Demystifying Enterprise Open-Source Model Expenditure: Denominators, Cost Misinterpretations, and TCO

    This article analyzes common misinterpretations in enterprise open-source AI model expenditure statistics, such as the claim that 89% of enterprise AI budgets flow to specific model vendors. It clarifies the importance of statistical denominators, subset normalization, and absolute amounts versus structural ratios. Furthermore, it outlines the true Total Cost of Ownership (TCO) for open-source models and provides a reusable enterprise cost accounting framework for accurate budget forecasting and procurement decisions.

    A common statistical conclusion often gets summarized in circulation as “More than half of the money enterprises spend buying open-source models flows to a few Chinese models.” While this conclusion might hold true, simplifying it to “89% of US enterprise AI budgets goes to Chinese models” changes the meaning by orders of magnitude. The issue isn’t whether the numbers look good, but whether the statistical denominators are fully preserved. Technical managers, procurement officers, and architects must answer three questions when reading such reports: What is the denominator? Within which subset is the share normalized? Are absolute amounts being mixed with structural ratios? This article does not evaluate specific vendors; it discusses how to correctly interpret such cost and share data and provides a reusable enterprise accounting method.

    1. Define the Denominator: What Does 89% Actually Mean?

    The most common misinterpretation of percentages occurs when the denominator is quietly replaced. A conclusion that “89% flows to Chinese models” could imply several statistical meanings:

    1. The denominator is the enterprise’s total AI budget, including compute, data, annotation, integration, O&M, and model invocations.
    2. The denominator is only the expenditure on model invocations or hosting services.
    3. The denominator is the newly purchased amount, not the total existing expenditure.
    4. The denominator is the expenses saved after migration, not the enterprise’s original total budget.
    5. The denominator is the reduced portion of the budget within a specific time window, not the total budget.

    The correct reading emphasized in the original material is: As more people use the models, the money spent decreases, and nearly 90% of that decreased amount went to Chinese models. The core of this statement is the subset of “decreased money.” Reading it as “89% of the total enterprise AI budget” incorrectly elevates a subset ratio to a global ratio.

    Let’s illustrate the magnitude of difference with an example. Suppose an enterprise’s annual total AI budget is 1,000 units, of which 50 units are directly spent on model invocations and hosting. Due to unit price drops and invocation structure optimization, this expenditure decreased by 10 units compared to the previous year. Of those 10 reduced units, 89% (8.9 units) flowed to Chinese models. Here, 8.9 units accounts for only 0.89% of the enterprise’s total AI budget—a difference of about two orders of magnitude from 89%. Even if we change the denominator to the 50 units of model-related expenditure, 8.9/50 is only 17.8%, still not 89%.

    Therefore, when seeing “89%,” you must at least clarify four fields: the numerator, the denominator, the time window, and whether it has been normalized. Missing any of these makes the data unsuitable for direct use in budget forecasting or procurement decisions.

    2. Share Summation: The Meaning of 61+15+11=87 and 89-87=2

    The original material presents a formula: Out of a total of 89% flowing to Chinese models, three vendors hold shares of 61%, 15%, and 11%. Together, they sum to 87, and 89 minus 87 equals 2. While this math seems simple, you must first confirm if the denominators are consistent.

    If 61%, 15%, and 11% are all “percentage points of the total sample,” then the three combined hold 87 percentage points, and the remaining 2 percentage points can be understood as the share captured by other similar models. In this case, 89% is the total caliber, the three combined equal 87%, and the remaining 2% belongs to other models in the same category. To compare the relative structure of the three within the “Chinese models” subset, normalization is required:

    • Leading Model Family A: 61 / 89 ≈ 68.54%
    • Leading Model Family B: 15 / 89 ≈ 16.85%
    • Leading Model Family C: 11 / 89 ≈ 12.36%
    • Other similar models: 2 / 89 ≈ 2.25%

    After normalization, the total is roughly 100%. You can then say, “In the expenditure flowing to Chinese models, A accounts for about 68.5%,” but you cannot reverse it to say, “A accounts for 68.5% of the enterprise’s total AI budget.” If 61%, 15%, and 11% are already shares within the “Chinese model subset,” then the three combined equal 87%, and the remaining 13% are other similar models. In this case, you should not directly subtract from 89. In other words, 61+15+11=87 and 89-87=2 only hold true if the denominators are consistent and the shares represent percentage points of the total sample.

    This is the most common error in share analysis: adding percentages with different denominators, or treating a ratio within a subset as a global ratio. Percentage change and percentage point change are also not the same concept. A share increasing from 15% to 16% can be described as an increase of 1 percentage point, or a relative increase of about 6.7%, but the two cannot be mixed.

    Below is a sample code for caliber verification, used to check share summation, residuals, and subset normalization:

    1. total_share_pct = 89
    2. parts_pct = {
    3. "Leading Model Family A": 61,
    4. "Leading Model Family B": 15,
    5. "Leading Model Family C": 11
    6. }
    7. sum_parts = sum(parts_pct.values())
    8. residual = total_share_pct - sum_parts
    9. normalized = {}
    10. for name, value in parts_pct.items():
    11. normalized[name] = value / total_share_pct * 100
    12. normalized["Other similar models"] = residual / total_share_pct * 100
    13. # Output:
    14. # sum_parts = 87
    15. # residual = 2
    16. # normalized = {
    17. # "Leading Model Family A": 68.54,
    18. # "Leading Model Family B": 16.85,
    19. # "Leading Model Family C": 12.36,
    20. # "Other similar models": 2.25
    21. # }

    The value of this code lies not in the calculation itself, but in forcing the denominator to be labeled. Any share report should output the “original share,” “denominator definition,” “normalized result,” and “residual explanation” simultaneously. Otherwise, readers can easily misinterpret subset ratios as total budget ratios.

