The market structure of the AI value chain (compute → foundation models → inference API/MaaS → orchestration/routing/agents → distribution), assessed against the adtech sector as the reference case. Source dossier: Zwei Inversionen und eine Leerstelle.
Method note. The comparison is property-by-property, not layer-by-layer. Mapping layers onto each other (“compute is like publishers, routers are like exchanges”) always finds a fit somewhere and yields nothing. The eleven properties are compared with one of three verdicts each: same mechanism, analogous but different magnitude, structurally absent.
The three headline findings
1. The stickiness inversion
In adtech the rule-setting layer (the ad server) was also the stickiest layer — which is what allowed it to hold a rent. In the AI stack the two are on different layers:
- Compute is sticky but sets no rules over third parties. The FTC documents
that OpenAI had to “port to run on Microsoft Azure,” with code changes showing
the migration still under way “a total of 17 months later”
(
sources/ai-market/2026-08-22-ftc-6b-ai-partnerships-report-2025.pdf, p. 34). Add proprietary toolchains (CUDA, AWS Neuron SDK), exclusivity, spend commitments, egress fees. The Autorité documents a harder case: AutoML services where “users do not have access to the model itself after the final model is created” and a switch requires rebuilding “from scratch” (sources/ai-market/2026-08-22-autorite-concurrence-generative-ai-opinion-2024.pdf, 24-A-05 ¶ 251). - Routing sets selection rules but is not sticky. A switch is a changed base URL; LiteLLM measures a 99.3 % warm-cache hit rate across 4,684 provider switch-backs (LiteLLM, 2026-07-31).
⟨Inference: the adtech pathology — a rule-setter that can also hold you — has no single home in the AI stack. This is the strongest falsifying finding against transferability, and it is load-bearing for everything else on this page.⟩
2. The opacity inversion
Adtech: the buyer knew what it bought (an impression) but not what it paid (the ISBA/PwC study could not trace 15 % of spend). AI stack: the buyer knows what it pays (posted per-token list price) but not what it got — model composition changes behind a stable name, with automatic version updates as the default (Azure AI Foundry model router), and 27 private Llama-4 variants were tested pre-release (arXiv 2504.20879).
⟨Inference: this upgrades Transfer 2 from adtech-as-regulatory-precedent — opacity as change frequency behind a stable interface — from analogy to literal documentation. But it also weakens the “design lever” reading: evidence that appears in the vendor’s own documentation is disclosure, not concealment.⟩
3. The missing auction
The AI stack has no auction anywhere. Clearing happens via posted list prices, subscriptions, multi-year capacity contracts (CoreWeave discloses an unsold-capacity obligation running to 2032-04-13 with a $6.3 bn initial value, 8-K), and rate limits — i.e. an allocation quota rather than a price.
⟨Inference: roughly half the transferable adtech findings are auction findings (bid shading, reserve prices, last look, header bidding, first vs. second price) and therefore have no target here. Anyone citing adtech as precedent must subtract that half explicitly.⟩
Forming exception, again at compute rather than routing: CME Group and Silicon Data announced H100/B200 rental-index compute futures for 2026-10-05 on NYMEX, “pending regulatory review” (CME, 2026-08-11). ⚠ ⟨As of the 2026-11-02 lint the announced launch date is four weeks past and the wiki holds no source on what happened — whether the product launched on 5 October, was delayed, or is still in regulatory review. Treat the “forming exception” above as a reading current to the August 2026 announcement, not to today. This is the single check that would most change the page: a launched contract is a price-formation mechanism, an abandoned one is not.⟩
Router taxonomy
No agency document contains one; this is our own. Only type (c) has self-preferencing structurally built in.
| Type | Example | Rent | Adtech analogue |
|---|---|---|---|
| (a) buyer-controlled proxy | self-hosted LiteLLM | none | the buyer’s trading desk |
| (b) neutral aggregator | OpenRouter, Not Diamond, Requesty | fee or spread | the exchange |
| (c) integrated first-party router | Azure Model Router, Bedrock, Vertex | bundled | the ad server with last look |
| (d) hosting/CDN gateway | Cloudflare, Vercel AI Gateway | fee or none | — |
Where the chokepoint sits: four agencies, four answers
In adtech the authorities converged on the ad server. Here they do not:
- Bundeskartellamt — the vertically integrated stack owner. Microsoft is
“the only provider active on all vertically integrated levels of the software
stack for businesses,” “not only sets the framework conditions for third-party
developers but also competes with them,” and has “considerable power to set the
rules”
(
sources/ai-market/2026-08-22-bundeskartellamt-microsoft-19a-case-summary-2024.pdf, B6-26/23, 2024-09-27). - Autorité de la concurrence — the MaaS layer; recommends the Commission consider designating MaaS providers as DMA gatekeepers specifically for those services (24-A-05).
