How to Segment the AI Value Chain: Four Layers Defined by Deliverables

    This article explains why slicing the AI industry by technical sophistication, visibility, or a coarse upstream-midstream-downstream model tends to break down, and instead proposes a stable segmentation based on what the customer actually receives. It divides the AI value chain into four segments — compute foundation, data and platforms, models and algorithms, and applications and agents — with a practical four-question decision tree for classifying any company, an analysis of how maturity differs across the four, and a look at where each segment is heading.

    1. Why Layering by “Technical Sophistication” Almost Always Goes Wrong

    Technology is fluid. A capability that sits in the application layer today may be absorbed into the model as a built-in feature six months from now, and a year later settle down into an instruction on a chip. If you use technical level as the basis for segmentation, the industry map has to be redrawn every year — and every redraw makes last year’s comparisons meaningless.

    The genuinely stable dimension is not “whose technology is stronger,” but “what is the deliverable.” To determine where a company sits in the value chain, the most reliable question is: what does it ultimately sell to its customer? The deliverable determines the business model, cost structure, customer profile, and gross margin room — and those stay stable over long time horizons. Technology flows; deliverables do not change day to day. That is the fundamental reason to anchor the segmentation on deliverables.

    2. Three Common Cuts, and How Each One Fails

    Before looking at the right approach, consider why the three most common alternatives fail.

    1. By technology stack: chips — frameworks — models — applications

    This chart looks very clean. The problem is the “framework” layer. Development frameworks mostly exist as open source and have no independent revenue structure of their own, so they barely constitute a real industry. Forcing them into the value chain turns the “representative players” of that segment into a collection of foundations and research labs. Once a layer contains a large number of non-commercial entities, every subsequent comparison based on revenue, gross margin, or market share becomes distorted.

    2. By visibility

    Putting the companies with the largest funding rounds and the most frequent launch events at the center of the map produces a popularity ranking, not a value chain. The fatal flaw of a popularity ranking is that the same name appears in several places at once — a company that builds compute, trains models, and ships applications is entirely normal. When one entity spans multiple segments, segmentation loses its discriminating power.

    3. By upstream / midstream / downstream: upstream hardware, midstream platforms, downstream applications

    This cut is not wrong, but it is too coarse. The so-called “midstream” may contain companies selling training corpora, companies selling compute rentals, and companies selling model inference capacity at the same time. These three businesses have completely different gross margin structures, customer profiles, and cost structures; putting them in the same box amounts to no segmentation at all.

    All three approaches share one consequence: the same company gets assigned to different segments in different reports, and the baseline for horizontal comparison collapses. Two companies nominally “doing compute” may occupy completely different positions in the value chain, with different investment logic, growth curves, and risk exposure. That kind of judgment cannot be assembled from news coverage and funding amounts.

    3. The Core Method: Four Segments Cut by “What the Customer Buys”

    Drop the technology dimension and switch to the deliverable dimension, and the AI value chain can be cut stably into four segments, from front to back:

    1. Compute foundation
    2. Data and platforms
    3. Models and algorithms
    4. Applications and agents

    The further forward a segment sits, the closer it is to hardware and infrastructure; the further back, the closer it is to concrete business scenarios.

    Translated into a practical test, you only need to ask a few questions in sequence:

    • Does the customer receive hardware? Accelerator cards, integrated servers, optical modules, data center facilities — if yes, classify it as compute foundation.
    • Is the customer buying the right to use a resource? Training corpora, data labeling services, compute rental, training and inference frameworks — if yes, classify it as data and platforms.
    • Is the customer buying the right to call a capability? Model APIs, model licensing, inference services — if yes, classify it as models and algorithms.
    • Is the customer buying a finished product that can go straight to work? Office software, industry solutions, complete robotic units — if yes, classify it as applications and agents.

