% of total AI spending in a sample, and that 89% of those purchases are attributed to Chinese models:

    2% × 89% = 1.78%

    The 2% figure is not an actual survey result. It is only an assumed value used to demonstrate the denominator relationship. If O ÷ T changes, the resulting share changes as well.

    Without knowing the proportion of total AI budgets represented by open-model-related purchases, it is impossible to state precisely how many times an interpretation has been exaggerated. The rigorous conclusion is that applying a share from a narrow purchasing category to the entire AI budget expands the scope substantially; the extent of that expansion depends on the actual size of the category.

    3. Why More Users Can Still Mean Lower Spending

    Adoption, usage, and spending are three different metrics.

    Adoption measures how many companies or teams use a particular type of model. Usage may be measured by request volume, tokens processed, runtime, or completed tasks. Spending measures the amount companies actually pay. These metrics may move in the same direction, but they can also diverge.

    1. Lower Unit Costs Can Offset Growth in Usage

    Under comparable conditions, spending can be approximated as:

    Spending = Usage volume × Average unit cost

    Assume usage grows by 50% while the average unit cost falls by 40%:

    1.5 × 0.6 = 0.9

    Usage has increased, but spending has fallen by 10%. This is a hypothetical calculation that illustrates the mechanism, not an observed market result.

    In real operations, lower unit costs may result from better inference efficiency, request batching, cache reuse, workload routing, or shifting lower-complexity requests to smaller models.

    2. New Adopters May Be Lightweight Users

    A company beginning to use a model does not necessarily mean that the model has been deployed at scale. A team may be running a proof of concept, processing only a small number of internal tasks, or using the model as a backup route for an existing solution.

    If new adopters generally consume fewer resources while existing high-volume users optimize their costs, the market can show more adopting companies but lower total purchasing volume.

    3. Deployment Changes Can Alter How Spending Is Recorded

    The same model capability can be purchased through a hosted service or deployed by the enterprise itself.

    With a hosted API, costs are often recorded as model-service spending. After switching to self-hosting, related costs may be recorded under compute, equipment depreciation, electricity, operations, or research and development.

    The model may still be running, and business usage may continue, while the observable amount of “model-service spending” declines. A reduction in model-service purchases therefore cannot automatically be interpreted as a decline in actual usage.

    4. Why Three Market Shares Cannot Simply Be Combined

    The source material also lists shares for three anonymous model providers: 61%, 15%, and 11%. Arithmetically:

    61% + 15% + 11% = 87%

    Compared with 89%, this appears to leave a two-percentage-point difference. But this calculation alone does not provide a statistical interpretation.

    Several percentages can be added or subtracted directly only when all of the following conditions are met:

    1. They use the same statistical period;
    2. They come from the same sample;
    3. They use the same definition of spending;
    4. They have the same denominator;
    5. Their categories are mutually exclusive and do not involve double counting;
    6. They use consistent rules for geographic attribution.

    If 61% represents a share within a narrow category while 89% represents a share within a broader category, the two figures cannot be directly compared. Likewise, percentages based on spending cannot be combined with percentages based on usage volume.

    Even when the methodology is consistent, rounding can affect the total. The two-percentage-point difference therefore cannot automatically be interpreted as the market share of “other Chinese models.”

    The statement that “one provider accounts for more than half” also requires a qualification. Only after the denominator of the 61% figure is confirmed can we say what category it represents more than half of. A majority share within a narrow sample does not equal a majority share of the entire enterprise market.

    5. Model Origin Is Not the Same as the Final Recipient of Revenue

    When discussing where the money goes, at least three layers should be distinguished:

    • Model origin: which organization developed and released the base model;
    • Transaction counterparty: which organization receives the enterprise invoice or payment;
    • Revenue distribution: how the money is ultimately divided among service providers.

    These three parties or concepts may overlap, but they do not have to.

    For example, a company may access an open-weight model through a hosting provider. The company pays the hosting provider, and that bill may include compute, networking, storage, scheduling, and operations. Whether the model developer receives revenue directly, and how much, depends on the specific commercial arrangement. It cannot be inferred from the model’s origin alone.

    Therefore, “spending associated with Chinese models represents a high share” and “Chinese model developers receive the same share of revenue” are not equivalent conclusions.

    It is also important to note that “open source” is sometimes used broadly in industry discussions to refer to open-weight models. Open weights, open training code, open training data, and permission for different types of commercial use are not automatically equivalent. For enterprise procurement, the relevant license and usage restrictions should be reviewed rather than relying only on a category label.

