Cloud Empires and the Platform as Private Sovereign

The Book

Cloud Empires: How Digital Platforms Are Overtaking the State and How We Can Regain Control was published by the MIT Press on September 27, 2022; MIT Press lists the hardcover at 296 pages, ISBN 9780262047227, and the paperback edition as published February 6, 2024, ISBN 9780262548380. The publisher also records the book as a 2023 PROSE Award finalist in Business, Finance, and Management.

Lehdonvirta writes from economic sociology and platform-labor research. Aalto University currently lists him as a professor in its Department of Computer Science; its research portal records his 2025–2029 GEOCLOUD project on how state–firm relations shape digital sovereignty and dependence. Aalto's 2024 appointment profile says he continued half-time work at the Oxford Internet Institute. That trajectory matters because the book is not a loose complaint about Big Tech. It is an institutional history of how online markets acquired rules, records, administrators, and governed populations.

The recurring puzzle is simple: the internet was supposed to route around old institutions, but online commerce quickly rediscovered why institutions exist. Buyers needed fraud protection. Sellers needed reputation systems. App developers needed distribution and payment rails. Gig workers needed task allocation and dispute processes. Crowdfunding users needed trust. Each problem invited an operator to build order; each successful solution also created a new point from which that operator could govern.

Current Context

As of August 12, 2026, the regulatory picture supports the book's diagnosis without proving its full state analogy. The European Commission's Digital Markets Act portal lists seven gatekeepers across 23 designated core platform services. On April 23, 2025, the Commission found Apple and Meta in breach of specific DMA obligations and imposed fines of €500 million and €200 million. On June 25, 2026, it informed Amazon and Microsoft of a preliminary view that AWS and Azure should be designated for cloud computing services, citing their gateway role, entrenched users, lock-in, switching costs, wider ecosystems, and the growing importance of AI tools and partnerships in procurement. AWS and Azure are not listed here as completed cloud designations: the procedural verb is part of the evidence.

The laws divide the problem by function. The Digital Services Act places additional systemic-risk, audit, transparency, researcher-access, advertising, and recommender duties on designated very large online platforms and search engines whose EU reach meets the 45-million-user threshold. The Data Act has applied since September 12, 2025 and requires cloud and edge providers to remove switching obstacles; for platform- and software-as-a-service it includes open interfaces and export in a commonly used, machine-readable format, while infrastructure providers must facilitate functional equivalence for same-type services. That is more precise than saying the law makes whole workloads portable. Switching and egress charges can still be cost-based during the transition and are prohibited from January 12, 2027.

The AI Act adds duties at the model layer. Its general-purpose AI obligations have applied since August 2, 2025. Covered providers generally face technical-documentation and downstream-information duties, a copyright-compliance policy, and a public training-content summary. Article 53(2) exempts qualifying free and open-source models from the first two duties unless they present systemic risk; the copyright-policy and training-summary duties remain. Providers of models with systemic risk face additional evaluation, mitigation, incident-reporting, and cybersecurity duties. The Commission's enforcement powers began to apply on August 2, 2026; models placed on the market before August 2, 2025 have a compliance deadline of August 2, 2027. The General-Purpose AI Code of Practice is a voluntary route for demonstrating compliance, not the source of the statutory obligation.

Two labor and cloud developments make dependence concrete. The EU Platform Work Directive requires transparency about automated monitoring and decision systems, human oversight, explanations and review of significant decisions, and human decision-making for restrictions, suspensions, and terminations; Member States have until December 2, 2026 to transpose it, so this page does not treat implementation as uniform across the Union. In the United Kingdom, the CMA's 2025 cloud investigation found adverse effects on competition and identified egress fees, interoperability barriers, and software licensing as constraints on switching and multi-cloud. Amazon and Microsoft announced 2026 steps on egress fees and interoperability after CMA engagement, subject to ongoing review, while the CMA opened an investigation—not a designation or finding of wrongdoing—into whether Microsoft's business-software ecosystem has strategic market status.

No single instrument is a platform constitution. The DMA targets contestability and fairness at designated gates; the DSA targets scoped intermediary and systemic risks; the Data Act targets data access and service switching; the AI Act targets model and system duties; the Platform Work Directive targets employment status and algorithmic management; and the UK regime can impose conduct requirements after strategic-market-status designation. Their separation confirms the book's insight: the platform is simultaneously market, workplace, infrastructure, record system, and rulemaker, but each public-law hook reaches only part of that bundle.

