Two competing buyers can rent the same computing infrastructure and still make different commercial decisions. Two buyers using different infrastructure can also arrive at the same choice, or coordinate it. The useful question is not simply what they share, but who controls the economic decision and what evidence supports that answer.
This is the fifth piece in a series. Earlier pieces cover bidder counts, auction transparency, independent assurance and agent identity.
The wrong shortcut
A modern advertising marketplace is assembled from shared components: cloud services, exchange protocols, measurement systems, data providers, software libraries and, increasingly, specialized optimization services. These dependencies can lower fixed costs, make systems interoperable and allow smaller companies to compete. The OECD's 2025 review of AI in downstream markets describes both potential entry and innovation benefits and new competitive risks; its assessment is explicitly context dependent.[1]
Calling a common technology supplier a common decision-maker skips several necessary questions. So does treating matching bids as evidence of an agreement. But the opposite shortcut is equally weak: different corporate owners, vendor contracts or bidder IDs do not guarantee distinct economic choices.
To investigate responsibly, separate four things that are often bundled together under the word "shared":
- Infrastructure: the equipment, hosting, protocol or software on which decisions run.
- Information: the market observations, forecasts and private customer data used as inputs.
- Decision service: the system that recommends, optimizes or executes an action.
- Control rights: who determines objectives, constraints, updates, exceptions and whether recommendations are followed.
These can overlap, but none is interchangeable with the others. A shared server need not set a common objective. A separate server does not prevent competitors from receiving the same nonpublic competitor-sensitive signal. A hosted service may execute tenant-specific policies; alternatively, a single provider may exercise substantial discretion for multiple customers. The last two possibilities require evidence about the service's actual role, not a guess based on its name.
Four hypothetical markets, four different conclusions
- Same infrastructure; separately chosen rules. Isolates: infrastructure. Two advertisers use the same cloud host and the same bidding interface. One values reach and has a strict frequency cap; the other values completed purchases and refuses certain inventory. Each sets and changes its own economic rules. The shared host creates operational dependence, including possible common outages, but does not establish shared commercial control.
- Same market signal; independently similar reactions. Isolates: information. Buyers using different vendors all lower bids after the same verifiable deterioration in inventory quality. Their responses are correlated because they face a common fact and may have similar incentives. Even synchronized decisions can be independently rational. This does not rule out coordination in a different case; it defeats the inference that similarity alone proves it. U.S. antitrust law distinguishes parallel conduct from an agreement under Sherman Act Section 1.[2]
- Same optimizer; authority divided or delegated. Isolates: decision service and control rights. This is the familiar demand-side platform (DSP) model, in which one bidding system serves many advertisers. If it implements separate customer objectives, isolates competitively sensitive inputs and leaves substantive pricing authority with each customer, common sourcing does not by itself show that the customers' decisions lack diversity. If it instead chooses material objectives or constraints centrally, pools competitor-sensitive data or makes decisions customers cannot meaningfully override, there may be a real shared-control or information-flow issue. Even then, a measurement of operational dependence is not automatically a finding of unlawful agreement.
- Technology used to implement an arrangement. Isolates: the arrangement itself, not any technical layer. Competitors allegedly agree to align prices or feed nonpublic strategic information through an intermediary. Where competitors are linked through a common intermediary rather than directly, enforcers and plaintiffs sometimes describe the structure as "hub-and-spoke." That label names a theory to be proved, not a property of any shared service. A common algorithm could implement or facilitate such an arrangement, but proving one requires evidence concerning communications, knowledge, conduct, market context and the relevant law. A shared vendor, a common price movement or an observed behavioral cluster alone is not that proof.
All four situations are illustrations, not accounts of actual ad exchanges, buyers or Exchange Optics experiments. Their purpose is to isolate the fact that changes the interpretation.
Where the real control question begins
A buyer's contract may say it retains "control." Operationally, that word can cover very different arrangements. The more useful inquiry concerns decision rights in practice.
Who specifies the economic objective: the advertiser, the DSP or a cross-customer optimizer? Who can change targeting, price ceilings, pacing, exclusions or reward functions? Can a recommendation be rejected without penalty or technical friction? Are rival customers' sensitive inputs isolated or used in a common optimization process? Who authorizes updates, and are changes synchronized across customers by design? Could a buyer switch providers or operate independently without losing the ability to make the relevant choice?
These are questions, not a single checklist that certifies competitive independence. Answers may differ by transaction, customer, time and service feature. A provider can control software deployment while customers retain economic discretion; a nominal customer override can exist on paper but rarely function in practice. One shared input can make decisions move together even when each buyer remains free to act differently.