    6. Why Narrow Spending Categories May Concentrate Around a Few Models

    Once the statistical definition is confirmed, a high concentration level can still be analytically meaningful. However, it should not be explained simply as proof of absolute technical superiority or the lowest price.

    Enterprise model selection typically reflects several factors.

    1. Task Fit Determines Effective Output

    Companies do not need to use the most capable and expensive model for every task. For structured extraction, document classification, information organization, and other rule-defined assistance, cost and stability after meeting the required quality threshold may matter more.

    A model that fits high-frequency business tasks may therefore obtain a large share of calls and purchases. This mechanism explains how concentration can emerge, but it does not replace causal evidence for a specific sample.

    2. Integration Costs Influence Continued Selection

    After a model is integrated into a business process, teams often build prompt templates, evaluation sets, exception handling, and output-validation mechanisms around it. Switching models requires renewed testing and may introduce changes in formatting, quality fluctuations, and compatibility issues.

    These migration costs create purchasing persistence. If a model performs well during early validation, it may later be assigned more adjacent tasks.

    3. Data Boundaries and Deployment Constraints Narrow the Candidate Pool

    Some businesses need to control where data is processed, who can access it, and how logs are retained. The ability to deploy within existing infrastructure, support auditing, and meet licensing requirements can all influence the final choice.

    However, “can be self-hosted” does not mean “costs less to deploy,” nor does it automatically mean “compliant.” Enterprises still need to handle capacity planning, patch management, access isolation, disaster recovery, and security reviews.

    7. Evaluate Models by Cost per Successful Task

    For developers and procurement teams, market share can be a useful signal, but it should not be the final basis for model selection. A more practical metric is cost per successful business task:

    Cost per successful task = Total investment ÷ Number of tasks that meet business quality requirements

    Total investment should include more than API or compute charges. It may also include:

    • Integration, adaptation, and ongoing maintenance;
    • Retry costs caused by incorrect outputs;
    • Human review and correction;
    • Business losses caused by service failures;
    • Security governance and compliance reviews;
    • Future migration and replacement costs.

    A model with a lower per-call price may not have a lower actual cost if it requires more retries and manual correction. Conversely, a more expensive model may provide a lower cost per successful task if it substantially improves completion rates.

    Evaluation should use anonymized test sets derived from real business workflows. Results should be recorded by task type, including quality, response time, failure rate, and cost. Difficult cases and abnormal inputs should be retained to avoid deciding whether a model is suitable for production based only on a small number of demonstration examples.

    8. A Checklist for Reviewing Reports of This Kind

    Before citing a claim that “a certain class of models accounts for most enterprise spending,” verify the following:

    First, who does the sample represent? Is it all enterprises, customers of a particular procurement channel, or a small group of surveyed organizations? A convenience sample does not automatically represent the overall market.

    Second, what does the measurement cover? Does it include self-hosting, compute procurement, and indirect purchases? Excluded categories may materially change the overall conclusion.

    Third, what does the spending figure mean? Is it actual payment, contract value, estimated consumption, or net spending after discounts? These measures should not be mixed.

    Fourth, what is the denominator? Is the percentage based on total AI budgets, model-service budgets, or spending on hosted open-model calls?

    Fifth, how is geographic attribution determined? Is it based on the model developer, service provider, payment recipient, or another rule? Different attribution methods can produce different conclusions.

    Sixth, is there a comparable time series? If the sample composition changes, a shift in market share may result from a change in the sample rather than a change in market behavior.

    When this information is incomplete, the safest approach is to retain the relevant qualifications. Clearly distinguish which figures come from the source material, which explanations are possible mechanisms, and which conclusions cannot yet be verified.

    Conclusion: Market Share Is a Starting Point, Not the Final Answer

    Whether U.S. enterprise purchases related to open models are concentrated in Chinese models is first a data question that requires clear definitions of the sample and measurement framework. Only then does it become a question of technical competition and business models.

    More adopters do not necessarily mean higher purchasing volume. A high share within a specific purchasing category does not equal a high share of total AI budgets. Spending attributed by model origin does not equal the revenue actually received by the model developer.

    For technical teams, the more important question is whether a model can complete business tasks reliably and sustainably under existing data, security, and deployment constraints. Clarifying the statistical methodology, then evaluating quality, cost, and risk within the same framework, is the best way to extract dependable decision-making insight from noisy market figures.