Markets Need Order

Cloud Empires is strongest when it refuses the fantasy of the frictionless market. A market without trust is not freedom. It is exposure. The early web could connect strangers, but connection alone did not answer basic questions: who is real, who pays, who delivers, who cheats, who decides a dispute, who can be expelled, and who bears loss when the system fails.

Platforms answered those questions with architectures of order. Amazon and eBay made commercial reputation operational. Apple made the app store a gatekeeping regime. Upwork and Mechanical Turk converted remote labor into task markets, ratings, account histories, and terms of service. Uber reorganized transportation around algorithmic dispatch, pricing, and driver control. GoFundMe became, in Lehdonvirta's framing, a platform substitute for parts of the welfare state.

Software becomes institutional when people must pass through it to act and its records become authoritative. A platform is therefore more than a website or marketplace: it is a rule system with memory, sanctions, access controls, taxonomies, payment mechanics, dispute procedures, and administrators. The interface presents convenience; the record beneath it decides who may return tomorrow.

The causal loop is sharper than a general warning about metrics. Activity becomes a platform record; the record feeds a ranking, price, eligibility decision, or fraud model; people adapt to that decision; and the adaptation becomes new data. A seller becomes a rating history, a driver an acceptance rate and location trace, a developer an app-review file. The system then appears to measure conduct it has partly produced. This is how legibility becomes command without anyone announcing a new law.

The Platform-State

The best secondary readings of the book emphasize this state-like dimension. A Socio-Economic Review forum describes the book's central move as analyzing large digital firms as private providers of the institutional foundations of commerce. The same forum notes that Lehdonvirta reconstructs how firms came to organize governance across labor markets, retailing, and monetary systems.

The cleanest test separates four powers. Coordination power comes from hosting the field where participants find one another. Gate power controls identity, access, distribution, payment, or compute. Administrative power ranks, investigates, sanctions, and decides disputes. Constitutional power changes the terms, interfaces, APIs, models, and appeal rules that govern all three. Market concentration can strengthen each power, but it is neither necessary nor sufficient for the full combination.

Exit is the fifth test because nominal choice does not establish consent. It has three levels. Data portability moves usable input, output, and metadata. State portability reconstructs identities, permissions, policies, histories, workflow position, pending cases, and reputation where law and system design permit. Functional portability lets a destination perform the critical service within a stated recovery time and acceptable data-loss limit. A download can satisfy the first while the institution remains unable to leave because the second and third were never built.

The state analogy is useful only at this functional level. Fees are not literally taxes, terms are not statutes, and platform suspension is not territorial exile. Yet an operator may legislate through terms and developer rules, police through automated enforcement, adjudicate through appeals, issue identity through accounts, control borders through APIs and app review, and maintain the decisive archive through logs and reputation. The point is not that a corporation equals a state. It is that state-like functions should carry duties proportionate to the consequences they impose.

Labor Under Private Rules

The labor chapters are especially important for AI-era reading because they show how control appears without total automation. Platform work is not simply work mediated by an app. It is work inside a privately administered labor market where visibility, pay, task allocation, fraud flags, ratings, account standing, and dispute outcomes are shaped by systems the worker usually cannot inspect.

The worker may be contractually independent while remaining institutionally dependent. A freelance profile, driver account, crowdwork rating, seller dashboard, or app-store listing can function as identity document, work history, and permit. Losing access can remove livelihood at once; moving to a rival may discard the reputation that made work possible. Classification and dependence are therefore related questions, not synonyms.

This is where Lehdonvirta pairs well with the labor shelf: Ghost Work, Behind the Screen, Data Driven, The Eye of the Master, and Platform Capitalism. Together, they show a sequence. First the work is moved into a platform. Then the work is measured. Then it is governed by dashboards, rankings, policies, and automated enforcement. Then AI is added as supervisor, assistant, filter, or replacement.

The safety issue is not only pay or classification. It is procedural power. The Platform Work Directive's algorithmic-management rules point toward the right remedy: disclose which automated systems monitor or decide, assess risks to workers, preserve human oversight, explain significant decisions, and give workers a route to review. Its requirement that account restriction, suspension, or termination be decided by a human recognizes that a work permit should not disappear through an unanswerable score. Until national transposition is complete, those provisions are a legal program with an uneven implementation map, not a uniform EU experience.