The public literature supports caution in both directions. In controlled repeated-pricing simulations, Calvano and coauthors found reinforcement-learning algorithms could learn supracompetitive outcomes without communication. That is important scientific evidence about a possible mechanism, not proof of such behavior in any particular advertising exchange, and not a legal determination.[3] Conversely, the UK Competition and Markets Authority has documented both the efficiency benefits of algorithms and the challenges of detecting algorithmic competition harms.[4]
The lesson is more precise than "shared AI means collusion": economic convergence, shared operational control and an unlawful agreement are distinct hypotheses.
Competition law does not treat these as synonyms
In the United States, Section 1 of the Sherman Act addresses concerted conduct. In Bell Atlantic Corp. v. Twombly (2007), the Supreme Court held that allegations of parallel conduct, without more, do not plausibly suggest an agreement, because such conduct is equally consistent with lawful, independent reactions to the same commercial incentives.[2] That is a pleading standard grounded in longer-standing doctrine on parallelism; it is not a statement that parallel outcomes can never justify investigation.
U.S. agencies have also emphasized the other side of the boundary. In a 2024 statement of interest in private hotel-pricing litigation, the Justice Department and Federal Trade Commission argued that using an algorithmic intermediary does not shelter an otherwise unlawful price-fixing agreement and that retaining some discretion over a recommendation is not necessarily a defense. That was the agencies' litigating position, not a court judgment resolving that case.[5]
European Union guidance frames part of the issue through indirect exchanges of commercially sensitive information. The European Commission's 2023 Horizontal Cooperation Guidelines (§6.2.4.2, paragraphs 401–403) address exchanges through platforms, third parties and shared optimization algorithms; they distinguish publicly available data from sensitive competitor information and identify circumstances in which indirect exchange may infringe Article 101. These are EU guidelines, not a universal legal test for U.S. markets.[6]
An example outside advertising shows why the distinctions matter. In August 2024 the U.S. Justice Department filed a civil antitrust case alleging that RealPage and landlords used nonpublic competitor information and algorithmic pricing in an anticompetitive arrangement. The Justice Department announced a proposed RealPage consent decree on November 24, 2025, and the U.S. District Court for the Middle District of North Carolina entered a final judgment against RealPage on May 19, 2026. That resolved the federal government's claims against RealPage, but not the whole case: certain state claims and the claims against other defendants continued on separate tracks. A complaint is an allegation, a proposed settlement is not an adjudication, and a consent judgment is not a finding against every defendant. The case concerns rental housing and is not evidence of misconduct by any advertising company.[7]
What advertising standards can, and cannot, tell us
The industry's transparency standards answer real questions. IAB Tech Lab's sellers.json identifies direct sellers and intermediaries; the OpenRTB SupplyChain object describes parties involved in selling or reselling a bid request. Those are supply-path representations, not an inventory of who set a buyer's optimization objective or whether rival buyers exchanged information.[8] [9]
OpenRTB 2.6 defines bid requests, bids, seats, responses and no-bid conventions. A compliant message can describe a transaction without recording the business rules behind it, all eligible participants not invited to respond, or the provenance of a recommendation. Data retained by one operator may be much richer, or poorer, than the wire protocol describes. The standard is neither a universal event ledger nor a decision-independence certification.[10]
IAB Tech Lab's Agentic Real Time Framework v1 describes interoperable agent services operating under a hosting/orchestration model, with declared interfaces for bidstream operations. Its ability to support delegated execution is useful architectural context. It is not, by itself, evidence that two hosted agents make either independent or jointly controlled economic choices.[11]
The difference matters because a market may be transparent about sellers and still have incomplete information about buyers' decision rights. It may also be transparent about a shared vendor without having evidence that the vendor coordinates anyone.
What a responsible investigation would ask
A marketplace operator or customer need not begin by alleging collusion. It can ask three narrower questions:
What is actually shared? Distinguish hosting, code, public inputs, confidential customer inputs, model updates, recommendation services and binding execution authority. Obtain contractual descriptions and technically supportable confirmations only with proper permissions; do not presume that a marketing label establishes the answer.
Who can make a different choice? Understand where independent objectives, constraints and override powers are set and exercised. A formal ability to override is not the same as evidence that independent decisions occur; conversely, a common software release does not prove a common commercial policy.
What can the observations distinguish? Bid records may show what responding systems submitted. They generally do not reveal unexpressed choices by uninvited participants, all internal decision rules, data provenance, communications or the legal elements of an agreement. Common market shocks, selection and changes in opportunity mix remain plausible alternatives. Additional lawful records or controlled review may narrow uncertainty; sometimes the appropriate result is unresolved.
