Who Controls the Models?

The most powerful artificial intelligence systems arrive through a deceptively simple interface. A person opens an application, types a request, and receives an answer. The model appears to be the product.

Behind that answer is a chain of organizations that financed the system, fabricated the chips, supplied the cloud, trained the model, licensed the intellectual property, distributed access, set the defaults, enforced the usage rules, and decided whether the customer could leave without rebuilding everything.

The visible model may be the least useful place to begin asking who has power. Control is a stack, not a logo.

The Model Is Only the Visible Layer

Chapter Nine followed access down to electricity and compute. Its point was physical: the cloud has a power cord, and access to intelligence depends partly on who can obtain the infrastructure to run it. This chapter does not repeat that allocation problem. It asks what happens after infrastructure, capital, models, and distribution channels become connected through a small number of organizations.

The 2026 Stanford AI Index gives the scale of the shift. Industry produced more than 90% of notable AI models in 2025. Global AI compute capacity grew rapidly, reaching an estimated 17.1 million H100-equivalents. Nvidia accounted for more than 60% of that compute, while Google and Amazon supplied much of the remainder through their own accelerators. Stanford also identifies TSMC as the foundry fabricating almost every leading AI chip.

None of those facts proves monopoly. They show something different: frontier AI depends on several concentrated layers at once. A model company can compete vigorously with another model company while both depend on the same chip designer, the same foundry, one of a few cloud platforms, and a small number of distribution channels. Plurality above can coexist with concentration below.

The Control Stack

There are at least five places where control can accumulate. The first is compute. Frontier development requires access to large clusters of specialized hardware, networking, memory, power, and data-center capacity. Chapter Nine examined who gets that capacity. Here the relevant question is whether a model developer has more than one realistic path to obtain it.

The second is capital. Frontier systems consume extraordinary amounts of money before they produce a finished model. Capital providers can therefore become strategic partners rather than passive investors, especially when financing is tied to infrastructure purchases, commercial rights, or long-term commitments.

The third is the model itself: what it learned during training, how it is refined afterward, its safety controls, and the connections that let other software use it. Contracts determine who may use or change it.

The fourth is distribution: how the model reaches customers. An excellent model may have less influence than a slightly weaker one that is already built into a computer, search service, office software, or cloud contract people use.

The fifth is exit: whether a customer or developer can realistically change providers. An alternative may exist, but moving to it can mean rebuilding software connections, transferring data, retraining staff, and paying to end a contract. Those burdens are switching costs. A company can gain control through these dependencies, or through contracts and distribution, without owning a share of the model developer.

The choke point may sit one layer away from the model.

The FTC Saw the Partnership Problem Early

In January 2025, the Federal Trade Commission published a staff study of three major relationships: Microsoft and OpenAI, Amazon and Anthropic, and Google and Anthropic. The Commission did not find that the partnerships were illegal. Its staff report was explicitly a market study, not an enforcement judgment.

What the report exposed was the architecture of dependence. Across the partnerships, the FTC identified equity and revenue-sharing rights, consultation or control provisions, exclusivity terms, large cloud-spending commitments, discounted compute, technical and business information sharing, and opportunities to integrate models into the cloud providers’ own products.

The FTC highlighted three competition questions. Could the relationships affect access to scarce inputs such as compute and engineering talent? Could contractual and technical switching costs make it harder for model developers to change cloud providers? Could a cloud partner gain sensitive information that would be unavailable to competitors while also developing its own models and applications?

Those questions matter even if every contract is lawful. The concern is structural. When the financier, infrastructure provider, distribution channel, model partner, and potential model competitor are connected, ordinary labels such as investor, supplier, and customer stop describing the whole relationship. A dependency becomes a governance problem when the dependent party cannot realistically leave.

The OpenAI Partnerships Show Control Moving in Real Time

The OpenAI and Microsoft relationship is useful precisely because it has not remained static. In October 2025, the companies described an arrangement that included Microsoft investment, intellectual-property rights, Azure API exclusivity, revenue sharing, and a large OpenAI commitment to purchase Azure services. By April 2026, the companies amended the agreement again.

Under the April 2026 terms, Microsoft remained OpenAI’s primary cloud partner and retained a license to OpenAI model and product intellectual property through 2032, but that license became non-exclusive. OpenAI gained the ability to serve all of its products to customers across any cloud provider. Revenue-sharing terms changed, while Microsoft remained a major shareholder.