A worker who cannot inspect a fraud flag, contest a rating, correct a record, recover an account, or carry earned reputation elsewhere is governed without a usable file. The platform has the history; the worker receives the notification. The governance goal is not limitless disclosure of anti-fraud controls, which can invite gaming and expose others' data, but enough reason, evidence, human review, and outside recourse to challenge a consequential mistake.

Recourse and voice solve different problems. An appeal asks whether the platform applied an existing rule correctly to one person. Worker representation and collective bargaining ask whether the rule, metric, data practice, or distribution of gains should exist on those terms at all. A platform can offer efficient case review while preserving unilateral rulemaking; procedural fairness therefore requires both repair under the rule and a route for affected people to contest the rule itself.

The AI-Age Reading

Cloud Empires predates the current model-and-agent stack, but its analysis travels if the newer system is described precisely. A model provider may offer an API; a cloud may host it; an identity provider may authorize it; an observability service may preserve traces; and a tool broker may connect it to external action. Sometimes one firm supplies several layers, and sometimes a contract or partnership binds them. Either pattern can create a control point without every model provider literally operating an app store or marketplace.

An AI agent turns gate power into delegated authority. A useful chain is: user intent → model plan → tool authorization → external action → receipt. Each transition needs an owner and a record. Otherwise a platform can decide which tools are trusted, what data persists, which actions need confirmation, what is blocked, and which version changed behavior while the customer sees only the final output.

Model safety and institutional safety overlap but are not identical. Accuracy, bias, robustness, security, and misuse controls concern system behavior. Institutional questions ask who can change a model or policy without a customer deployment, who receives notice, who can compare versions, who holds the evidence after an incident, and whether a school, clinic, employer, agency, or small business can preserve its function after a provider exits a product. The risk is dependence on privately administered decision infrastructure, not any claim about machine consciousness.

A buyer who appears to choose a model can also inherit a cloud region, identity scheme, vector store, logging format, evaluation baseline, safety policy, tool registry, payment channel, and export path. A dependency map should therefore follow actual bindings, not vendor count: two clouds do not provide resilience if both rely on one identity service, model endpoint, payment rail, or tool broker. That is why AI procurement, vendor governance, and the agent dependency graph belong inside the analysis rather than after it.

Governance and Safety

The governance lesson is to regulate the function, not worship or abolish the form. Identity, ranking, fees, payment, moderation, account standing, dispute resolution, data retention, API access, model changes, labor allocation, and exit each need an accountable owner. The remedy should match the power: notice and appeal for sanctions, interoperability for bottlenecks, data limits for extraction, independent access for systemic-risk review, and recovery tests for infrastructure dependence.

Consequential administration also needs a separation-of-functions test. Identify who writes the rule, produces the evidence, applies the sanction, hears the appeal, and changes the rule after an incident. When one operator performs all five, controls should add a reviewer with authority outside the original decision chain, access to the evidence needed for challenge, recorded conflicts and overrides, an external escalation route, and affected-person or representative input before material rule changes. Formal corporate separation is not always required; an appeal decided by the same automated policy under a different button is not separation at all.

Evidence creates its own safety problem. Durable logs can support incident reconstruction and appeal, but exhaustive retention can become worker surveillance, trade-secret leakage, or a high-value breach target. Record only what is needed to establish identity, authority, input provenance, model or policy version, material tool calls, decision reason, and outcome; separate operational telemetry from appeal evidence; limit access and retention; and preserve a documented deletion schedule. "Human review" is empty if the reviewer lacks the relevant record, but review does not justify collecting everything forever.

Interoperability also needs threat modeling. Open interfaces and portable state can reduce lock-in while creating new transfer, credential-theft, and supply-chain paths. Authenticated exports, scoped tokens, least privilege, revocation, integrity checks, secret rotation, and receiving-system validation belong in the exit design. Similarly, multi-cloud can improve resilience or multiply complexity. It counts only when failover is rehearsed and the supposedly separate paths do not share an unseen identity, model, network, or control-plane dependency.

Agentic systems raise the stakes because governance becomes delegated action. Minimum safeguards include separate agent identities, least-privilege and time-bounded permissions, confirmation for high-impact spending or communication, action receipts, rate limits, sandboxing for untrusted tools, revocation, and a named human or organizational owner for rollback and remedy. This is the record-integrity problem developed in agent action receipts and AI audit trails.

