The Upstream AI Economy Is Concentrated in Different Places

To judge whether the consumer AI assistant market is closed, we also need to understand the production system behind the interface. The question here is not which consumer platform has the most users, but where the firms that make AI capabilities are located, which supply-chain layers they occupy, and how capital and investment connect them. This distinction helps us recognize upstream concentration without incorrectly treating it as proof of downstream consumer control.

A chat window conceals a chain of capabilities. Hardware has to support computation; infrastructure has to host and scale workloads; data and software tools have to help build systems; models have to be developed; and applications have to reach end users. These layers can be concentrated in different economies and can have different economic bottlenecks. A country that is strong in chips is not necessarily strong in consumer applications, and a large application market is not automatically a source of model or infrastructure independence.

The Bank for International Settlements study is useful because it maps this production structure rather than counting only famous consumer brands. It identifies 1,246 AI-producing firms across 32 economies, including public and private companies valued above $500 million. The sample begins with commercial database tags, uses an LLM-assisted screen to separate producers from AI users or unrelated companies, and applies manual verification. The result is a structured map of large producers, not a complete census of every AI company.

Five layers behind an AI product

The first layer is compute: the hardware ecosystem that designs, manufactures, packages, tests, or sells chips, boards, servers, and networking equipment used for AI workloads. Compute is a physical bottleneck. Its firms are exposed to manufacturing capacity, specialized equipment, and the investment needed to expand supply.

The second is cloud and related infrastructure: data centers, colocation, cloud services, managed compute clusters, and other services that host and scale workloads. Infrastructure sits between hardware and use. A firm may not design a chip or a model, but it can determine how easily a model is trained, served, and made available to users.

The third is data tools: data, licensing, labeling, reinforcement learning from human feedback, and software used to train, deploy, monitor, or evaluate models. This layer is less visible than a data center but still affects how systems are built and maintained. It includes services that turn raw information and human judgments into inputs a model-development process can use.

The fourth is AI models: firms developing general-purpose machine-learning models or model services. Models are the component most closely associated with the public debate about advanced AI, but they are only one layer. A model needs compute, infrastructure, data, and an application or interface before it becomes a consumer experience.

The fifth is AI applications: end-user products that integrate models into services such as healthcare, finance, legal work, education, or autonomous vehicles. Consumer assistants sit closest to this downstream layer, although a particular assistant can depend on firms in several layers at once. An application may differentiate itself through its task design, distribution, trust, or privacy controls while relying on production capabilities it does not own.

The layer distinction matters because “AI concentration” can mean several different things. It might mean that a few firms supply critical hardware. It might mean that cloud capacity is concentrated. It might mean that model development is controlled by a small number of organizations, or that applications and distribution are concentrated at the user interface. These are connected through dependencies, but they are not the same market. Counting firms across all five categories and adding the totals would conceal the bottleneck rather than reveal it.

The BIS classification assigns each firm a primary role and, where clearly supported, a secondary role. That is necessary because some companies operate across layers. It also introduces judgment.