How to Segment the AI Industry: A Four-Layer Framework Based on Customer Deliverables

    A practical way to analyze the AI industry is to classify businesses by what customers actually receive. This article introduces a four-layer framework—compute infrastructure, data and platforms, models and algorithms, and applications and agents—and explains how to distinguish them, avoid common classification errors, and interpret their different business models and operating metrics.

    Why Common AI Industry Classifications Often Fail

    The same company is often placed in different parts of the AI value chain across different industry reports. Some reports use categories such as chips, frameworks, models, and applications. Others divide the market into upstream, midstream, and downstream segments. Some simply rank companies by visibility or market attention.

    Each approach has limitations. More importantly, they share the same underlying problem: the classification framework is inconsistent, so the basis for comparison is lost.

    Classifying companies by technical sophistication may appear straightforward, but technology moves across layers. A problem solved at the application layer today may become a built-in model capability over time and eventually be embedded in a chip’s instruction set. Frameworks are also difficult to classify. Most are open-source projects without an independent revenue structure. Forcing them into the industry chain may result in foundations or research laboratories being treated as conventional commercial companies.

    Ranking companies by brand recognition creates a popularity leaderboard rather than a useful industry map. Companies with substantial funding and frequent product launches may appear simultaneously in hardware, models, and applications. This can create the impression that they dominate every layer, while concealing the differences in cost structures and competitive dynamics.

    The upstream–midstream–downstream framework is also too broad. The “midstream” category may include companies selling training data, rented compute, and model access. These businesses have very different gross margins, customer profiles, and cost structures. Placing them in the same category provides little analytical value.

    A Four-Layer Framework Based on Customer Deliverables

    A more stable approach starts with one question: What does the customer actually receive?

    Under this standard, the AI industry can be divided into four layers. The further forward a layer is positioned, the closer it is to hardware and infrastructure. The further back it is, the closer it is to specific use cases and end-user problems.

    1. Compute Infrastructure

    The customer receives a physical hardware product, such as accelerator cards, complete servers, optical modules, or data centers.

    The core of this business is manufacturing and systems integration. Gross margins are heavily affected by material costs and economies of scale. Customer decision cycles are often long, and once a supplier becomes part of a customer’s supply chain, replacement costs can be relatively high.

    2. Data and Platforms

    The customer purchases the right to use a resource. Typical offerings include datasets and data-labeling services, rented compute, and managed platforms for model training and deployment.

    The core capability in this layer is the organization and scheduling of resources. Many businesses charge by usage or subscription, which can make revenue relatively predictable.

    It is important to distinguish open-source training frameworks from the commercial services built around them. The framework itself typically does not generate direct revenue. Revenue is more likely to come from hosted services, operations and maintenance, technical support, or data services. These revenue sources should be analyzed separately.

    3. Models and Algorithms

    The customer purchases the right to access a capability. Common forms include APIs, model licenses, and inference services.

    This layer packages research and development results into standardized services. Each inference request still consumes compute, bandwidth, and operational labor. As a result, the cost per request may be low, but it is not zero.

    Gross margins depend on usage volume, model efficiency, and pricing strategy. Companies in this layer also require strong technical differentiation and engineering capabilities.

    4. Applications and Agents

    The customer purchases a finished product that can perform a task directly. Examples include productivity software, industry-specific solutions, and complete robots.

    The value of this layer comes from understanding the use case and delivering a working solution. Technology is the means rather than the product itself. Customers pay to have a problem solved; the underlying technology is simply how the solution is implemented.

    A complete robot may have a physical hardware form, but it still belongs in the applications layer when customers are paying for the finished functionality rather than the underlying components.

    Summary of the Four Layers

    Layer Customer deliverable Common pricing model Gross-margin characteristics Key metrics
    Compute infrastructure Physical hardware Unit sales or volume procurement Highly affected by material costs and scale Product shipment mix, delivery cycle, capacity utilization
    Data and platforms Resource access rights Usage-based pricing or subscription Depends on resource advantages and operating efficiency Resource scale, scheduling efficiency, customer retention
    Models and algorithms Capability access rights Usage fees or licensing Low marginal cost but high R&D investment Usage volume, unit inference cost, pricing changes
    Applications and agents Ready-to-use products Subscription, project-based fees, or complete-system sales Varies widely with use-case depth Renewal rate, project delivery cycle, implementation cost

    How to Classify a Business and Avoid Common Mistakes

    The classification process can follow three questions in order:

    1. Does the customer receive hardware?
      If yes, the business belongs to compute infrastructure.

    2. Does the customer purchase the right to use a resource?
      If yes, it belongs to data and platforms.

    3. Does the customer purchase the right to access a capability?
      If yes, it belongs to models and algorithms.

    If none of these apply and the customer is paying for a finished product that can perform work directly, the business belongs to applications and agents.

    Two issues frequently lead to incorrect classifications.

    First, a company may operate across several business lines. In that case, the business should be broken down by revenue source. Assigning a single label to the entire company can hide important differences between its operations.

    Second, the word “compute” can refer to two very different businesses. Selling accelerator cards is a hardware business, while selling rented compute is a resource-access business. Classifying companies according to their actual deliverables makes it easier to place each business in the correct layer.