    These four questions are essentially a decision tree that drills down level by level. The following illustrative pseudocode expresses the classification process:

    1. def classify_ai_segment(deliverable):
    2. """
    3. Classify an AI value chain segment based on what the customer buys.
    4. Rules are evaluated front to back and return on the first match,
    5. which avoids misclassification across segments.
    6. """
    7. rules = [
    8. (is_hardware, "Compute foundation"), # accelerator cards, servers, optical modules, data centers
    9. (is_resource_usage, "Data and platforms"), # corpora, labeling, compute rental, frameworks
    10. (is_capability_call, "Models and algorithms"), # APIs, model licensing, inference services
    11. (is_finished_product, "Applications and agents"), # office suites, industry solutions, robots
    12. ]
    13. for predicate, segment in rules:
    14. if predicate(deliverable):
    15. return segment
    16. return "Unclassified" # deliverable unclear; gather more information first

    Note the evaluation order: hardware first, then the right to use a resource, then the right to call a capability, and finished products last. The order itself carries the industry logic of moving from heavy to light, from the underlying layer to the scenario layer. Reversing it causes the same deliverable to be assigned to the wrong segment.

    4. How Maturity Differs Across the Four Segments

    Only after the classification is done does it make sense to talk about “progress.” Measuring stage with concrete indicators is far more accurate than describing it with adjectives like “mature” or “exploding.”

    Compute foundation: the most concentrated and heaviest of the four segments. It is characterized by enormous capital expenditure, a limited number of players, and a visible shift toward domestic chip substitution. In recent years, a batch of domestic AI chip vendors have moved into a ramp-up phase on revenue, with some posting year-over-year doubling or even multiples of growth and successively crossing the break-even point. This indicates that the compute foundation is transitioning from an “investment phase” to a “payoff phase,” but its capital-intensive nature brings a strong cyclicality — once downstream demand slows, capacity utilization transmits rapidly into revenue volatility.

    Data and platforms: this segment addresses how resources get used efficiently. It covers the collection and cleaning of training corpora, data labeling, elastic rental of compute resources, and the frameworks and scheduling platforms that support training and inference. Its customers include both model teams that want to build their own capabilities and application developers who only want to pay by usage, making it more like a utility-style middle layer whose gross margin is driven by both economies of scale and resource utilization.

    Models and algorithms: what this segment sells is the right to call a capability. Customers no longer care which chips or clusters run underneath; they care only whether the API is stable, whether latency is acceptable, and whether the unit price is competitive. The core of competition is cost-effectiveness and reliability, and pricing power will keep coming under pressure as open source models mature. Players therefore have to strike a balance between capability leadership and cost leadership.

    Applications and agents: the segment closest to money and closest to the scenario. Customers buy a finished product that can go straight to work — an office suite, an industry solution, a robot that can actually do the job. Success here does not depend on how strong the model is, but on the depth of understanding of the specific scenario, delivery and operations capability, and the ability to package unstable model capabilities into a stable, deliverable product experience.

    5. Where Each Segment Is Heading

    From the deliverable perspective, the evolution of the four segments can be summarized roughly as follows:

    • Compute foundation: the cost per unit of compute keeps falling, making energy efficiency a key metric; scheduling efficiency, cluster interconnect, and software ecosystem adaptation will matter more than peak compute alone.
    • Data and platforms: a shift from “selling resources” to “selling efficiency,” as training frameworks, data governance, and evaluation systems become progressively productized and platformized, helping customers train better models with fewer resources.
    • Models and algorithms: a shift from “selling per-call usage” to “selling a trusted bundle of capabilities,” as multimodal, long-context, reasoning, and tool-calling capabilities are integrated into a unified interface layer, with price competition and capability competition running in parallel.
    • Applications and agents: a shift from “assistive tools” to “autonomous execution,” as agents gradually take over complete multi-step, cross-system task flows, with industry know-how and engineering delivery capability becoming the core moat.

    One commonly overlooked pattern is that technology flows from back to front. Problems solved well in the application layer today may be abstracted into built-in model capabilities; scheduling strategies the model layer relies on today may sink into instruction-level support in chips tomorrow. Any segment’s “lead” is therefore temporary, and the only stable anchor remains the deliverable itself.

    6. Conclusion

    Back to the original question: how should the AI value chain be segmented? Not by technical sophistication, not by visibility, and not by a coarse upstream-midstream-downstream split, but by “what the customer buys” — into compute foundation, data and platforms, models and algorithms, and applications and agents. Deliverables are the stable anchor; technology is just the constantly shifting backdrop. Once you master this deliverable-based method, you have a ruler that will not expire every year — whether you are analyzing a company’s position, assessing the maturity of a track, or anticipating where a capability is heading.