Nor is every shared decision service economically harmful. Centralization may improve efficiency, reduce operational failures and enable entry. A common model can simultaneously create resilience concerns, such as correlated errors or synchronized disruptions, without any illegality. Those potential effects must be assessed separately from actual price, allocation or welfare effects. A behavioral signal is not a loss estimate.
Independence is not a vendor count
The useful operating principle is neither to condemn common infrastructure nor to take comfort in a list of independent company names. It is to identify which decisions remain under independent economic control, which dependencies could make them converge, and which questions the evidence cannot yet answer.
For exchanges, publishers and buyers, this creates a practical distinction: supplier due diligence asks what a service is and who operates it; market-integrity inquiry asks how that service changes decisions and outcomes across participants. Competition-law analysis, when warranted, asks a further question about conduct and legal elements. These inquiries can inform one another. None substitutes for the others.
Independent market verification examines who retains economic decision authority, how shared dependencies shape observed behavior, and how much effective decision diversity the evidence supports. Its output is not a verdict on competition law or a certificate of independence. It is a bounded statement: for a defined set of records and a defined question, what the evidence supports, what it cannot distinguish, and what additional information would narrow the gap. Neither common infrastructure nor separate company names can settle that question on their own.
Sources
[1] OECD (14 November 2025), Artificial intelligence and competitive dynamics in downstream markets (https://www.oecd.org/en/publications/artificial-intelligence-and-competitive-dynamics-in-downstream-markets_ccf0624a-en.html) (policy analysis).
[2] U.S. Supreme Court (21 May 2007), Bell Atlantic Corp. v. Twombly, 550 U.S. 544 (https://supreme.justia.com/cases/federal/us/550/544/) (judicial opinion reproduced by Justia; Sherman Act §1, parallel conduct and pleading).
[3] Calvano, Calzolari, Denicolò & Pastorello (October 2020), "Artificial Intelligence, Algorithmic Pricing, and Collusion," American Economic Review 110(10) (https://pubs.aeaweb.org/doi/10.1257/aer.20190623) (controlled model/simulation study).
[4] UK Competition and Markets Authority (19 January 2021), Algorithms: How they can reduce competition and harm consumers (https://www.gov.uk/government/publications/algorithms-how-they-can-reduce-competition-and-harm-consumers/algorithms-how-they-can-reduce-competition-and-harm-consumers) (regulatory research).
[5] U.S. Federal Trade Commission (28 March 2024), FTC/DOJ statement of interest in hotel-room algorithmic price-fixing litigation (https://www.ftc.gov/news-events/news/press-releases/2024/03/ftc-doj-file-statement-interest-hotel-room-algorithmic-price-fixing-case) (agency litigation position, not a merits judgment).
[6] European Commission (2023), Guidelines on horizontal cooperation agreements, §6.2.4.2, paragraphs 401–403, OJ C 259/85 (https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52023XC0721%2801%29) (EU Article 101 guidance on information exchange, including indirect exchange).
[7] U.S. Department of Justice, U.S. and Plaintiff States v. RealPage, Inc. case docket (https://www.justice.gov/atr/case/us-and-plaintiff-states-v-realpage-inc); DOJ (24 November 2025), announcement of proposed RealPage consent decree (https://www.justice.gov/opa/pr/justice-department-requires-realpage-end-sharing-competitively-sensitive-information-and); and final judgment against RealPage, Doc. 194 (M.D.N.C., 19 May 2026) (https://business.cch.com/ald/USvRealpage-wn-194052126.pdf) (pleadings, settlement proposals, party-specific judgments; do not equate them).
[8] IAB Tech Lab (specifications adopted 31 July 2019; webpage updated 7 October 2026), sellers.json guidance (https://iabtechlab.com/sellers-json/) (seller/intermediary identities).
[9] IAB Tech Lab (living specification, accessed 25 October 2026), OpenRTB SupplyChain object (https://github.com/InteractiveAdvertisingBureau/openrtb/blob/main/supplychainobject.md) (supply-side transaction chain).
[10] IAB Tech Lab (OpenRTB 2.6 specification, accessed 25 October 2026), OpenRTB 2.6 (https://github.com/InteractiveAdvertisingBureau/openrtb2.x/blob/main/2.6.md) (request/response schema).
[11] IAB Tech Lab (ARTF v1 released 12 November 2025; renamed March 2026; page updated 12 August 2026), Agentic Real Time Framework (https://iabtechlab.com/standards/artf/) (hosted-service interface; not a certification of economic independence).
Scope note: Public research and conceptual analysis, not legal advice or a factual accusation about any named advertising participant. Hypotheticals are expressly illustrative. No Exchange Optics customer evidence, proprietary methodology or unreleased experiment is reported.