- CMA — distribution and ecosystem controllers, who could “set the rules of the game for FM deployment to restrict customer choice” (Technical Update 2024, p. 56).
- FTC — not a layer but the contractual structure of the partnerships themselves (6(b) 2025).
⟨Inference: a chokepoint whose location is contested is probably not yet a chokepoint. Counter-reading that cannot be ruled out: adtech convergence also took years, so this may be immaturity of investigation rather than of the market.⟩ Nobody names the routing layer — the one layer whose stated purpose is setting selection rules.
Rent: real, an order of magnitude smaller, differently shaped
Adtech’s chain took 30–50 %. Routing-layer rents: OpenRouter 5.5 %/5 % with inference passed through “without any markup” (FAQ); Cloudflare 5 % on credit purchase; Vercel zero markup; Bedrock $1 per 1,000 routing requests; Not Diamond $0.05/M routed tokens; Requesty +5 %; Portkey/Unify/Helicone pure subscriptions.
⟨Inference: the rent is shaped as a payment fee, not a spread. A spread requires an information advantage over the execution price; a fee only requires the payment to pass through. Stripe’s acquisition of OpenRouter fits that reading — the buyer is a payments company. This supersedes the flat “no measurable intermediation rent” claim in Dossier 1 and the subsequent full retraction of it; both were wrong.⟩
Network effects: structurally absent
“The role of network effects is significantly smaller than for platforms — a
given ChatGPT user does not derive significant direct benefits from others
joining ChatGPT” (Korinek & Vipra 2025,
sources/ai-market/2026-08-22-korinek-concentrating-intelligence-2025.pdf,
p. 19). Adtech is a two-sided market where the network effect is the mechanism.
Substitutes named by the same authors: scale and scope economies, data feedback,
user inertia, and an AI-specific “intelligence feedback loop.” On the
monopoly tendency they are explicitly undecided: “the number of players that a
market of a given size can support is shrinking fast,” but “it is unclear whether
investment requirements or expectations of future profits will rise faster.”
⟨Inference: regulation aimed at network effects — which is the whole DMA architecture — targets the wrong mechanism here. What concentrates is capital, not linkage.⟩
Entry barrier: capital and geography, and measurable
Adtech’s barrier was information and distribution, unquantifiable from outside. The AI barrier is capital and publicly priced: xAI’s Colossus, the most performant AI supercomputer as of March 2025, at an estimated $7 bn hardware cost; leading-system hardware cost doubling at 1.9×/year; extrapolated to ~$200 bn by 2030 plus ~$90 bn of data-centre construction at 9 GW (Epoch AI 2025, pp. 15, 19). Plus a jurisdictional barrier adtech did not have in this form: ~75 % of aggregate AI supercomputer performance is in the US, private owners are ~85 % in 2025, and the largest public system reaches only 22 % of the largest known industry system (ibid., pp. 10, 16, 20).
⟨Inference: this undercuts the Autorité’s recommendation to open public supercomputers as an alternative — that layer’s share has already collapsed to roughly a sixth. And a capital barrier is externally checkable where an information barrier is not, which is the one respect in which the AI stack is easier to regulate than adtech.⟩
Measurability: the one row where AI is behind adtech
Adtech at least had attempts — DSA Art. 40 as a data-access right, the ISBA/PwC
study. The joint statement of EC, CMA, DOJ and FTC
(sources/ai-market/2026-08-22-comp-cma-doj-ftc-joint-statement-genai-2024.pdf,
2024-07-23) names three principles (fair dealing, interoperability, choice) and
zero measurement instruments. No
router publishes an auditable routing policy. And there is no documented
accusation, complaint, audit, agency investigation or benchmark alleging that any
router preferences its own or an affiliated model — absence of evidence, not
evidence of absence, since without the instrument the claim could not be raised
even if true.
