Who Is Building AI
AI Companies Get Breached Too
In one respect, AI companies are no different from any other supplier: they get breached.
In the past two years, a model developer left a database of user chat logs open to the internet, a training-data provider exposed client project documents through public links, a model-hosting platform detected an intruder on its systems, the best-known AI company in the world lost customer contact data through its analytics vendor, an AI productivity tool’s integration became the way into a customer’s environment, and Straive, an AI data services company, turned up on a ransomware leak site.
Six companies, six different doors. The question for anyone buying AI is not whether their suppliers can be breached, but what they are handing those suppliers, and what happens when one of them is.
The AI Ecosystem Is Larger Than Its Most Visible Names
When people think of AI, they often picture the companies behind the world's best-known models and chatbots. But those companies represent only one part of a much larger ecosystem.
AI is also being developed for everyday business needs and specialized industries. A law firm may use AI to review documents, a manufacturer to inspect production lines, and a customer service team to respond to thousands of inquiries. Behind each application is a company with its own technology, infrastructure, and suppliers.
An AI Supplier Is Not One Kind of Company
Black Kite identified more than 18,000 AI companies spanning this ecosystem. To understand what these companies represent, the research examined not only where they operate, but what they actually build and the role they play in delivering AI products and services.
Not all AI companies serve the same purpose. Some develop the models, computing infrastructure, and data that make AI possible. Others help businesses develop and implement AI systems. A third group builds AI-powered products for everyday work or specific industries.
Black Kite classified the companies into six broad groups based on their primary AI activity; the percentages below are shares of all AI companies in the study. This distinction helps explain the different roles companies play in the AI supply chain, and why assessing them all in the same way can overlook important differences.
Companies Building the Foundations of AI
These companies develop the technologies that other AI products depend on. They include companies building AI models and platforms (6.9%), providing the computing infrastructure needed to train and run AI systems (2.0%), and specializing in the data and labeling used to develop and improve models (1.1%).
This group includes the companies creating the models themselves, the hardware and computing systems that power them, and the data services that support their development.
Not every business has the expertise or resources to build its own AI systems. This group includes companies that help organizations develop, integrate, and operate AI technologies.
The largest subgroup consists of AI consulting and development companies (13.8%) that build or implement AI solutions for clients. The remaining companies specialize in software development (3.2%), data and analytics (3.1%), and cloud solutions (2.7%).
Unlike companies selling a finished AI product for a specific task, many businesses in this group provide the technical expertise, software, or infrastructure that other organizations need to build their own AI capabilities.
Companies Helping Businesses Develop and Implement AI
AI for Everyday Business Operations
This is the largest group in the research. These companies develop AI-powered products that help businesses carry out everyday work, from serving customers and managing sales to creating marketing campaigns, recruiting employees, and automating internal processes.
The largest categories are business productivity and workflow automation (8.5%), customer service (7.3%), and marketing and advertising (6.0%). Others focus on sales (3.2%), HR and recruiting (2.3%), cybersecurity (1.8%), legal services (1.5%), and finance and accounting tasks (0.7%).
Depending on their purpose, these applications may process customer conversations, business documents, employee information, or other operational data.
This group includes companies bringing AI to consumers, creators, and digital entertainment.
Content creation and design is the largest category, with 6.0% of the study population developing AI-powered creative tools. Another 4.0% focus on consumer and lifestyle applications, while the remaining companies work in media and entertainment (1.5%), travel and hospitality (0.9%), and gaming (0.6%).
Many of these products are designed around creative work, personal use, and consumer-facing experiences rather than internal business operations. Their security considerations may involve personal information, user-generated content, and connections to external digital services.
AI for Content, Entertainment, and Consumers
AI for Industry and the Physical World
AI development is not limited to software applications and digital services. Companies are also applying AI to physical operations, equipment, and infrastructure.
The largest category focuses on manufacturing and industrial applications (2.1%), followed by real estate and construction (1.6%), transportation and automotive technologies (1.2%), and supply chain and logistics (0.7%). Smaller groups work in energy and utilities (0.4%), food and beverage (0.4%), agriculture (0.3%), and telecommunications (0.2%).
These applications can support equipment monitoring, production processes, transportation, and operational decision-making.
For customers, cybersecurity risks in these environments may extend beyond information security to the availability and reliability of the operations supported by AI.
Some AI companies develop products for industries with distinct operational needs, sensitive information, or specialized requirements.
Healthcare and life sciences is the largest category in this group, representing 4.6% of the study population. Financial services and insurance accounts for 3.2%, followed by retail and e-commerce (2.6%), education (2.4%), and government, defense, and scientific research (1.5%).
These applications serve very different environments. An AI product supporting a hospital, a financial institution, a school, or a government agency may process different types of sensitive information and interact with different systems.
Understanding the supplier's intended industry and use case helps determine which security requirements matter most.
A further 1.7% of companies could not be assigned to a sufficiently specific AI activity category and are not included in the six groups above.
AI for Healthcare, Financial Services, Education, and Other Specialized Sectors
The distribution reveals an important characteristic of the AI ecosystem: Most companies are not developing foundational AI models or infrastructure. They are building applications, developing solutions for other businesses, or applying AI to specific industries.
Together, companies developing AI for everyday business operations and those helping other organizations build and implement AI account for more than half of the companies identified in the study.
