AI statistics
Reference, maintained weekly from primary sources
How much AI research is actually published, what the companies building these systems spend on research, how many annual reports now mention artificial intelligence, what the US government pays for AI work, and how much AI rulemaking is being written. Counted from arXiv, from company filings and from the government’s own records. No vendor numbers, no benchmark tables, no forecasts. Quote any of it with credit.
How much AI research is actually being published?
The artificial intelligence category alone went from 7,418 papers in 2020 to 45,161 in 2025, 6.1x in 5 years. Not everything grew: neural and evolutionary computing is smaller than in 2020.
| arXiv category | 2020 | 2025 | Scale in 2025 | Change |
|---|---|---|---|---|
| Machine learningcs.LG | 25,892 | 46,022 | 1.8x | |
| Artificial intelligencecs.AI | 7,418 | 45,161 | 6.1x | |
| Computer visioncs.CV | 15,340 | 34,987 | 2.3x | |
| Computation and languagecs.CL | 7,125 | 23,762 | 3.3x | |
| Roboticscs.RO | 3,972 | 10,564 | 2.7x | |
| Neural and evolutionary computingcs.NE | 1,751 | 1,613 | 0.9x |
Source: arXiv submission counts by calendar year
What do the companies building this spend on research?
| Company | Latest year | Scale | Five years earlier | Change |
|---|---|---|---|---|
| AlphabetFY2025 | $61.1B | $27.6BFY2020 | 2.2x | |
| MetaFY2025 | $57.4B | $18.4BFY2020 | 3.1x | |
| MicrosoftFY2026 | $35.6B | $20.7BFY2021 | 1.7x | |
| AppleFY2025 | $34.5B | $18.8BFY2020 | 1.8x | |
| NvidiaFY2026 | $18.5B | $3.9BFY2021 | 4.7x | |
| IntelFY2025 | $13.8B | $13.6BFY2020 | 1.0x | |
| IBMFY2025 | $8.3B | $6.3BFY2020 | 1.3x |
Source: SEC EDGAR, each company’s own Form 10-K
Amazon is absent because it reports “technology and infrastructure” rather than a research and development line, so there is no comparable figure to show. Leaving it out is more honest than approximating one.
How many companies even talk about AI?
The number of annual reports filed with the SEC whose text contains the phrase “artificial intelligence”. In 2020 it was 558. In 2025 it was 3,324. A mention is not a business, though: companies name AI in their risk factors at least as often as in their plans.
| Year | Annual reports mentioning AI | Count |
|---|---|---|
| 2020 | 558 | |
| 2021 | 848 | |
| 2022 | 1,156 | |
| 2023 | 1,295 | |
| 2024 | 2,436 | |
| 2025 | 3,324 |
Source: SEC EDGAR full-text search, Form 10-K
How fast did agents arrive?
The word arrived in company filings in a single year
How many filings of any kind lodged with the SEC contain each phrase. Not annual reports only: prospectuses and quarterly filings pick up new language first.
| Phrase in SEC filings | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|---|
| “agentic AI” | 0 | 0 | 0 | 0 | 16 | 893 |
| “AI agents” | 1 | 1 | 1 | 5 | 85 | 719 |
| “AI agent” | 1 | 7 | 21 | 27 | 43 | 417 |
And the research got there first
arXiv papers whose abstract mentions an LLM agent, an autonomous agent or anything agentic: 3,947 in 2020, 14,883 in 2025, 3.8x in 5 years. The papers were climbing for years before the word reached a single annual report.
| Year | Agent papers on arXiv | Count |
|---|---|---|
| 2020 | 3,947 | |
| 2021 | 4,382 | |
| 2022 | 4,686 | |
| 2023 | 5,788 | |
| 2024 | 8,199 | |
| 2025 | 14,883 |
Sources: SEC EDGAR full-text search, and arXiv abstracts
A filing that names agents is not a company that has shipped one, and a paper is not a product. These count how far the subject has travelled, not how well anything works.
What is the US government paying for AI?
$429.6M in federal contract awards matching “artificial intelligence” in the fiscal year to 30 September 2026, across 10 agencies. Department of Defense accounts for 61 percent of it. Read this as a floor rather than a total: it catches contracts that describe AI work in their text and misses every one that does not say so.
| Agency | Share | Awarded |
|---|---|---|
| Department of Defense | $260.7M | |
| Department of Homeland Security | $39.0M | |
| Department of Health and Human Services | $37.8M | |
| Department of the Treasury | $27.5M | |
| Department of Commerce | $25.8M | |
| General Services Administration | $14.6M | |
| Department of Veterans Affairs | $12.8M | |
| Department of Labor | $6.3M | |
| National Science Foundation | $2.5M | |
| Department of Agriculture | $2.5M |
Source: USAspending.gov, federal fiscal year 2026
How much AI rulemaking is being written?
