Why U.S. Enterprise Spending on Open-Source Models Is Concentrated in Chinese Models
This article explains why reports of high U.S. enterprise spending on open-source models from China must be interpreted carefully. It distinguishes model usage volume, model service spending, and total enterprise AI budgets; examines why high usage can coincide with lower average costs; clarifies what the reported 89% share actually represents; and explains why enterprises often concentrate on a few models. It also outlines a practical evaluation framework covering quality, cost, deployment flexibility, routing strategy, and long-term total cost of ownership.
1. Usage Volume and Spending Volume Are Not the Same
When a report says that one category of models accounts for a large share, the easiest misreading is to treat three different metrics as if they were the same thing:
- Call volume or request volume
- Model service spending
- Total enterprise AI budget
These three metrics do not move in a simple, proportional relationship.
At minimum, enterprises generate three different kinds of data when they use models:
Call volume or request volume
This reflects how often a model is invoked. It shows usage frequency, not necessarily cost.Model service spending
This reflects how much enterprises actually pay for inference, API calls, or hosted model services. It is closer to the revenue generated by model usage.Total AI budget
This includes a much wider set of costs, such as compute procurement, data processing, personnel costs, software subscriptions, model training, application development, and security governance.
Because these categories are different, it is risky to infer one from another.
For example, one model may be used heavily for lightweight tasks such as text classification, summarization, or code completion. These tasks can generate many requests, but each request may involve relatively little input or output, so the cost per call remains low.
Another model may be used mainly for complex reasoning, long-context analysis, or multimodal processing. It may be called less frequently, but each call can cost much more.
Therefore:
- More users do not necessarily mean more spending.
- A high spending share does not mean the model consumes the enterprise’s entire AI budget.
If a report measures payments in the model service market, the more accurate statement is: within the sample of enterprises and the reporting period covered by the report, a certain category of models received a relatively high share of model invocation spending. It would be incorrect to say that these models account for all AI investment by U.S. enterprises.
2. Why High Usage May Lead to Lower Average Spending
The economics of model services often change as usage scales. Enterprises with larger usage volumes may achieve lower average costs for several reasons.
1. Unit inference costs continue to decline
Model service costs are usually related to input length, output length, computational complexity, and concurrency. For repetitive and standardized tasks, providers can reduce average per-request costs through batching, caching, routing optimization, and improved hardware utilization.
Enterprises can also reduce unnecessary computation by using prompt compression, context trimming, result caching, and task tiering.
2. Lightweight tasks generate a large share of calls
After enterprises begin deploying AI, the first workloads to spread widely are often not high-difficulty autonomous decisions, but a large number of assistive tasks, such as:
- Document summarization and classification
- Customer support Q&A
- Internal knowledge retrieval
- Content moderation and information extraction
- Code explanation and test generation
- Table field cleaning
- Ticket categorization and workflow routing
These tasks tend to have high request frequency, relatively short outputs, and stable workflows. They can significantly increase call volume without increasing total spending at the same rate.
3. Enterprises tend to choose models that are “good enough”
When purchasing or deploying models, enterprises usually do not assign every task to the most capable and most expensive model. A more common approach is to build a tiered strategy:
- Lightweight models for simple tasks
- General-purpose models for tasks requiring stronger comprehension
- High-capability models for complex reasoning or high-risk tasks
- Human review or rule-based validation for critical outputs
This architecture helps control costs while maintaining quality. It can also create a pattern where call volume is very large, but the average price per call remains relatively low.
3. How to Interpret “Nearly 90% Flowing to Chinese Models”
The core data point in the original report is that, within a specific sample and statistical scope, approximately 89% of relevant spending went to Chinese models. The key is to understand what “relevant spending” means.
It does not usually mean any of the following:
- 89% of the total AI budget of U.S. enterprises
- 89% of all model procurement spending by U.S. companies
- 89% of all cloud service or compute spending
- 89% of the global model market
A more rigorous interpretation is that the report first defines a relatively narrow statistical set. For example, it may look at spending by some U.S. enterprises on open-source models or related model services. It then compares the amount spent on different model categories within that set.
If total spending within the statistical scope is represented as T, and spending on Chinese models is represented as C, the report is essentially saying:
C ÷ T ≈ 89%
If the total AI budget of an enterprise is represented as B, and model service spending is only one part of that budget, then:
T < B
Therefore, even if C ÷ T is close to 89%, it does not follow that:
C ÷ B = 89%
The denominators are different, so the meaning of the conclusion is completely different. Many disputes do not come from arithmetic errors, but from the statistical scope being expanded during dissemination.
4. Why Three Models May Account for Most of the Sample Share
The report also provides shares for three models: 61%, 15%, and 11%. If these numbers belong to the same statistical scope, the same time range, and the same share definition, a simple addition gives:
61% + 15% + 11% = 87%
If the overall share flowing to Chinese models is 89%, the remaining portion is:
89% - 87% = 2%
This “2 percentage points” only suggests that, under this presentation, a small share may belong to other models or fall within a statistical margin. It should not be treated directly as a standalone business conclusion.
