The $4.7 trillion filing puzzle - CharitySense Data

CharitySense Data

11 min read Original article ↗

01/What the $4

What the $4.69 trillion measures

The current sitewide revenue sum is $4.69 trillion. It selects one primary filing for each public profile and adds the revenue reported on Forms 990, 990-EZ and 990-PF. The selected filings span tax years. The sum is drawn from2,222,250 profiles containing 8,868,546historical 990-series filings, but those are coverage counts, not the number of organizations with an exact revenue figure. Form 990-N does not report an exact revenue amount, and churches generally do not have to file an annual 990-series return. We cannot turn this visible sum into a total for the entire tax-exempt sector.

For a like-for-like comparison of revenue, assets and program expenses, we narrow the view to the 485,109 current full Form 990 records, one latest filing per organization. Together they report $4.50 trillion in revenue and $10.04 trillion in year-end assets. Program-service expense fields are present on 445,199 of those records and sum to $3.29 trillion. These are filed expense categories, not a measure of outcomes or direct aid to recipients.

Full Form 990 financials, in trillions

All three bars use the same current full-990 cohort; the program-expense field is absent on some filings.

Chart source values; units and population are described in the surrounding figure.
labelvalueU.S. dollarsfilers
Reported revenue4.54,498,864,184,992485,109
Program expenses3.293,287,342,647,334445,199
Year-end assets10.0410,042,708,243,163485,103
CharitySense analysis of current full Form 990 filings, one latest return per organization, September 25, 2026.

We parsed IRS Form 990 filings. The corpus contains 1,540,498 Form 990-N filings, about 17.4% of its filing rows. An eligible small organization may use this electronic notice when gross receipts are normally $50,000 or less. The notice asks whether it meets that threshold, but not for an exact revenue figure. This is one reason profile coverage is much wider than exact-dollar coverage.

A tax form was never built for this

A hospital's payroll is part of its service, and a small nonprofit also needs people to run programs. Form 990 sorts spending into accounting categories. Those categories are useful, but they cannot tell us what changed for the people an organization serves.

We keep two questions separate: how much revenue these filers disclose, and how much comes from charitable gifts. A hospital can earn most of its revenue from patient care; another organization may rely chiefly on donations. The $4.69 trillion combines those models. The source of revenue matters before anyone treats that headline as a measure of giving.

02/average hides the typical

The average hides the typical

A trillion-dollar headline can make every nonprofit sound large. Among the 485,109 latest full Form 990 records, annual revenue at the midpoint is $361,386. The arithmetic mean is $9.27 million, about 25.7 times as much. This group excludes the 990-N postcard filers, so its median should never be described as the median of all public profiles.

A single sector-wide number describes almost none of the organizations inside it.

The largest 1% of full-990 organizations report 71.03% of this group’s revenue. The top 10% report 92.96%. At the other end, 342,197 organizations with nonnegative revenue below $1 million account for only 2.18%of the total. The 626 organizations reporting $1 billion or more account for 42.44% of nonnegative revenue. Click a bar to see every revenue band; the bands omit 3,096 filings with negative revenue.

Revenue is held at the very top

The top slices use all 485,109 full-990 records; the revenue bands separately show nonnegative amounts. Click for the split.

Chart source values; units and population are described in the surrounding figure.
labelPercent
Top 1% of orgs71.03
Top 10% of orgs92.96
Under $1M2.18
CharitySense analysis of 485,109 latest full Form 990 filings, September 25, 2026.

The $100,000-to-$1 million band contains over half of the nonnegative-revenue organizations in this full-990 sample, yet it accounts for only about 2% of the reported dollars. The $1 billion-plus band contains 0.13% of the sample and accounts for more than 42% of the dollars. Counting organizations and counting dollars answer quite different questions.