Documents published in the Federal Register, the daily journal of the US government, whose text mentions artificial intelligence. These are rules, proposed rules, notices and presidential documents. It counts documents, not laws, and a single policy can run to several of them.
| Year | Federal Register documents mentioning AI | Count |
|---|---|---|
| 2020 | 105 | |
| 2021 | 107 | |
| 2022 | 113 | |
| 2023 | 189 | |
| 2024 | 309 | |
| 2025 | 182 |
Source: federalregister.gov
Questions people ask about these numbers
How many AI research papers are published each year?
45,161 preprints were posted to arXiv's artificial intelligence category in 2025, against 7,418 in 2020. Machine learning saw 46,022 and computer vision 34,987. Papers are often cross-listed in several categories, so these overlap and should never be added into one total.
Is AI research still growing, or has it levelled off?
It is still growing, but unevenly. arXiv's artificial intelligence category is 6.1x its 2020 size, and papers about agents are 3.8x theirs. Neural and evolutionary computing, an older corner of the field, is smaller than it was in 2020. Growth in "AI" is not growth everywhere in AI.
How much do the big technology companies spend on AI?
Nobody can tell you honestly, because no company reports an AI line in its accounts. What is filed is total research and development: $229.2B across 7 listed companies in their latest annual reports, led by Alphabet at $61.1B. Treat any published "AI budget" for these firms as an estimate someone has made up.
What percentage of companies mention AI in their filings?
3,324 annual reports filed with the SEC used the phrase "artificial intelligence" in 2025, up from 558 in 2020, about 6.0x. We publish the count rather than a percentage because the number of filers changes each year, so a percentage would move for reasons that have nothing to do with AI.
How many AI agents have been launched?
There is no trustworthy count, and this page does not publish one. Launches are company announcements, and the trackers that collect them disagree on names and dates. What can be counted is spread: "agentic AI" appeared in 893 SEC filings in 2025 and none at all before 2024.
How much is the US government spending on AI?
$429.6M in federal contract awards matching "artificial intelligence" in the fiscal year to 30 September 2026, across 10 agencies, with Department of Defense accounting for 61 percent. Read it as a floor: it finds contracts that describe AI work in their text and misses every one that does not.
Where does this data come from, and can I quote it?
arXiv for research, SEC EDGAR for company filings, USAspending.gov for federal contracts and the Federal Register for rulemaking. All are primary sources, all are free to check, and each figure on this page carries its own date. Quote any of it with credit to The Super Intelligence News and a link to this page.
Why does this page stop at 2025 instead of showing this year?
Because a part year set beside whole years invents a decline that has not happened. Papers, filings and rulemaking all arrive unevenly through a year, and most annual reports land in the first months of the next one. Every table here ends at the last complete calendar year for that reason.
Method, and what these numbers are not
Research. Counts of preprints submitted to each arXiv category in each calendar year. A paper can be cross-listed in more than one category, so the categories overlap and must not be added together.
Company spending. Total research and development expense as filed. It is not an AI budget: no company reports one, and this page does not invent one. Only full year periods are counted: the raw data mixes quarterly rows, segment rows and restatements, so a row is used only when its start and end span a full year, and where a year was restated the newest filing wins.
Filing mentions. The number of annual reports whose text contains the phrase. A mention is not a business: companies name AI in risk factors as often as in plans.
Federal spending. Contract awards whose text matches the phrase, for the federal fiscal year that ends 30 September. Keyword matching catches contracts that describe AI work and misses ones that do not say so, so read it as a floor, not a total.
Rulemaking. Documents published in the Federal Register whose text matches the phrase: rules, proposed rules, notices and presidential documents. It counts documents, not laws.
The current year is excluded from every table on this page. Research, filings and rulemaking all arrive unevenly through a year, so setting a part year beside whole ones invents a decline that has not happened.
This page states what the numbers are, never what they will be. No forecasts, no predictions about capability, and nothing here is investment advice.
Every figure is regenerated from the source by a script in our repository and refreshed weekly.
Citing this page
Free to quote with credit to The Super Intelligence News and a link to this page. If you need a cut we do not show, or you think a figure is wrong, write to us and we will check it against the source and correct it in the open.