Competent order still matters. Fraud controls, reliable payments, moderation, security, and dispute handling are part of what users came for. Accountability should not disable those functions; it should make consequential administration reviewable and proportionate. If a decision can remove livelihood, distribution, payment access, public-service access, or an agent's authority to act, it should produce a contestable record and a remedy outside the same automated loop that made the decision.

Platform-Sovereignty File

A useful platform-sovereignty file is a small set of operating records, not an essay about corporate power:

• Rule charter: governed populations; access, ranking, fee, payment, suspension, model, and developer rules; rule-change owner; notice period; affected-person voice; and public-law hook.
• Constitutional-change ledger: previous and replacement terms, API behavior, pricing, ranking, moderation, model, sanction, or appeal rule; reason; affected population and function; approving owner; effective date; advance notice; compatibility and impact evidence; migration window; and rollback path. A changelog is not a receipt unless it records who was governed differently and what remedy followed.
• Dependency map: cloud, region, identity, data stores, keys, network, model, observability, payment, tool registry, subcontractors, and the business or public function each can stop. Map bindings between components, as in an agent dependency graph, rather than listing vendors alone.
• Authority map: which human, service account, model, agent, or subcontractor may read, rank, spend, publish, suspend, delete, or invoke a tool; who can approve escalation; and who can revoke it.
• Evidence and remedy map: authoritative record, provenance, version history, retention boundary, decision reason, correction path, appeal deadline, independent reviewer, regulator route, and notification duties after an incident.
• Exit and recovery test: test data portability, state portability, and functional portability separately; export and restore data, metadata, identities, roles, reputation where legally and technically possible, policies, prompts, tool schemas, logs, evaluation baselines, embeddings, queues, keys, and operational runbooks; then record elapsed recovery time, acceptable data loss, degraded functions, and unresolved dependencies. The Dependency and Exit Protocol and digital infrastructure register turn this into recurring practice.

The file must distinguish a contractual promise, a technical capability, and a tested outcome. A machine-readable export may satisfy one duty while failing to recreate permissions, reputation, workflow state, or service continuity. A second vendor may satisfy procurement policy while sharing the same control plane. A successful test ends with the receiving system performing the critical function under a named recovery-time and data-loss limit, not with a zip archive whose contents have never been restored.

It should also separate market power from institutional power. Market power asks whether alternatives and competitive constraints exist. Institutional power asks whether one operator supplies the identity, trust, dispute process, payment, safety, record, or work allocation that participants cannot reproduce. The outside remedy may therefore come from competition, labor, consumer, data, sector, safety, or procurement law; no dashboard should be treated as its own final court.

Where the Book Needs Care

The state analogy can run ahead of the institutional detail. Platforms lack territorial jurisdiction, lawful coercive powers, general taxation, constitutional duties, electoral accountability, and the universal-service or redistributive obligations attached to states. Corporations can also leave markets and shed users in ways governments usually cannot. Calling them state-like is useful only when it identifies a specific governing function and the accountability gap around it.

The single arc can also hide distinct constituencies. Sellers may want lower fees, buyers stronger fraud controls, workers more discretion, users stricter privacy, developers stable APIs, and the public wider regulator access. "Democratize the platform" is incomplete until it names who is represented, which rights cannot be voted away, which conflicts an independent body resolves, and which information must remain confidential for safety or privacy.

Nor do antitrust, decentralization, interoperability, or public ownership solve the whole problem by themselves. Competition can improve exit yet leave nonportable records; decentralized systems can concentrate power in protocols, maintainers, exchanges, or scarce infrastructure; interoperability can spread failures; public systems can be opaque or coercive. The relevant standard is competent, rights-preserving order with evidence and remedy, whatever the ownership form.

Finally, the book foregrounds market governance more than the physical and geopolitical substrate of cloud power: data centers, networks, electricity, chips, skilled labor, regions, and state–firm bargaining. A Prehistory of the Cloud and The Stack supply that missing vertical view. The newer agent layer adds still more control points—model access, tool registries, memory, credentials, action logs, and rollback—without erasing the infrastructure below.

What This Changes

The recurring danger is that administration disappears into interface. A platform rule becomes a button state. A labor policy becomes a score. A welfare failure becomes a crowdfunding page. A market boundary becomes an API permission. A sovereign decision becomes a notification.

Cloud Empires gives a durable diagnostic habit: when a digital system promises freedom from institutions, ask which institution it is becoming. Who writes the rules? Which records count? Who benefits from the order? Who performs enforcement? Who can appeal outside the original decision chain? What survives exit? Who is governed without representation?