03/filing year changes

The filing year changes the denominator

The historical full-990 series contains a different set of filers in each tax year. Watch reported revenue and assets move beside the number of filing organizations. In 2021, all three lines dip together: the filing record contains 221,910 full 990 filing rows for that year, compared with 317,551for 2020 and 349,371 for 2022. We cannot call the revenue dip a sector contraction without accounting for that change in coverage.

Reported totals move with the filing cohort

Each observed full-990 tax year is indexed to 2010 = 100. These are changing cohorts, not a same-organization growth panel.

2024reported full-990 assets258 vs 100filing organizations283 vs 100data status346,477

reported full-990 assetsreported full-990 revenuefiling organizations

Chart source values; units and population are described in the surrounding figure.
yearassetsrevenuecoveragecount
2,010100100100122,011
2,011107108131159,492
2,012119118147179,674
2,013127123163198,843
2,014135131179218,603
2,015142136191233,505
2,016153148200243,835
2,017165155208254,486
2,018172163218265,949
2,019189172231283,314
2,020229191259317,551
2,021144127181221,910
2,022246215283349,371
2,023263233291358,718
2,024258223283346,477
CharitySense analysis of observed full Form 990 filing rows by tax year, 2010–2024; checked September 25, 2026.

The latest substantially populated year in this series is 2024, with346,477 full-990 filing rows. The 2025 slice has only 72,691 rows so far, so we leave it off the trend chart. A falling total in an unfinished filing cohort is a warning about the dataset’s calendar, not evidence that nonprofits suddenly became smaller.

04/What the filing record

What the filing record cannot see

The exact-dollar story starts with the form an organization files. Our corpus has 3,860,065 full Form 990 filings and2,180,247 Form 990-EZ filings. It also has1,540,498 Form 990-N notices, which confirm a small-filer threshold rather than disclose precise revenue. These counts span years, so they are filing rows, not unique organizations. A reader who divides dollars by every row would manufacture a meaningless average.

The IRS also exempts churches and certain related religious organizations from the annual Form 990-series filing requirement. Churches may still file voluntarily. A profile database built from annual filings cannot count the exempt organizations that do not file, so the religious activity outside the record is unknown.

Even the detailed form has limits. It reports spending categories and grants, but it cannot on its own establish whether a program changed lives. We examine one particularly visible part of this record in what nonprofit CEOs earn: compensation is public, yet the meaning of any one pay figure depends on the role, organization and year behind it.

05/Some dollars have

Some dollars have a named destination

A program-expense total is an accounting classification. A filed grant can say something different: who reported sending it, who was named as the recipient, the amount and where it appeared on the return. Our expanded analysis of filed money flows now includes 1,588,539 source-cited money-flow records, alongside about 456,000 organizations with named relationships. Of those records, 1,544,399 (97.2%) come from private-foundation grant lists or the domestic-grant schedule attached to Form 990. That tells us what the public return is most able to name. It does not mean grants account for 97.2% of nonprofit spending. A recorded link reflects what a filer disclosed; it is not an independent audit that the transfer happened as described.

That new linkable evidence helps us ask sharper questions about grants. It still cannot turn every line of program expense into a payment to a named beneficiary. Staff time and services often produce the program itself. For a closer look at the difference between spending inside an organization and money sent out, see where the money goes.

06/Why a raw program-efficiency

Why a raw program-efficiency percentage is unreliable

A frequently cited number is the program-efficiency ratio: the share of spending that goes to programs rather than overhead. Among 436,970 latest full-990 records with positive total expenses and a reported program-expense value, 7.89% have a program value of zero and 15.08% have a program value equal to total expenses. Together, 22.98% sit at an exact endpoint. These are values on the filings, not a verdict on how the organizations operate.

Reported program share clusters at the round extremes

The two exact endpoints account for 22.98% of current eligible full-990 records.

Chart source values; units and population are described in the surrounding figure.
labelPercentcount
Exactly 0% to program7.8934,497
All other values77.02336,576
Exactly 100% to program15.0865,897
CharitySense analysis of latest full Form 990 filings; 436,970 with positive total expenses and a program-expense value, September 25, 2026.