    3. Cost Reduction vs. Usage Growth

    “More people using it, but less money being spent” sounds counterintuitive, but it is not uncommon in software infrastructure. Model-related expenditure can be roughly broken down as:

    Total Cost = Invocation Volume × Unit Invocation Cost + Fixed Cost + Integration & O&M Cost + Compliance & Security Cost.

    When invocation volume increases, total expenditure can still decrease, provided that the unit invocation cost drops faster. Common reasons include inference optimization, cache hits, batch processing, small model routing, model distillation, context compression, maturity of hosting services, and price drops driven by market competition. Enterprises may also use multi-model routing, delegating simple tasks to low-cost models and complex tasks to high-capability models, thereby lowering the average unit price.

    However, note that cost reduction and share concentration are metrics of two different dimensions. Cost reduction describes changes in absolute amounts or unit costs, while share concentration describes expenditure structure. A model having a high share does not mean the enterprise’s total budget is high. Similarly, a certain model class having a high share in “savings” does not mean it holds a high share in existing expenditure. If a report only tracks “where the saved fees from migration went,” it cannot directly deduce “where the total enterprise AI budget went.” Technical managers should simultaneously review absolute amounts, growth rates, base numbers, invocation volumes, and unit costs to avoid focusing solely on a single percentage.

    4. TCO Boundaries of Open-Source Models

    “Open-source models” easily create the illusion of “zero cost.” Open weights do not mean zero cost of enterprise usage. Actual enterprise expenditures typically include:

    • Model invocation or hosting service fees;
    • Inference compute resources, including container platforms, elastic scaling, load balancing, and monitoring/alerting;
    • Fine-tuning, training, and evaluation costs;
    • Data collection, cleaning, annotation, and governance costs;
    • Application integration, gateways, routing, caching, and logging service costs;
    • Security compliance, permissions, auditing, and content governance costs;
    • Engineering costs from migration, rollbacks, and multi-person collaboration.

    If the statistical caliber only covers model invocation fees, it misses the costs of self-built inference and O&M. If it only covers cloud resources, it is affected by deployment architecture, resource utilization, and billing methods. Therefore, the boundary of “money spent buying open-source models” must be clarified: is it licensing fees, service fees, compute fees, or the Total Cost of Ownership (TCO) including data and integration? Shares under different boundaries cannot be directly compared.

    5. A Reusable Enterprise Cost Accounting Framework

    It is recommended that enterprises establish a lightweight but auditable model cost ledger, encompassing at least the following steps:

    1. Define objectives: Is it to evaluate the total budget, optimize unit costs, or compare migration savings?
    2. Determine boundaries: Clarify whether it includes compute, data, integration, O&M, and compliance.
    3. Standardize denominators: Create separate tables for total AI budget, model-related expenditure, newly purchased amounts, and savings amounts.
    4. Collect data: Record invocation volume, unit cost, model family, business scenario, and time window.
    5. Normalize: Label subset shares separately and do not mix them with global shares in calculations.
    6. Conduct sensitivity analysis: Measure changes under varying invocation growth, unit price drops, and migration ratios.
    7. Continuously monitor: Use logging services, monitoring/alerting, and cost reports to observe deviations.

    Reports should retain “original percentages,” “percentage points,” “absolute amounts,” and “denominator explanations.” For instance, a model family might account for 61% of “savings” but only 10% of “total model expenditure.” Both are true, but they answer different questions.

    6. Practical Recommendations

    For technical leaders, the most important thing is not remembering a specific percentage, but building an awareness of statistical calibers. When seeing “89% flows to Chinese models,” first ask: Is this the total budget, model expenditure, new procurement, or savings? When the three vendors’ shares total 87%, is the remaining 2 percentage points other similar models or statistical error? Is the 61% within the total sample or within the 89% subset? Only when these questions are clearly answered can data feed into budget, architecture, and procurement decisions.

    On the engineering side, unit costs can be reduced through multi-model routing, caching, batch processing, small model routing, and observability building. For example, on Baidu AI Cloud or any general cloud platform, you can use container platforms to deploy inference services, leverage object storage to manage models and data, and use logging services and monitoring tools to observe invocation volume, latency, and cost anomalies. These are general cloud product categories and do not constitute assumptions about specific vendor capabilities. Enterprises should also bind model expenditure to business metrics, avoiding chasing low unit prices at the expense of quality, stability, and compliance costs.

    7. Conclusion

    “89% flows to Chinese models” is a statistical conclusion that requires a complete caliber to hold true. It more likely describes the expenditure structure within a specific subset, rather than 89% of the total enterprise AI budget. The formula 61+15+11=87 and 89-87=2 only makes sense when the denominators are consistent; if the shares are inherently subset ratios, they cannot be directly subtracted from 89. Enterprises should simultaneously focus on absolute amounts, unit costs, invocation volumes, migration savings, and TCO, replacing single-point percentages with unified denominators, normalized labeling, and sensitivity analysis. This prevents misreading “nearly 90% of the reduced money” as “90% of the total budget,” and ensures open-source model selection truly serves cost, quality, and sustainable operations.