On April 28, 2026, OpenAI and Amazon announced that OpenAI frontier models and Codex were coming to Amazon Bedrock. The February strategic partnership had already tied OpenAI to major Amazon investment and infrastructure commitments. OpenAI’s own description of its 2026 financing strategy reduced the problem to three requirements: compute, distribution, and capital.

That formulation is unusually revealing because it describes the control stack from the developer’s side. A frontier lab needs intelligence, but intelligence alone is not enough. It needs money to build, infrastructure to run, and channels through which customers can obtain it. The 2026 changes also show why concentration should not be described as permanent captivity. Large model developers can use their own scale to renegotiate dependencies, add providers, and create competing paths. Power in the stack is concentrated, but it is also bargained over.

Distribution Is Power

The easiest form of control to miss is the default. Most people do not choose technology from a neutral shelf. They encounter whatever appears in the device, browser, search product, office suite, cloud marketplace, or procurement system already in front of them.

The Google search antitrust case turned this distribution issue into a legal remedy. In September 2025, after finding unlawful monopolization in search, the federal court prohibited Google from maintaining specified exclusive distribution arrangements involving Google Search, Chrome, Google Assistant, and the Gemini app. The remedy also barred agreements that prevented partners from simultaneously distributing another general search engine, browser, or generative AI product.

The case was about Google’s search monopoly, not a judicial finding that Gemini itself was monopolized. That distinction matters. The AI lesson comes from the remedy: distribution channels can reinforce power before a user ever compares model quality.

The default is a distribution decision with market consequences. This is the softer authority-jump logic from Chapter Fourteen applied to markets: the architecture chooses the menu. In competition, the menu can determine which models get the opportunity to become habits, platforms, and standards.

Concentration Is Not the Same as Monopoly

There is an obvious danger in overstating the case. Frontier AI is not a market with one uncontested model provider. The 2026 Stanford AI Index found leading models from Anthropic, xAI, Google, OpenAI, Alibaba, and DeepSeek clustered near the top of human-preference rankings. For its March 2026 snapshot, Stanford reported that Anthropic’s top model led the leading Chinese model by 2.7%, while the leading closed model held a 3.3% advantage over the leading open model. These figures describe the reported ranking comparisons, not universal measures of capability.

Anthropic offers another counterexample to simple lock-in. In April 2026 it described Claude as available across AWS, Google Cloud, and Microsoft Azure while training and running on AWS Trainium, Google TPUs, and Nvidia GPUs. Amazon remained its primary cloud and training partner, but Anthropic was deliberately diversifying hardware and distribution. The April expansion also paired new Amazon investment with long-term AWS spending commitments. Broader hardware and distribution choices therefore coexisted with deeper financial ties to a major infrastructure provider.

Several developments can weaken these control points. Customers may use more than one cloud provider. Open-weight models make the model’s learned numerical settings available, although licenses and other restrictions can still apply. Alternative chips can offer another route to computing power. International competitors and lower costs for running models can broaden choice. A valuable market can attract new suppliers.

The argument of this chapter is not that a handful of companies secretly control one machine called AI. The claim is narrower: competition at the model layer does not by itself guarantee independent paths through the layers beneath and around the model. Competition is not only how many models exist. It is how many independent paths exist to build, run, distribute, and leave them.

Exit Is Part of Accountability

The Delegation Doctrine usually asks whether a human owner can stop a machine. At the institutional level, another stop question appears: can an organization stop relying on a provider?

An organization may keep the legal right to choose while losing the practical ability to switch. Its records, automated tasks, sign-in systems, software connections, and staff routines may all depend on one platform. Canceling the contract remains possible on paper. Keeping the work running after cancellation may be much harder.

This is not an argument against integration. Deep integration can improve security, reliability, governance, and usefulness. It is an argument for treating reversibility as part of procurement rather than as a problem to discover after dependence has formed.

A serious buyer should know what can be exported, what can be moved, which interfaces are portable, how long migration would take, what functions would fail during a provider outage, and whether a second model or cloud can perform critical work. Chapter Eighteen’s outage test protected human competence. Here the institutional outage test applies the same maintenance instinct: selected consequential workflows should occasionally prove that the organization still has a viable alternative path.

