The AI Value Chain

A model call looks like a software service, but it is the visible end of an industrial chain. Depending on the deployment, an accepted outcome may require chip fabrication, memory and packaging, deliverable power, accelerator capacity, a model artifact or hosted service, a gateway, an application, data rights, and a working organizational process. Those links form a dependency graph, not one market. A local model can bypass a hosted API, for example, while still depending on hardware, power, software, and licensed data.

Chapter 76 compared the cost of feasible alternatives from one buyer's perspective. This chapter asks what shapes the terms offered to that buyer. The answer is specific to a workload: which dependency is scarce or difficult to replace, who controls it, and which outside options can deliver the same accepted outcome. Market structure matters because bargaining power grows when credible alternatives disappear, not simply because a company appears at several layers.

Map dependencies before naming a market

A value-chain diagram describes technical and contractual dependence. It does not establish that every box is a market, that every edge is a purchase, or that one company owns the whole path.

The OECD's 2025 review describes AI compute infrastructure as a multilayer supply chain with concentrated segments, high entry barriers, vertical relationships, partnerships, switching barriers, and occasional shortages. It also warns that its reported supplier shares are bottleneck indicators, not the result of one global market-definition exercise (Organisation for Economic Co-operation and Development 2025). The UK Competition and Markets Authority (CMA) similarly identifies compute, data, technical expertise, funding, and routes to market as possible constraints on entry or expansion in foundation-model services (Competition and Markets Authority 2024). These are risks and dependencies to investigate, not proof that one company has power everywhere.

A bottleneck is a dependency whose available capacity or terms constrain the workload. A control point is narrower: a supplier controls a scarce or hard-to-replace input, has the ability and incentive to change access or terms, and faces too few effective outside options for the buyer to respond. A long lead time can reveal a bottleneck. It does not by itself prove market power. Neither does technical superiority, a large company, or a high market share calculated over an undefined denominator.

The same supplier can appear in several rows, and two rows can be complements rather than substitutes. That is why the market boundary must come next.

Define the market before calculating shares

Concentration measurement has a long history. Albert Hirschman in 1945 and Orris Herfindahl in 1950 independently developed closely related share-squared indexes; Rhoades later summarized their origin and regulatory use (Rhoades 1993). The arithmetic is easy. The difficult part is deciding which sales and suppliers belong in the denominator.

Record a candidate market as

$$\mathcal{M} = (P, G, T, U, Q).$$

Here, $P$ means the product or service being tested, $G$ its geography, $T$ the time window, $U$ the customer group, and $Q$ the minimum quality and service threshold. The tuple is an analytical record, not a claim that only one market definition is valid. A candidate belongs in $\mathcal{M}$ only if customers in $U$ would treat it as a reasonable substitute for $P$ under the declared geography, period, and threshold.

For example, a downloadable model and a managed frontier API are both ways to obtain model output. They are not necessarily substitutes for a hospital that requires a particular accuracy, region, audit trail, support contract, and response-time objective. Conversely, two APIs with different model names may be close substitutes if both pass the same evaluation and operational contract. Define the decision first; do not let a vendor category define it for you.

A share also needs a unit. Revenue can reflect what customers pay. Delivered capacity can better describe a homogeneous input whose competitive significance comes from the ability to expand. Usage can describe an operating service, while new-customer wins can reveal current competition better than a large installed base. The 2023 U.S. Merger Guidelines explicitly permit different share metrics when they better match market reality (U.S. Department of Justice and Federal Trade Commission 2023). For an AI service, report at least two defensible proxies when the data allow, such as revenue and accepted workload volume, and explain omissions.

Given a defined market, the Herfindahl-Hirschman Index (HHI) is

$$\operatorname{HHI}(\mathcal{M}) = 10{,}000 \sum_{i=1}^{n} s_i^2, \qquad \sum_{i=1}^{n} s_i = 1.$$

Here, $n$ is the number of suppliers in $\mathcal{M}$ and $s_i$ is the share of supplier $i$, expressed from zero to one; all supplier shares sum to one. The factor $10{,}000$ puts the result on the familiar zero-to-10,000 scale. Squaring gives larger suppliers more weight. The equal-sized equivalent count $N_{\mathrm{eff}}=10{,}000/\operatorname{HHI}$ answers a limited intuitive question: how many equal suppliers would produce the same index?

The HHI is a screening statistic, not a finding of market power. Under the 2023 U.S. Merger Guidelines, a post-merger HHI above 1,800 combined with an increase above 100 points creates a structural presumption in that merger framework (U.S. Department of Justice and Federal Trade Commission 2023). Those thresholds are not a universal grading scale for technology markets. A result changes when the market boundary, share metric, missing suppliers, or time window changes.

