No single ranking of the "giants" of artificial intelligence exists, because leadership depends on what is counted. Stanford's 2026 AI Index reports that industry produced over 90 percent of notable AI models in 2025, and that the United States produced 59 of them against China's 35. A second Stanford project, the Foundation Model Transparency Index, found that the average score among 13 major developers fell to 41 out of 100 in December 2025.
Key takeaways
- Companies, not universities, now build nearly all notable AI models: over 90 percent in 2025, according to the 2026 AI Index.
- Global AI computing capacity has grown about 3.3 times per year since 2022 and reached 17.1 million H100-equivalents.
- Training details are no longer disclosed for several of the most resource-intensive systems, including some from OpenAI, Anthropic, and Google.
- Developer transparency scores diverge widely: IBM scored 95 and xAI and Midjourney scored 14 in the 2025 Transparency Index.
- Counts of models, compute, and transparency can each produce a different ordering, so a single "top" list is a choice of measure.
Why "leading" needs a measure
Lists of the biggest names in AI usually mix unlike things: company size, model quality, user numbers, research output, and fame. Each can be counted, but the answers differ. A firm may build many models but disclose little, or run the most users but release few papers. Rather than rank organizations by an unstated blend, the sections below use three published measures from Stanford University, each with its own definition and limits: who produces notable models, who controls computing capacity, and who explains how the models were built.
Who produces notable models
The Stanford Institute for Human-Centered AI publishes the AI Index each spring. The research and development chapter of the 2026 edition states that industry produced over 90 percent of notable AI models in 2025, with academia and industry-academia collaborations making up the rest. For countries, it reports that the United States produced 59 notable models in 2025 and China 35, and the report's summary notes that models from the two countries have traded the lead on benchmarks several times since early 2025.
A count of notable models is only as neutral as the rule that decides which models are notable. It also treats a small specialized model and a very large general one as one unit each, so a high count does not show the highest-performing system.
Who has the computing capacity
The same chapter reports that training compute, which can be estimated independently of company statements, has continued to rise. Global AI computing capacity has grown 3.3 times per year since 2022 and reached 17.1 million H100-equivalents, a unit that expresses different chips in terms of one widely used AI processor. Parameter counts, by contrast, have stayed near 1 trillion for three years.
Capacity is a measure of potential, not ownership of the best product. The report counts capacity worldwide and does not rank companies by it, so the figure cannot be turned into a list of leaders.
Who explains how the models were built
Disclosure is where the picture is least flattering. The AI Index notes that training details are no longer disclosed for several of the most resource-intensive systems, and it names OpenAI, Anthropic, and Google as examples. Disclosure of training code, parameter counts, dataset sizes, and training duration has dropped for these models.
The Foundation Model Transparency Index, from Stanford's Center for Research on Foundation Models, puts numbers on this. Its December 2025 edition scored 13 companies. The mean score was 41, down 17 points from 2024, and training data remained the least transparent area. Only 30 percent of the companies it contacted agreed to submit transparency reports, down from 74 percent in 2024.
| Developer | Score in the December 2025 Index (out of 100) |
|---|---|
| IBM | 95 |
| DeepSeek | 32 |
| Alibaba | 26 |
| xAI | 14 |
| Midjourney | 14 |
The Index reports that open-weight developers outscore closed-weight developers overall, yet high-profile open developers such as Alibaba, DeepSeek, and Meta still land in the bottom half. A low score means little public information on the Index's indicators. It does not by itself mean a model is poor or unsafe.
What the three measures show together
- Production is concentrated in companies, so claims about AI progress mostly rest on corporate reporting.
- Capacity is growing quickly, while the public record of how the largest systems are trained is thinning.
- Leadership by output (notable models), by capacity (computing), and by openness (transparency) points to different organizations.
- Rankings from magazines or vendors that omit their method cannot be compared with these sources.
How to read a "top AI companies" list
A list is easier to judge if it states what it counts, the year, the data source, and who funded the ranking. A list of the most valuable companies reflects investor expectations, a list of most-cited labs reflects research output, and a list of popular apps reflects consumer habits. None says which organization builds the most capable or the most trustworthy systems. Rankings that change with each publication, or that offer no definition of "leading," deserve less weight than public datasets with documented methods.
FAQ
Which company is the biggest in AI?
That depends on the measure. By notable models, industry as a whole leads academia by a wide margin, but the sources used here do not rank individual companies, and rankings by revenue, users, or valuation are different measures.
Do universities still matter in AI?
Yes, but mostly through research, training, and public interest work rather than by building the largest models. The AI Index counts over 90 percent of notable models as industry-made.
What is an H100-equivalent?
It is a unit that expresses the computing power of various AI chips in terms of Nvidia's H100 processor, which allows capacity from different hardware to be added up in one figure.
Does a low transparency score mean a model is unsafe?
No. The Transparency Index measures how much developers disclose, such as about training data and methods. It does not test whether a model works well or safely.
Related reading
More in AI Basics.
Last verified: October 2026
Sources
- The 2026 AI Index Report, research and development chapter, Stanford Institute for Human-Centered AI, 2026
- The 2026 AI Index Report, overview page, Stanford Institute for Human-Centered AI, 2026
- Foundation Model Transparency Index, December 2025 edition, Stanford Center for Research on Foundation Models, 2025