Exit is part of control.

Public Capacity Can Be a Counterweight

Private concentration is not answered only by antitrust. One response is to create additional capacity that does not depend on a frontier company deciding a project is commercially attractive.

The National Science Foundation’s National Artificial Intelligence Research Resource is an early example. NAIRR began as a pilot connecting researchers and educators to compute, data, models, software, and expertise contributed by federal agencies and private partners. In its two-year pilot snapshot, NSF reported support for more than 600 research projects and 6,000 students across all 50 states, Washington, D.C., and Puerto Rico.

On September 1, 2026, NSF reported that the pilot had supported more than 800 research projects and announced an operations center intended to move NAIRR toward sustained national infrastructure. The initiative still relies heavily on public-private partnership, so it does not remove private control from the stack. It creates another path into the stack for researchers who would otherwise lack access.

That distinction is important. The goal of competition policy and public infrastructure should not be to guarantee that every organization can build a frontier model. It should be to prevent the frontier from becoming reachable through only one institutional doorway. Plural access is a form of resilience.

Current Evidence: Concentration and Diversification at Once

Reviewed October 20, 2026. DeepSeek procurement notice added; earlier evidence retained. Cloud access and policy evidence added October 22, 2026.

The 2026 evidence points in two directions simultaneously. Frontier AI remains capital- and infrastructure-intensive, with industry producing more than 90% of notable models and critical hardware concentrated among a small number of designers, cloud providers, and foundries. At the same time, model performance is competitive across several U.S. and Chinese companies, open models remain relevant, and the largest developers are adding infrastructure partners rather than relying on a single path.

The FTC’s 2025 partnership report remains the clearest institutional map of the competition risks created when cloud providers invest in model developers while supplying compute and distributing their products. The report should be treated as a structural baseline rather than a description frozen in 2025 because the contracts have already changed materially.

OpenAI’s April 2026 Microsoft amendment and its expanded AWS availability demonstrate that bargaining power can alter the stack. Anthropic’s 2026 expansion across AWS, Google, Microsoft, Trainium, TPUs, and Nvidia hardware shows the same diversification strategy from another frontier developer, alongside continuing financial ties to the infrastructure providers supplying those routes. OpenAI systems are used in this Living Edition’s production workflow, and Anthropic systems are used in the external critique workflow, so both companies’ descriptions of their partnerships are treated as disclosed provider evidence rather than neutral market analysis.

On October 21, 2026, OpenAI announced plans to bring its models and Codex workflows to Google Cloud. In a separate announcement, it said partners must maintain consistent safeguards and coordinate enforcement across access routes. Updated usage policies, effective November 20, require prior approval for services offering direct access to OpenAI models. The policies distinguish those intermediaries from applications that merely use a model to provide a feature. These are company announcements and terms, not proof of effective enforcement.

A buyer should distinguish another cloud supplier from an independent source of model access. Several routes can still depend on one company's permission. Before calling a backup independent, check whether it uses the same model, shares the same access restrictions, and can continue the essential work if that provider withdraws access. More places to buy a service do not necessarily create another place to turn when permission ends.

The Google remedies provide a separate signal at the distribution layer. The federal court’s restrictions on exclusive arrangements involving Gemini recognize that control over defaults and placement can matter to competition in generative AI even when the underlying liability case arose from search.

The public-capacity side is also moving. NSF’s September 2026 NAIRR Operations Center announcement turns a pilot access program into the beginning of a longer-lived institutional counterweight for research and education. The pattern is not simple consolidation or simple decentralization. It is a race between concentration at expensive control points and diversification around them.

California’s September 2026 data-center package shows the physical side of AI governance becoming explicit law. The state required more disclosure and oversight around electricity use, water consumption, land use, workforce needs, grid costs, and local review for data centers tied to AI expansion. That matters because frontier AI is not only a model running in the abstract. It is also a demand on power, water, land, transmission, and public tolerance. When private computation requires public infrastructure, accountability cannot stop at the model provider’s terms of service. The delegation chain must include who pays for capacity, who bears environmental and utility risk, and who has authority to say no when local systems are asked to carry private scale.

