AI unicorns: half of 317 have no qualifying papers

Bibliometric analysis of 317 AI unicorn startups, 1998 to 2025

The preprint is a bibliometric analysis, which means it counts publications and the citations they receive, for 317 AI unicorns between 1998 and 2025. The authors defined an AI unicorn as a privately held AI company valued at US$1 billion or more, and built the list by merging two public rankings of AI startups, one from Failory and one from Eqvista, both read on February 4, 2026. One company appeared twice under an alias, which left 317 companies headquartered in 25 countries.

For each company the authors searched Clarivate databases such as Web of Science, plus the Preprint Citation Index, for papers whose author affiliations name it. Five filters cut 144,185 records down to 3,410. A paper then counted as a qualifying output only when one of these held:

• a researcher from the company was first or last author
• the authors were listed in alphabetical order
• two companies tied for the most authors on the paper

That left 2,077 qualifying outputs, made up of 1,389 peer-reviewed publications and 688 preprints. The other 1,333 records, where company staff appeared only in middle author positions, were set aside as collaborative-only. Most figures below rest on this counting rule, but the 2025 share of the literature also includes the collaborative-only records.

Half of billion-dollar AI unicorn companies have no qualifying papers

Of the 317 companies, 166 (52.4%) produced no qualifying paper or preprint. The other 151 did publish, and their median was 4 papers, though one company reached 243. Fewer than half of the companies publish at all, and a paper that other researchers cite heavily is rarer still.

The authors call a paper highly cited when it has at least 200 citations, meaning at least 200 other papers cite it. Only 24 companies (7.6%) have one, and 293 have none. Among those 24, the median is 1.5 highly cited papers per company.

The top 10% of AI unicorn startups hold 96.8% of citations

The top 10% of the 317 companies hold 96.8% of all citations. The bottom half account for essentially none.

The authors measure this gap with the Gini coefficient, a score from 0 (everyone has an equal share) to 1 (one company has everything). For citations it is 0.97, close to the maximum of 1, so nearly all citations sit with a few companies.

A few names carry most of that weight. OpenAI has the largest single share of all citations, followed by MEGVII and Hugging Face.

The abstract adds that three startups produced 92 of the 134 firm-attributed highly cited papers. So saying that half publish nothing understates the picture: a handful of companies account for most of the highly cited record.

AI startup valuation does not predict papers at AI unicorns

A company's valuation shows no link to how many papers it publishes. The authors tested this with the Spearman correlation, which asks whether two rankings move together: 0 means no pattern and 1 means the rankings match exactly. Across all 317 companies, valuation against paper count gives 0.017, effectively zero and not statistically significant, and valuation against highly cited papers gives 0.042, also not significant.

Funding raised tells a slightly different story: it has a weak but statistically significant link to paper count, at 0.181, and the authors say it explains little of the variation. The authors suggest in their discussion that investors apparently pay little attention to the peer-reviewed record of these companies. That is their reading of the numbers, not something the study measured.

AI unicorns made up 0.1% of the AI literature in 2025

AI unicorns accounted for 0.103% of AI publications in 2025, up from 0.019% in 2016. Papers involving these startups rose from 23 in 2016 to 950 in 2025. Over the same years, the AI literature in the same databases grew from 123,383 to 919,486 publications.

Startup-led papers climbed from 18 to 534, and collaborative-only papers from 5 to 416. The companies' output grew fast, yet it stayed tiny next to the whole field, which is why the authors call their participation negligible.

Chinese AI unicorn startups publish more than US ones

Chinese companies publish the most: 26 of 40 produced at least one qualifying paper, with a median of 3.5 papers per company. US companies publish the least: of 209, only 90 produced one, and the median is 0. Even the Chinese figure is low, and the regional gap is statistically significant (p = 0.003).

Qualifying papers by region of headquarters

Region	Companies	Published	Share	Median papers
China	40	26	65.0%	3.5
Europe	35	20	57.1%	1.0
Other regions	33	15	45.5%	0
United States	209	90	43.1%	0

The authors suggest that organizational context and scientific norms may differ by country. The data cannot say why, only that the pattern exists.

What the preprint on billion-dollar AI startups cannot show

The study counts formal publishing, not how capable or successful a company is. The authors state the limits themselves, and the paper itself has not yet been peer reviewed. They list these:

• Matching on company names in author affiliations may undercount work tied to undisclosed affiliations, informal collaboration, or internal industrial research.
• Citation counts depend on how common citing is in a field.
• Publications and citations do not directly measure technological capability, commercial success, or societal impact.

Not publishing is also a choice the authors treat as defensible: they write that organizations may reasonably protect proprietary innovations, and that whether firms should publish as a societal requirement is a matter of debate. What worries them is transparency, reproducibility (other teams being able to repeat the work) and accountability. They also point to Theranos, named in the paper as a case of fraud, as a highly valued startup that produced little peer-reviewed evidence.

The preprint is also available as a PDF. To read the methods and exclusion criteria instead of only the abstract, try the tool that converts a PDF to Markdown.