The reason is that percentages in a report may involve several complicating factors.
1. Rounding errors
The original data may have been 60.6%, 14.7%, and 10.8%. When displayed as 61%, 15%, and 11%, rounding can cause the summed total to differ from the overall proportion.
2. Different statistical dimensions
The overall proportion may be calculated by model origin, while individual model shares may be calculated by payment amount, call volume, or number of customers. If the denominators differ, subtracting one from another has no strict meaning.
3. Different sample coverage
The overall proportion may include all identified models, while the ranking of the top three models may only include models that reached a certain threshold. Models not listed separately may be grouped into an “other” category.
4. Inconsistent time windows
The model market changes quickly. A report may use an overall proportion from one period and model rankings from another period, or it may combine monthly and quarterly data. When the time ranges are not aligned, the numbers can only serve as approximate references.
For this reason, professional analysis should not focus only on a few percentages. It should also confirm:
- What the denominator is
- Which sample is being measured
- How long the time window is
- Whether amounts have been normalized
- Whether the indicators use a consistent definition
5. Why Enterprises Concentrate on a Few Models
Enterprise model procurement often shows a clear concentration trend. This does not necessarily mean other models lack technical value. Instead, it is often driven by engineering and management costs.
First, enterprises prefer models that have already been validated. Before a model enters production, it typically needs to pass tests related to accuracy, stability, latency, concurrency, data security, content compliance, and failure recovery. Once a model passes internal evaluation, subsequent applications will reuse the existing capabilities and operational knowledge.
Second, migration costs create path dependence. Once an enterprise builds prompt templates, evaluation sets, invocation gateways, logging systems, and human review workflows around a model, switching models is not simply changing a configuration item. The enterprise must also re-test performance, adjust context handling, assess risks, and retrain users.
Third, open-source models offer strong deployment flexibility. Enterprises can deploy them in private environments, dedicated resources, or controlled networks depending on data sensitivity. This helps avoid sending all data directly to an external service. This is especially important for finance, healthcare, manufacturing, government systems, and large internal enterprise environments.
Finally, model ecosystems generate scale effects. The more documentation, evaluation tools, adapter components, and development experience accumulate around a model, the lower its overall cost of use becomes. Even if another model performs better on a single metric, the total migration cost may still make enterprises keep their existing choice.
6. Enterprises Should Not Look Only at Market Share
For enterprises preparing to adopt open-source models, market share can only serve as a reference. It cannot replace technical evaluation. A more reliable approach is to build a layered and reproducible evaluation process.
1. Define the task set
Enterprises should not rely only on generic question-answering tests. They should build task sets based on real business needs, including:
- Knowledge retrieval
- Structured extraction
- Long-text processing
- Code generation
- Classification and judgment
- Multi-turn interaction
2. Evaluate quality and cost together
Model evaluation should include at least the following dimensions:
- Accuracy
- Factual consistency
- Output format stability
- Average response latency
- Concurrency capability
- Cost per request
- Peak resource consumption
- Retry costs after failures
3. Calculate total cost of ownership
Total cost should not be judged only by API pricing. It should also include:
- Deployment resources
- Network transfer
- Operations personnel
- Monitoring and alerting
- Data masking or desensitization
- Version upgrades
- Security audits
4. Design a model routing strategy
Enterprises can dynamically route requests based on task difficulty, data sensitivity, and real-time requirements. Simple tasks can first use lower-cost models, while complex tasks can be escalated to higher-capability models. Caching, batching, and result reuse can further reduce duplicate computation.
5. Establish a continuous evaluation mechanism
After models go into production, business data, user questions, and risk types will continue to change. Enterprises should continuously collect failure cases, update test sets regularly, and run regression tests when model versions change. Evaluation should not be a one-time activity limited to the procurement stage.
7. Conclusion: Check the Denominator Before the Ratio
The statement that “most U.S. enterprise spending on open-source models goes to Chinese models” should not be judged only by how striking the number sounds. The more important question is: which market segment is it describing?
A more accurate analytical path is:
- First confirm the statistical object: is it call volume, payment amount, or total budget?
- Then confirm the sample scope: is it some enterprises or the entire market?
- Then check the time window to avoid mixing data from different periods.
- Finally verify whether the percentages use the same denominator, and consider rounding errors and sample omissions.
If the report measures a specific set of model service spending, the 89% figure can show structural concentration within that set. However, it should not be expanded into a conclusion about the overall allocation of U.S. enterprise AI budgets.
For technical decision-makers, the most valuable takeaway is not simply judging which category of models “won.” It is understanding the real logic behind enterprise procurement: model quality, unit cost, deployment flexibility, ecosystem maturity, and migration cost together determine the final share. Only by combining statistical definitions with engineering practice can a set of market numbers become a reliable technical judgment.