A U.S. Department of Homeland Security notice filed October 19, 2026, for publication October 20 described a July 24 order requiring covered federal civilian agencies to remove DeepSeek products, including third-party products and services developed using DeepSeek implementations. The order provides an exceptions process. It does not create a general ban on private use.

The delegation boundary therefore reaches past the company name on the contract. An organization needs to know which models its suppliers actually use, who can change them, and what must happen when a component is no longer permitted. A removal rule cannot be enforced reliably if nobody can find the model inside the software.

On October 30, 2026, the UK Competition and Markets Authority opened an investigation into Meta's commercial terms. The regulator said the terms could let Meta collect content from publishers' and businesses' websites for its own products, including AI features, without a choice about that use. Meta had paused the revised terms pending consultation. The investigation had not established that Meta held a dominant position or had broken competition law.

The question is whether permission for one relationship is being used to authorize another. A publisher may need a platform to reach readers without wanting to supply material for its AI. Before accepting a platform agreement, an organization should identify which additional uses it permits and what refusing them would cost. A formal choice offers less practical control when exercising it means losing access to an essential audience.

Accountability Analysis

Principles 1, 3, 4, 5, and 6 carry this chapter. Principle 1 prevents an organization from treating dependence on a model or cloud provider as a transfer of accountability. A hospital, court, company, school, or government agency still answers for the decisions it delegates through the system it chose.

Principle 3 requires the institutional boundary to include the provider relationship. The contract should specify not only what the model may do, but what the provider may change, what data it may retain, what access it may restrict, what notice accompanies material changes, and what the customer can take elsewhere.

Principle 4 makes exit and override practical questions. An organization that cannot stop, substitute, or meaningfully challenge a critical provider has less control than its contract may imply.

Principle 5 moves accountability upstream to architecture and procurement. By the time a model is embedded across identity, data, workflows, and distribution, a nominal choice to switch may arrive years too late. The consequential decisions were the earlier ones that created the dependency.

Principle 6 asks where benefits and consequences land. Concentrated control becomes an accountability problem when the economic benefit of scale is retained by the controlling parties while the costs of failure, exclusion, migration, or abrupt policy change fall on dependent users who have little leverage.

The Accountability Chain remains the same: Who answers? What is the boundary? Can the owner see it? Can the owner stop it? Do the consequences reach the owner? In this chapter, the owner must be able to answer those questions not only about the model, but about the stack that makes the model available.

Trace an exit route

Choose one AI service used by an organization you know. Ask what would have to move if the provider became unavailable: records, software connections, staff routines, or all three. Identify an alternative and one obstacle to using it. This is a planning exercise; a real outage test should protect essential services and sensitive information.

Accountability Verdict

Concentration is not automatically abuse, and scale is not automatically illegitimate. Frontier systems require enormous investment, specialized infrastructure, and integration. Those facts can produce better products, lower costs, stronger security, and faster innovation.

The accountability problem begins when concentrated control becomes hidden dependency. The relevant question is not merely who owns the model company. It is who can constrain compute, finance development, set access terms, shape distribution, change the rules, observe sensitive information, and make exit expensive.

The practical answer is to keep independent paths available. That means using multiple suppliers where feasible, choosing systems that can exchange information, and making it possible to move data and software connections. Buyers should count the cost of changing providers before signing. Public research access and competition enforcement can help where suppliers can block alternatives. The institutions choosing the systems remain accountable for those choices.

The model is the visible layer. Power lives in the dependencies.

Chapter Twenty-Three turns the question around. What happens when the institution with the power to choose, buy, deploy, and compel is the state itself?

Sources and Notes

1. Stanford Institute for Human-Centered Artificial Intelligence, “Research and Development,” The 2026 AI Index Report, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development. Accessed September 12, 2026. Stanford reports that industry produced more than 90% of notable AI models in 2025; global AI compute capacity reached an estimated 17.1 million H100-equivalents; Nvidia accounted for more than 60% of total compute; and TSMC fabricates almost every leading AI chip. Chapter Nine retains the electricity-allocation argument; this chapter uses the infrastructure record to establish concentration of control points. Source record: DA-CH22-S1. Archive reference: pending.