An endpoint can reflect a real operating model or an accounting choice. The form alone cannot separate those explanations. We also found 326 records in this eligible group where program expense is negative or exceeds total expense. Those filing-level inconsistencies are reasons to inspect the original return, not reasons to erase the filed value.

The better question is what the ratio means for a particular organization. A direct-service nonprofit and a grantmaking foundation have different expense structures. We explore that context in the overhead question.

07/record extends beyond

The record extends beyond the base form

Base-form totals are only the beginning of a filing. A grant disclosed on a schedule can carry an amount, recipient and source filing reference. That specificity is what lets a reader check an apparent connection rather than relying on a single grand total. Other schedules reveal different kinds of relationship. We explore two of them in the mysterious Schedules B and R.

We also compare filings with independently captured organization websites. We currently have captured public-web evidence for about 33,500 organizations. These captures can surface a published mission and policies, sometimes with descriptions of impact that a tax form does not convey. They are a selected, evolving subset of profiles and should never be mistaken for an all-sector sample or for filed financial facts.

The pattern we kept finding

More sources answer more questions, but each source has its own boundary. A filed transfer can name a counterparty. A website can explain an organization’s work. Neither substitutes for observing the outcome of a program.

08/Behind the Data

Behind the Data

We checked the current CharitySense filing corpus on September 25, 2026. The public-profile and historical filing counts come from the sitewide snapshot computed September 17. Its $4.69 trillion revenue figure selects the most recent primary Form 990, 990-EZ or 990-PF for each profile and adds the revenue those forms report. The brief Form 990-N and unrelated-business Form 990-T do not enter that sum. It measures revenue visible in those reported fields, with no estimate for organizations outside the filing record.

The like-for-like financial scan uses the most recent full Form 990 for each of 485,109 organizations. We read reported total revenue, year-end assets, and program-service expenses from those returns. Negative reported revenue remains in the median, mean and ranked top-share calculations. The revenue-band chart separately uses the 482,013returns with nonnegative revenue and their combined revenue, leaving out 3,096 negative values. The program ratio uses the 436,970 returns with positive total expenses and a numeric program-expense value.

The historical animation counts observed full Form 990 filing rows by tax year and indexes each series to 2010. Its cohorts change over time, and the incomplete 2025 set is omitted from the chart. We classified all 1,588,539 source-cited money-flow records by the filing section that disclosed them; 845,475 came from private-foundation grant lists and 698,924 from the domestic grant schedule. These are records, not unique annual transfers or a measure of all nonprofit spending. The website count describes public pages captured for a selected set of organizations. Neither source represents complete sector coverage. Churches are generally exempt from annual 990-series filing; Form 990-N gives a gross-receipts threshold rather than exact revenue. For a practical way to inspect an individual filing, see how to evaluate a nonprofit before you donate.

For how we turn e-files into comparable records, see the methodology. For the public-search argument behind this work, read public 990 data should be free to search.

09/Cite this analysis

Cite this analysis

Sources

What would you want us to pull from the filings? Tell us, and a future piece may answer it.

Final Thoughts

We began with a large number and expected it to tell us what a typical organization looks like. The filings changed that question. In the current full-990 cohort, half report no more than $361,386 in revenue, while the top 1% account for about 71% of the dollars. The forms also changed how much we could know: some disclose detailed expenses, while the smallest notice gives only a threshold.

We can now follow some transfers to named counterparties and compare filed facts with public websites. That gives us more useful questions, without providing a shortcut to an outcome measure. Each answer starts with a well-defined filing cohort and its tax year.

About the author

The CharitySense Data team

Research team

We read public nonprofit filings and other sources to understand how the sector works. This blog shares what we find, along with the limits of the evidence. Each article explains how its figures were measured and cites the records or primary sources behind them.

How we build the datasetSearch the corpus