The following hypothetical shares demonstrate that sensitivity. They are not estimates of any AI market.

markets = [
    ("narrow", [45, 30, 15, 10]),
    ("broader", [30, 25, 20, 15, 10]),
    ("equal-five", [20, 20, 20, 20, 20]),
]

for label, shares in markets:
    assert sum(shares) == 100
    hhi = sum(share * share for share in shares)
    effective_firms = 10_000 / hhi
    print(
        f"{label}: suppliers={len(shares)} hhi={hhi} "
        f"effective-firms={effective_firms:.2f}"
    )

A supplier count alone would call all three cases similar. The HHI distinguishes their size distributions, but it still cannot tell whether entry is timely, capacity is available, customers can switch, or providers compete on dimensions that the selected share misses.

Separate scale economies, concentration, and power

Large fixed cost can lower average cost without lowering marginal cost. In the simple cost function

$$C(q)=F+cq, \qquad AC(q)=\frac{F}{q}+c, \qquad MC(q)=c,$$

$q$ denotes output, $C(q)$ total cost, $F$ fixed cost, and $c$ constant cost for one additional unit. $AC(q)$ is average cost and $MC(q)$ is marginal cost. As output rises, the fixed term $F/q$ falls, but marginal cost remains $c$. The legacy shortcut that combined high fixed cost, falling marginal cost, and switching cost therefore mixed three distinct mechanisms.

High concentration can arise from innovation, capacity constraints, economies of scale, regulation, or exclusionary conduct. It can coexist with aggressive competition for the next contract, or with weak competition inside an installed base. Evidence of power comes from behavior and alternatives: persistent price or term changes that customers cannot avoid, denied or delayed access to an input, poor expansion by rivals despite demand, contractual restrictions, or failed switching attempts.

A worked example shows why scope matters. In its July 2025 UK cloud investigation, the CMA estimated 2024 infrastructure-as-a-service revenue shares of 40-50 percent for AWS, 30-40 percent for Microsoft, and 5-10 percent for Google. Platform-as-a-service shares were less concentrated. The same appendix says accelerated-compute revenue could not be separated consistently and was a small part of the measured total (Competition and Markets Authority 2025). This is evidence about specified UK cloud services, not a direct estimate of a global AI-compute or model market.

Do not infer profit from a price chart

Listed price is not production cost. Revenue is not gross profit, gross profit is not economic rent, and capital expenditure is not a layer's share of customer value. A valid margin comparison needs an audited segment or a reproducible allocation rule for shared hardware, labor, research, energy, and overhead. A return above the cost of capital can motivate further analysis, but it still does not identify which mechanism caused the return.

That boundary removes the legacy profit-pool table. Its source was one investor's unaudited estimate built from heterogeneous companies and layers. It could not show how one token's payment was divided, and a two-point trend could not support a decade-long forecast. Where segment reporting is unavailable, say that profit allocation is unknown and measure the contract terms the buyer can actually see.

Measure switching as a directed path

A buyer's outside option is credible only if it can be exercised. API syntax compatibility is one part of portability, not the whole migration. Record at least five layers:

1. Interface portability: request, response, tool, identity, and error schemas can be translated.
2. State and data portability: prompts, retrieval indexes, tool definitions, fine-tuning data, evaluation sets, logs, and user state can be exported under usable rights.
3. Behavioral portability: the replacement passes the same quality, policy, and service evaluation.
4. Artifact and runtime portability: weights, tokenizer, configuration, and inference code can run on another supported stack.
5. Operational and commercial portability: regions, quota, observability, support, price, and contract terms remain feasible.

Switching cost is directional. Moving from supplier $a$ to supplier $b$ can cost more than moving back because formats, rights, team knowledge, and capacity differ. A useful ledger is

$$\begin{aligned} C_{\mathrm{switch}}(a \rightarrow b) &= C_{\mathrm{export}} + C_{\mathrm{rewrite}} + C_{\mathrm{retest}} \\ &\quad + C_{\mathrm{parallel}} + C_{\mathrm{exit}} + \mathbb{E}[C_{\mathrm{failure}}]. \end{aligned}$$

Here, $C_{\mathrm{export}}$ covers state transfer and data egress; $C_{\mathrm{rewrite}}$ covers interface and workflow changes; $C_{\mathrm{retest}}$ covers evaluation, security, and compliance work; $C_{\mathrm{parallel}}$ covers running both paths during migration; and $C_{\mathrm{exit}}$ covers termination and stranded commitments. $\mathbb{E}[C_{\mathrm{failure}}]$ is expected loss from migration failures. The arrow records direction.

Estimate those terms with a migration drill. Export representative state, route a held-out workload to the replacement, measure quality and tail latency, remove the old path, record time to restore service, and test rollback. A fallback that has never passed this exercise is an aspiration, not an outside option.