2. Stanford Institute for Human-Centered Artificial Intelligence, “Technical Performance” and report overview, The 2026 AI Index Report, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance and https://hai.stanford.edu/ai-index/2026-ai-index-report. Accessed September 12, 2026. As of March 2026, leading models from Anthropic, xAI, Google, OpenAI, Alibaba, and DeepSeek occupied the top tier of Arena ratings; Stanford identified Anthropic’s top model as leading the top Chinese model by 2.7%; and the top closed model led the top open model by 3.3%. These are dated ranking comparisons, not general capability margins. Anthropic systems are used in this Living Edition’s external critique workflow; see the transparency statement. The chapter uses this as a counterweight to any claim that frontier-model competition has collapsed into a single provider. Source record: DA-CH22-S2. Archive reference: pending.

3. Federal Trade Commission, Office of Technology Staff, “Partnerships Between Cloud Service Providers and AI Developers,” FTC Staff Report on AI Partnerships and Investments 6(b) Study, January 2025. https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted.pdf. Accessed September 12, 2026. The staff study examines Microsoft-OpenAI, Amazon-Anthropic, and Google-Anthropic relationships and describes equity and revenue-sharing rights, consultation or control provisions, exclusivity terms, cloud-spending commitments, discounted compute, information sharing, switching costs, and access to scarce inputs. The report is a market study, not a finding that any party engaged in illegal conduct. Source record: DA-CH22-S3. Archive reference: pending.

4. OpenAI and Microsoft, “The next phase of the Microsoft-OpenAI partnership,” April 27, 2026. https://openai.com/index/next-phase-of-microsoft-partnership/ and https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/. Accessed September 12, 2026. The amended agreement states that Microsoft remains OpenAI’s primary cloud partner; OpenAI products ship first on Azure subject to the stated capability condition; OpenAI may serve products across any cloud provider; Microsoft’s model and product IP license runs through 2032 and is non-exclusive; revenue-sharing terms changed; and Microsoft remains a major shareholder. These are joint provider statements, not independent competition analysis. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH22-S4. Archive reference: pending.

5. OpenAI, “OpenAI models, Codex, and Managed Agents come to AWS,” April 28, 2026; Amazon Web Services, “Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents,” April 28, 2026; and AWS, general-availability update, June 1, 2026. https://openai.com/index/openai-on-aws/ and https://aws.amazon.com/about-aws/whats-new/2026/04/bedrock-openai-models-codex-managed-agents/ and https://aws.amazon.com/about-aws/whats-new/2026/06/amazon-bedrock-openai-models-codex-generally-available/. Accessed September 12, 2026. The companies announced OpenAI frontier models and Codex on Amazon Bedrock in April, with general availability following in June. The chapter uses current availability to describe diversification of OpenAI distribution. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH22-S5. Archive reference: pending.

6. OpenAI, “Scaling AI for everyone,” February 27, 2026. https://openai.com/index/scaling-ai-for-everyone/. Accessed September 12, 2026. OpenAI describes compute, distribution, and capital as the three requirements for scaling its products and announced major 2026 investment and infrastructure partnerships. The chapter uses that formulation as the developer’s own description of the resources required to scale frontier AI, not as neutral evidence about market structure. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH22-S6. Archive reference: pending.

7. Anthropic, “Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute,” April 6, 2026; and “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute,” April 20, 2026. https://www.anthropic.com/news/google-broadcom-partnership-compute and https://www.anthropic.com/news/anthropic-amazon-compute. Accessed September 12, 2026. Anthropic states that it trains and runs Claude across AWS Trainium, Google TPUs, and Nvidia GPUs, with Amazon remaining its primary cloud and training partner, and that Claude is available through AWS, Google Cloud, and Microsoft Azure. The April 20 statement also reports new Amazon investment and Anthropic’s commitment to substantial AWS spending over ten years. These are provider descriptions of both diversification and continued financial interdependence, not independent competition analysis. Anthropic systems are used in this Living Edition’s external critique workflow; see the transparency statement. Source record: DA-CH22-S7. Archive reference: pending.

8. U.S. Department of Justice, Antitrust Division, “Department of Justice Wins Significant Remedies Against Google,” September 2, 2025. https://www.justice.gov/opa/pr/department-justice-wins-significant-remedies-against-google. Accessed September 12, 2026. Following the search-monopolization judgment, the court barred specified exclusive distribution arrangements involving Google Search, Chrome, Google Assistant, and Gemini and prohibited partners from being barred from simultaneously distributing another general search engine, browser, or generative AI product. The chapter does not characterize this as a finding that Gemini itself was monopolized. Source record: DA-CH22-S8. Archive reference: pending.

