How to Read Enterprise AI Spending Reports: Denominators, Definitions, and Statistical Traps

    A widely shared survey suggests that 89% of US enterprise spending on open-source models goes to Chinese-developed models—a figure that is both informative and frequently misread. This article uses that dataset to demonstrate a repeatable method for interpreting market-share reports: verifying the publisher and sample, identifying the true denominator, checking internal consistency, and separating usage share from spending share. Developers and technology decision-makers will learn how to evaluate similar industry data and avoid conclusions that are off by orders of magnitude.

    1. A Number That Sparked Debate

    A recent survey of AI leads at US enterprises has been circulating in Chinese tech circles. Its most eye-catching finding: roughly 89% of US enterprise spending on open-source large models went to models developed by Chinese teams. Media coverage later added the individual shares of the three leading models, further fueling discussion that “Chinese models have taken over the open-source market.”

    The numbers themselves are informative, but much of the commentary misreads them. As engineers and technology decision-makers, we face industry reports and market-share figures every day. Knowing how to read the definitions and logic behind the numbers matters more than memorizing the numbers themselves. Using this dataset as an example, this article walks through a complete “reading method.”

    2. Step One: Understand Where the Report Comes From

    Before citing any data, confirm three things: who published it, what the sample is, and what the definitions are.

    For this report:

    • Publisher background: It comes from a US venture capital firm with decades of history that has long tracked investment targets in enterprise software and AI. The survey is conducted from an investment perspective, not as pure academic research.
    • Sample scope: The survey covers nearly 400 AI business leaders at US enterprises, who collectively oversee more than $4 billion in AI spending. Note that the sample is “US enterprises,” not the global market—the conclusions cannot be extrapolated to other regions.
    • Data nature: This is questionnaire data, not transaction data. Survey data reflects respondents’ subjective reporting and estimates, and inherently carries sampling and reporting bias.

    Only after confirming these three points does it make sense to look at the numbers themselves.

    3. Step Two: What Is the Denominator Behind 89%?

    This is the most commonly misread part of the dataset. The original statement is: of US enterprise spending on open-source models, 89% flows to Chinese models.

    The key phrase is “open-source model spending.” It is not “total US enterprise AI spending,” and certainly not “US enterprise IT budgets.”

    How big is the gap? The same report provides the overall picture of enterprise AI spending: US enterprise AI spending in 2026 is projected to reach the hundreds-of-billions scale, with the application layer accounting for the bulk—coding tools and enterprise conversational applications each at tens of billions. Open-source model–related spending is only a small slice of that total.

    An even more telling comparison: the report shows that open-source models’ token usage share is rising (from about 11% to 19%), while their spending share is actually falling (from about 6% to 4%). The reason is straightforward: average token prices for open-source models have dropped sharply, and volume growth has not offset the decline in unit price.

    A more accurate reading, therefore, is:

    More and more people are using open-source models, but less money is being spent on them; and within this shrinking pool of spending, nearly 90% goes to Chinese models.

    If 89% is interpreted as “89% of US enterprise AI budgets,” the error is close to two orders of magnitude. This is not because Chinese models are weak—it is determined by the open-source business model itself: the value of open source shows up more in adoption and ecosystem influence than in direct model-call revenue.

    4. Step Three: How to Verify Share Concentration

    The report also gives the shares of three Chinese models within open-source spending: one leading model at 61%, and the other two at 15% and 11% respectively.

    A simple addition check: 61 + 15 + 11 = 87, about 2 percentage points short of the 89% total.

    There are several possible explanations for the 2-point gap:

    1. Rounding error: the three shares are each rounded independently, and the sum deviates from the total—a normal phenomenon in survey data;
    2. Definition differences: the three shares may be based on different subsamples or different questions, and are not strictly a split of the same denominator;
    3. Long-tail distribution: the remainder may flow to other open-source models or uncategorized options.

    The rigorous approach: do not treat these 2 points as “the share of other models” to draw conclusions; use them only to check whether the data definitions are self-consistent. 87 versus 89 suggests the three shares most likely come from the same question, and internal consistency is acceptable.

    Further inference: a single leading model accounts for 61% of all open-source spending, meaning that in the specific segment of “US enterprises paying for open-source models,” more than half the money goes to one vendor—extremely high concentration. But note again that this is concentration at the level of “willingness to pay,” not “deployment volume” or “inference call volume.” In enterprise settings, the two often diverge (many enterprises self-deploy open weights and do not go through vendors’ paid channels).

    5. Step Four: Recognize the Divergence Between Spending and Usage

    The most valuable insight in this dataset is what it reveals about the structural characteristics of the open-source model market:

    Dimension Change Implication
    Token usage share Rising (11% → 19%) Open-source models are being called in practice more and more
    Average token price Sharp decline Intense competition; open-source models push prices down
    Spending share Falling (6% → 4%) Price cuts outpace usage growth
    Spending concentration Very high (top vendor at 61%) The paid ecosystem is converging on leaders

    The implication for enterprise technology selection: if you call open-source model APIs on a pay-per-use basis, the market choices are highly concentrated; if you self-host inference services with open weights, the choice set is much broader. The two paths differ completely in cost structure, operational complexity, and compliance risk, and should be evaluated separately in architecture decisions rather than being swayed by a single number like “89%.”

    Consider how enterprises typically consume open-source models: they can access them through cloud model services billed by call volume, or self-deploy model weights on container platforms or GPU compute. The former reduces operational overhead; the latter offers greater control. The choice should weigh data sensitivity and the team’s engineering capability.

    6. A Reusable Data-Reading Checklist

    Abstracting the process above into a general method, any market-share data can be run through these five questions:

    1. Who published it? The publisher’s position and motives shape the lens.
    2. Who is the sample? Conclusions hold only for the sample population—beware of geographic and industry extrapolation.
    3. What is the denominator? The denominator of 89% is “open-source spending,” not “total AI spending”—a difference of two orders of magnitude.
    4. Are the definitions self-consistent? When components do not sum to the total, find out why before drawing conclusions.
    5. Do usage and money diverge? In scenarios where usage grows and spending falls, “money share” is not the same as “momentum share.”

    7. Conclusion

    “89% of US enterprise open-source model spending flows to Chinese models” is a real and meaningful market signal. It shows that Chinese open-source models have come to dominate in international enterprise paid scenarios. But the correct magnitude of this number is: 89% of a slice that accounts for about 4% of total enterprise AI spending—far from “90% of enterprise AI budgets.”

    For developers and decision-makers, mastering the method of reading numbers is more valuable than spreading them: confirm the source, pin down the denominator, verify the definitions, and distinguish usage from spending. With this method, you will consistently see one layer deeper than most when navigating the flood of industry reports.