This distinction matters. An AI model provider, a company developing a custom AI solution, and a supplier offering an AI-powered customer service platform may have very different relationships with their customers. They may process different types of information, depend on different technologies, and require different levels of access to customer systems.
Understanding what an AI company actually provides is therefore an important first step in understanding the cyber risks associated with using its products or services.
Where AI Is Being Built, and for What
The global AI ecosystem is not developing in the same way everywhere. While AI companies can be found across a wide range of countries, the technologies and applications they focus on vary considerably between major regions.
The United States accounts for 39.7% of the companies in the study, followed by Europe at 25.0%, India at 7.2%, China at 6.7%, and Canada at 3.6%. Taiwan, which accounts for 1.7%, is considered separately from China.
The more interesting finding is what they are building.
The United States and Europe show a strong focus on AI for everyday business operations.
Nearly one-third of US AI companies
develop applications for activities such as customer service, sales, marketing, and workplace productivity. Europe shows a similar concentration at 33.4%.
India stands out for AI development and implementation
This is its largest category, at
followed by everyday business applications at 29.6%. Canada is more evenly divided between business applications, at 29.3%, and development and implementation, at 28.4%.
China presents a notably different picture.
While business applications dominate the AI company profiles of the United States and Europe, the Chinese companies in Black Kite's research show a much stronger concentration in foundational AI technologies and industrial applications.
Approximately 22.5% of the Chinese AI companies develop the underlying technologies that power AI, including models, computing infrastructure, and data services. That is nearly twice the corresponding share in the United States (11.3%).
The difference becomes particularly interesting when looking at the technologies within this category.
Computing and infrastructure specialists account for 10.7% of the Chinese AI companies in the study, compared with 1.7% in the United States. In absolute terms, the study includes 129 computing-infrastructure companies in China and 119 in the United States, even though the US group is six times larger. Model and platform developers represent more similar shares: 9.2% in China and 8.7% in the US. The larger difference within foundational AI is therefore in computing infrastructure, rather than in the share of companies developing models and platforms.
Another major difference appears in industrial applications. Among Chinese AI companies, 22.7% focus on industries connected to the physical economy, compared with just 5.3% in the United States.
Manufacturing and industrial applications account for 10.9% of the Chinese AI companies in the study, compared with 1.1% in the United States. Transportation and automotive applications represent 6.9% and 0.7%, respectively.
These results reveal different areas of concentration within the AI ecosystems identified by the research. Business applications represent a larger share of the US group, while computing infrastructure, manufacturing, and transportation-related AI represent larger shares of the China region.
For organizations evaluating AI suppliers, these differences are more than a geographic observation.
An AI company providing a customer service application, a company developing AI systems for manufacturers, and a provider of AI computing infrastructure occupy very different positions in the technology supply chain.
Their cybersecurity risks may therefore involve different assets, dependencies, and forms of customer access.
These distinctions matter more as AI products and services become part of increasingly complex technology supply chains.
Capital tells the same story from another angle. Funding is disclosed for roughly one in four companies in the study, 4,686 in all, and those disclosures total $632 billion. More than four-fifths of it sits with companies building models, computing infrastructure, and data, a group that makes up one in ten AI companies; three of them, OpenAI, Anthropic, and xAI, account for two-thirds of the total on their own. The rest of the ecosystem, the application and integration companies most organizations actually contract with, is built on a median of about $2 million in disclosed funding.
The country where an AI company is headquartered is only one part of that picture. The more important questions are what the supplier builds, which technologies it depends on, and what information or systems its customers entrust to it.
One-Third of the Ecosystem Has 10 or Fewer Visible Assets
In every region, the median AI company has 15 to 20 externally visible assets, about a third have ten or fewer, and the six activity groups spread similarly.
For 33.2% of the AI companies in the study (6,032 of 18,160), Black Kite can observe 10 or fewer externally visible assets: domains, subdomains, and IP addresses combined. The typical AI company has just under 20. The median is 18 assets in every region, and only a thin tail of companies reaches into the thousands.
The pattern is consistent across regions. The share with 10 or fewer assets runs from 29.3% in Taiwan to 38.9% in India, and the six activity groups are spread similarly in each. There is also no sharp line at 10 assets. Footprints form a continuum, so a company with 11 assets has little more to observe than one with 9. The threshold is a convenient cut, not a separate class of company.
Footprint doesn’t explain the regional differences. Because the typical footprint is alike across regions, it does not by itself explain the regional differences in security findings. The US and China, for example, both have a median of 19 assets, and 32.6% and 30.9% of their companies, respectively, have 10 or fewer. Yet critical vulnerabilities appear at 54.5% of China’s AI companies against 28.7% of the US companies, a gap the next sections examine.
At a small company, a short list of findings says less. Findings can only appear where there is something to observe, so at a company with few visible assets, a short list of findings, or none, says less than the same result would at a larger one. That doesn’t mean the company is secure, and it doesn’t mean it isn’t. It means less of it can be seen.
A footprint isn’t fixed. As a company adds products, domains, and infrastructure, what can be observed changes, and so do the findings. Continuous monitoring picks up that change as it happens, while a point-in-time review captures only the footprint that existed on the day it was done. That makes it practical to adopt an AI vendor now and keep watching as its footprint develops, instead of waiting for more evidence before deciding.
How findings change with company size, age, and funding is covered in the appendix.
Next: Knowing who builds AI is the starting point.
The next question is how well those companies protect themselves. Black Kite measured the first layer, the AI company itself, from the outside.