9. U.S. National Science Foundation, “National Artificial Intelligence Research Resource,” current program page; and “NSF establishes operations center for the National Artificial Intelligence Research Resource,” September 1, 2026. https://www.nsf.gov/focus-areas/ai/nairr and https://www.nsf.gov/cise/updates/nsf-establishes-operations-center-national-artificial. Accessed September 12, 2026. NSF’s two-year pilot snapshot reports more than 600 research projects and 6,000 students across all 50 states, Washington, D.C., and Puerto Rico. Its September 1, 2026 release separately reports more than 800 research projects and describes the new Operations Center as the operational backbone for sustained national research infrastructure connecting compute, data, models, tools, support, and training. The student count belongs to the earlier snapshot. Source record: DA-CH22-S9. Archive reference: pending.

10. Office of Governor Gavin Newsom, “Governor Newsom signs most comprehensive data center laws in the nation, providing communities more control on water, electricity, and land use,” September 21, 2026. https://www.gov.ca.gov/2026/09/21/governor-newsom-signs-most-comprehensive-data-center-laws-in-the-nation-providing-communities-more-control-on-water-electricity-and-land-use/. Reuters, “California governor signs broad data center oversight bill package,” September 21, 2026. https://www.reuters.com/legal/litigation/california-governor-signs-broad-data-center-oversight-bill-package-2026-09-21/. Accessed September 22, 2026. The governor’s announcement identifies seven signed bills and describes information on energy, water, workforce and land use, grid-cost allocation, water-supply upgrades and conditions for environmental-review streamlining. Enactment is distinguished from implementation and demonstrated effectiveness. Chapter Nine retains electricity allocation; this chapter addresses public authority and accountability over infrastructure dependencies. Source record: DA-CH22-S10. Archive reference: pending.

11. U.S. Department of Homeland Security, "Notice of Order Regarding Exclusion and Removal of DeepSeek Products from Federal Civilian Executive Branch Information Technology Systems and Procurements," docket DHS-2026-0628, Federal Register document 2026-21447, filed October 19, 2026, scheduled publication October 20, 2026; accompanying Exclusion and Removal Order #2026-01, original release date July 24, 2026. https://public-inspection.federalregister.gov/2026-21447.pdf and https://www.federalregister.gov/d/2026-21447. Accessed October 20, 2026. Government notice and signed order establish the stated procurement/removal scope and exceptions process, not proof that every security allegation is independently established or that removal has been completed. National-security and sensitive compartmented systems are outside this order's stated coverage. The second paragraph is the author's analysis of procurement visibility and enforceable dependency boundaries. Source record: DA-CH22-S11. Update 106. Archive reference: Federal Register document 2026-21447, signed-order appendix.

12. OpenAI, “OpenAI models are coming to Google Cloud,” October 21, 2026, https://openai.com/index/openai-models-google-cloud/; “Broad access through consistent standards,” October 21, 2026, https://openai.com/index/broad-access-through-consistent-standards/; “Usage policies,” updated October 21, effective November 20, 2026, https://openai.com/policies/usage-policies/. Accessed October 22, 2026. Company-described distribution plans and access terms, not independent validation of implementation, enforcement or market effects. Intermediary approval does not apply solely because an application integrates a model. The second paragraph is author analysis. OpenAI systems are used in this Living Edition's production workflow; see the transparency statement. Source record: DA-CH22-S12. Update 112. Archive reference: pending.

13. Competition and Markets Authority, “CMA opens investigation into Meta’s commercial terms,” October 30, 2026. https://www.gov.uk/government/news/cma-opens-investigation-into-metas-commercial-terms--2 Accessed October 31, 2026. Official opening of an investigation under the Competition Act 1998, not a finding of dominance or infringement. The revised terms were paused pending consultation; the earlier October30 and November20 implementation dates must not be treated as current operative dates. The regulator's concerns remain attributed. The second paragraph is author analysis of distribution dependence and scope of permission. Source record: DA-CH22-S13. Update 124. Archive reference: pending.