Read vertical integration through rights and incentives

Vertical integration can lower coordination cost, secure investment, improve hardware-software co-design, and make capacity planning more reliable. It can also create preferential access, raise rivals' input costs, restrict multihoming, or enable foreclosure from a route to market. The label alone does not determine the effect, and it does not make every integration anticompetitive.

The Federal Trade Commission's January 2025 staff report examined specific relationships among the cloud providers Alphabet, Amazon, and Microsoft and the developers Anthropic and OpenAI. It documented varying equity and revenue rights, consultation or control provisions, cloud-spend commitments, discounted compute, exclusivity terms, product integration, and access to sensitive technical and business information (Federal Trade Commission, Office of Technology Staff 2025). Staff identified possible effects on input access and switching. The report did not define every AI market or make an illegality finding, and its evidence should not be generalized beyond the studied arrangements.

For procurement, map the rights that change behavior:

• committed spend and the share of expected demand it covers;
• capacity priority, reservation, and shortage allocation;
• exclusivity, preferred-provider, and multicloud terms;
• ownership and permitted use of prompts, outputs, telemetry, and derived data;
• model, product, and intellectual-property licenses;
• governance, information, audit, price-change, termination, and exit rights.

Usage does not automatically become training data or a competitive flywheel. Separate product telemetry, evaluation data, training data, distribution feedback, and revenue. For each loop, record what information moves, under which right, how quickly it changes the product, and whether a rival can obtain an equivalent input.

Openness, transparency, and contestability are different

Available weights are not the same as Open Source AI. The Open Source Initiative's 2024 definition requires freedoms to use, study, modify, and share the system, together with data information, code, and parameters in the preferred form for modification (Open Source Initiative 2024). It is a community standard rather than a universal legal definition, but it makes the missing materials explicit.

Broadly available weights can enable inspection, adaptation, and self-hosting. They can improve a buyer's outside options when the license permits the use, the artifact is complete, a supported runtime exists, the candidate passes the evaluation contract, and sufficient capacity is obtainable. If any one of those conditions fails, weight access does not create a production substitute. Kapoor and colleagues present competition benefits of open-weight releases as plausible but qualified, alongside distinct benefits and risks (Kapoor et al. 2024).

Transparency is another axis. The 2025 Foundation Model Transparency Index reported a mean score of 40.69 out of 100, about 41, compared with 58 in 2024; training data, training compute, and post-deployment usage and impact were among the weakest areas (Wan et al. 2025). The indicator set changed substantially in 2025, the sampled companies changed, and only nine companies appeared in both editions. The headline decline is therefore not a clean like-for-like time series. The index measures public disclosure, not model quality, safety, market concentration, or switching cost.

Transparency can improve comparison, risk assessment, and accountability. It does not by itself create a substitute. Contestability requires challengers to obtain critical inputs, enter and expand, reach customers, and let customers switch or multihome without prohibitive technical or contractual friction.

Match each claim to evidence

Evidence should have a date. Supplier shares, model quality, licensing terms, capacity, and partnerships change at different speeds. Preserve the quote, contract, measurement method, and market boundary used for each conclusion.

Operate a market record

Treat the analysis as a dated control process:

1. Define the decision. State the outcome being procured and why a market question affects it.
2. Freeze the market boundary. Record product or service, geography, time window, customer group, and quality and service threshold.
3. Enumerate suppliers. Include current suppliers, committed entrants, self-supply where relevant, and the reason each candidate is or is not a reasonable substitute.
4. Measure more than one share proxy. Reconcile revenue, capacity, usage, accepted workload, or new-customer measures and disclose missing data.
5. Test matched alternatives. Run the same evaluation and service contract against candidates rather than counting names or endpoints.
6. Map dependencies and rights. Record input suppliers, ownership, spend commitments, information flows, capacity rights, price changes, and termination terms.
7. Exercise migration. Perform the directed switch, measure cash and time to restore service, and verify rollback.
8. Set review triggers. Reopen the analysis after a material price, model, license, capacity, partnership, regulation, or workload change.
9. Refresh the record. Replace forecasts with observed shares, incidents, migrations, and contract outcomes; retain the old record for comparison.

The record should contain the workload and acceptance-rule digest, candidate list, exclusion reasons, share data and denominator, HHI sensitivity, contract versions, dependency owners, migration evidence, unresolved assumptions, and review date. That package lets another team reproduce the conclusion instead of inheriting a label such as "concentrated" or "open."

Chapter 78 now moves from supplier choice to realized work. A market can offer several credible alternatives while none creates user value, or one supplier can hold a strong position while a narrow workflow still produces positive returns. The next chapter measures that organizational boundary. Bargaining power lives where credible alternatives end.

Further reading