According to a paper published by a Stanford University team on the preprint platform bioRxiv, the publication records of 317 AI unicorns between 1998 and 2025 were analyzed. The results showed that over half of AI unicorns have not publicly released any research papers, with scientific output highly concentrated among a few companies. The top 5% of companies accounted for over 90% of citations, with OpenAI alone contributing nearly 40%. In 2025, global AI papers exceeded 900,000, but papers related to these startups comprised only 0.1%. Notably, nearly two-thirds of Chinese AI unicorns have published research, while over half of U.S. companies have not published any papers—Chinese companies are more actively embracing open-source models.

Article author and source: Academic Headline

Recently, Science magazine published a news article titled “AI’s Top Startups Are Barely Publishing Their Research.”

This article focuses on a paper published by a Stanford University team on the preprint server bioRxiv. The research team analyzed the publication records of 317 AI unicorn companies between 1998 and 2025 and discovered a counterintuitive phenomenon:

More than half of AI unicorns have not published public research papers, and scientific output is concentrated among a few AI unicorns, with only a small number of core authors producing content.

Paper link: https://www.biorxiv.org/content/10.64898/2026.07.15.738744v1

At the same time, a higher company valuation does not necessarily mean more paper output. Although the number of papers published by these companies has increased over the past decade, their overall share remains very low. Among AI-related papers published in 2025, only one out of every 1,000 came from these AI companies.

The Science article also states that leading U.S. frontier labs more often adopt a "closed-source" model, while leading Chinese companies are actively embracing the "open-source" model. Nearly two-thirds of Chinese AI unicorns have published papers, while more than half of U.S. companies have not published a single qualifying paper.

The corresponding author of the paper, Stanford University meta-scientist John Ioannidis, expressed concern about this phenomenon: “It is a very strange paradox that, for a field claiming to be reshaping science and regarded as having tremendous scientific potential, there is almost no scientific literature.”

Only a few AI companies are producing output.

The text points out that most AI unicorn companies do not publish many papers. Of the 317 companies, more than half have no qualifying scientific output. Even among those with publication records, most publish only sporadically, and very few meet the threshold for highly cited work—accounting for just 6.4% of the entire sample.

The visible influence of AI unicorn companies is also dominated by a small number of leading firms. According to citation metrics, the top 5% of companies account for over 90% of all citations. OpenAI alone contributes nearly 40% of citations, followed by Megvii and Hugging Face. High-citation papers are concentrated among a few companies, with only 7.6% of companies producing such results.

In addition, companies differ in their choice of publication channels, particularly in the proportion of preprints used. Anthropic derives nearly 90.6% of its citations from preprints, while companies such as OpenAI, Megvii, Waymo, and Momenta primarily rely on peer-reviewed papers.

Figure | Distribution and concentration of scientific outputs at the company level.

Only a few people are producing.

A small group of core authors consistently produce paper content. Among nearly 2,000 affiliated authors from startups, more than half are affiliated with their companies; 38.0% of authors are also affiliated with universities or research institutions.

The papers are primarily produced by a small number of authors. Among the 132 highly cited papers, 27 prolific authors accounted for nearly 40% of the authorship, while most individuals contributed to only one paper. Specifically, the research teams at MEGVII and OpenAI each exceed 100 members, yet only eight authors from each have published more than five papers. Compared to their overall workforce of thousands, this proportion remains very low.

High valuation didn't lead to more papers.

A higher valuation for an AI company does not necessarily mean more papers or greater impact. Data shows no significant correlation between company valuation and the total number of papers or the number of highly cited papers. Although companies with more funding are somewhat more likely to publish papers, this association is weak. High-impact research remains concentrated among a few companies and does not scale proportionally with funding levels.

Accounts for only 0.1%

From a time perspective, the number of papers involving AI unicorns has indeed grown rapidly. In 2016, startup-led papers numbered just 18; by 2025, this figure had increased to 534. The growth in collaborative papers has been even more pronounced, rising from 5 in 2016 to 416 in 2025.

However, when viewed globally, these numbers are negligible. In 2025, the total number of AI papers worldwide exceeded 900,000, while papers related to these startups amounted to only 950, accounting for just 0.1%.

Chinese AI companies prefer to publish papers.

Regionally, Chinese AI unicorns are the most active in publishing papers, although the number remains relatively low.

Among the 40 Chinese AI unicorns, nearly two-thirds have published papers; in contrast, among the 209 U.S. companies, more than half of the AI unicorns have not published any papers. A Science article notes that leading U.S. frontier labs increasingly adopt a "closed-source" model, while leading Chinese companies are actively embracing "open-source" models.

Figure | Economic scale, temporal changes, and geographic distribution of scientific activities.

The impact of this situation extends beyond the company itself: only when research is publicly published do researchers, policymakers, and industry observers have the material to engage in discussion. If research remains confined within the company, only a select few will understand the direction of AI development.