Two side-findings showing there would be something to measure: Bedrock’s routing operates only within a chosen model family (AWS), and Microsoft’s own docs disclose that “if you use Agent service tools in your flows, only OpenAI models are used for routing.”
See binary-verifiable-remedies for the statutory self-preferencing reporting duty that does exist — at compute.
Same mechanism, better documented than in adtech: parity and MFNs
A cloud provider to the FTC: “We’ve sought parity in multiple ways… Parity on features and functionality… Parity on pricing… Parity on compensation for their sellers… Parity on marketing” (6(b) 2025, p. 22). And, harder: “In at least one case, a partnership effectively required the AI developer partner to keep the models on its partner’s cloud service at parity with the models on the AI developer’s first-party API… the AI developer would not be able to release a new model on its own before launching it on its partner’s cloud service” (ibid.) — an MFN operating against the developer’s own first-party channel.
Self-preferencing subsidy, with figures (Autorité 24-A-05 ¶ 249): Google for Startups Cloud Program up to $350,000 over two years for AI startups, $150,000 more than for other sectors; Microsoft up to $150,000; AWS Generative AI Accelerator up to $300,000. The decisive clause: “The amount of credits can also be increased on condition that the beneficiary start-ups use the suppliers’ new AI products” — AWS Activate $100,000 plus up to $300,000 more for using Trainium and Inferentia.
⟨Inference: structurally identical to the adtech pattern of discounts for buying through the vendor’s own stack. This is the one row where the transfer holds without qualification.⟩
Circular spending is the FTC’s term
“This feature of the partnerships—circular spending—is one avenue through which CSP partners may have potentially aimed to reduce the magnitude of potential loss” (6(b) 2025, p. 19). Instances: Nvidia’s $2 bn into CoreWeave at $87.20/share (Nvidia, 2026-01-26) and $2 bn into Nebius at $94.94/share for ~8.3 % (Nvidia, 2026-03-11). Adtech had discounts and bundles but no equity financing of its own customer — here the AI stack is escalation, not repetition.
Contradictions between the two literatures
Neither cites the other in any of these three cases:
- Data siloing. Korinek & Vipra recommend “that foundation model companies and cloud companies hold certain data in silos… as mandated, e.g., for gatekeepers under the EU’s Digital Markets Act” (p. 20). Witte & Krämer 2025 identify data siloing as unverifiable because ML pipelines “reuse partial data or inferred signals.” The leading AI-economics paper recommends precisely the instrument the adtech remedy literature discarded as unenforceable.
- Remedy vocabulary. Narechania reaches for natural-monopoly regulation — rate regulation and service regulation (Iowa L. Rev. 2021); that instrument class appears nowhere in the adtech remedy literature. Conversely, binary verifiability appears nowhere in the AI literature.
- Open-source exemptions. Widder, Whittaker & West 2023 argue the AI Act’s open-source exemptions “could serve to further concentrate, rather than challenge, Big Tech power.” ⟨Inference: extends Campbell, Goldfarb & Tucker 2014 by one step — not only do fixed compliance costs concentrate, so can the exemption from them.⟩
Falsifiable prediction sequence
If the adtech analogy holds despite the stickiness inversion, observable in this order: (1) routers move from flat/payment fee to spread or arbitrage; (2) integrated first-party routers default to their own models; (3) opacity emerges about which provider served a request; (4) model providers respond with direct deals and exclusivity to bypass routers.
⟨Inference: if step 1 does not occur, the analogy is refuted rather than merely delayed — without a spread there is no rent that makes distorting the selection rules worthwhile.⟩
Open
- The whole page depends on routing having near-zero switching costs. If stickiness forms there — accumulated observability data, non-portable agent state, embedded prompt/policy config, billing integration — the stickiness inversion falls. The Stripe acquisition points that way.
- No independent measurement of any router’s routing decisions exists.
- Martian’s pricing is unverifiable (vendor pages 404). Nvidia’s 10-K competition language was unreachable (SEC EDGAR 403).
- The evaluation layer is treated here as demand-side countervailing power (adtech had no continuous independent publication of take rates or execution quality), but it carries its own interest structure: LMArena sells paid evaluations to the labs it ranks; Epoch AI discloses a lab funding relationship; METR states it takes no lab funding.