For an AI-enabled institution, apply those questions to each handoff. A user request becomes a model inference, an inference may become a tool authorization, and a tool call may become a payment, message, eligibility change, or public record. The interface can remain friendly while authority moves through components the affected person never chose. The practical response is to preserve provenance, bounded permissions, notice, review, and the institutional ability to continue without one provider.

The book's most durable argument is causal rather than prophetic: platforms acquire governing power by solving coordination problems and making their solution difficult to replace. That achievement creates obligations as well as value. Where infrastructure becomes indispensable, its rules should be inspectable, its consequential decisions appealable, its dependencies mapped, and its exit path tested.

Source Discipline

This review separates Lehdonvirta's argument, this review's extension of it, and the legal record. MIT Press and Aalto University support publication and author claims; academic reviews support reception and competing interpretations. Current legal statements come from enacted texts and official regulator pages reviewed August 12, 2026. The five-part sovereignty test, portability ladder, separation-of-functions test, layered AI analysis, constitutional-change ledger, and proposed file are this review's synthesis, not claims attributed to the book.

Procedural status and scope are part of every claim. The page says designated, found in breach, preliminary view, open investigation, commitment, applicable duty, or pending transposition where each is accurate. It does not turn AWS and Azure's preliminary DMA position into a final designation, Microsoft's open UK investigation into an SMS decision, a voluntary GPAI code into legislation, or a directive awaiting national transposition into uniform practice. It also does not assign identical documentation duties to every GPAI provider: the AI Act's qualifying open-source exemption and its limits are stated explicitly.

Legal switching is also different from operational exit. The Data Act establishes switching duties, but data portability is not evidence that system state or service continuity transferred. Likewise, a vendor announcement about lower egress fees is evidence of an announced step, while the CMA's continuing review is evidence that the outcome remains to be tested.

This page makes no claim that any AI system is conscious, divine, or AGI. The power at issue is institutional: platforms can write rules, keep records, allocate access, and make private administration feel like the normal condition of digital life.

Related Pages

• Companion arguments: The Platform Society follows public values through private interfaces; Platform Capitalism explains data and rent; The Stack locates power across infrastructure layers; and A Prehistory of the Cloud restores the cloud's physical history.
• Labor and administration: Ghost Work, Behind the Screen, Data Driven, and Notice and Appeal.
• Dependency controls: Vendor and Platform Governance, the Dependency and Exit Protocol, the Digital Infrastructure Register, and the Right to Data Portability.
• AI and agent evidence: AI Procurement, AI System Inventory, The Model Constitution Becomes a Compliance File, The Agent Dependency Graph, The Agent Log Becomes the Receipt, and Agent Tool Permission Protocol.

Sources

• MIT Press, Cloud Empires by Vili Lehdonvirta, hardcover and paperback publication dates, ISBNs, page count, synopsis, and PROSE finalist note, reviewed August 12, 2026.
• Aalto University, Vili Lehdonvirta, current professorship and department, reviewed August 12, 2026.
• Aalto University, "The digital world isn't a separate dimension in some virtual cloud", April 4, 2024 appointment profile and continuing Oxford role, reviewed August 12, 2026.
• Aalto University Research Portal, The Geopolitics of Cloud Computing (GEOCLOUD), project scope and 2025–2029 dates, reviewed August 12, 2026.
• Timur Ergen et al., "On Vili Lehdonvirta's Cloud Empires", Socio-Economic Review 22, no. 4, October 2024, pp. 2025–2038, DOI: 10.1093/ser/mwae023, reviewed August 12, 2026.
• Timo Seidl, review of Cloud Empires, Regulation & Governance, 2022, DOI: 10.1111/rego.12493, reviewed August 12, 2026.
• European Commission, DMA Gatekeepers Portal, current gatekeepers, designated core platform services, and Facebook Marketplace undesignation, reviewed August 12, 2026.
• European Commission, "Commission finds Apple and Meta in breach of the Digital Markets Act", April 23, 2025 non-compliance decisions and fines, reviewed August 12, 2026.
• European Commission, "Commission reaches preliminary position that Amazon's and Microsoft's market leading cloud services should be designated under the DMA", June 25, 2026 procedural status and stated cloud factors, reviewed August 12, 2026.
• European Commission, "DSA: Very large online platforms and search engines", threshold and additional systemic-risk, audit, researcher-access, advertising, and recommender duties, reviewed August 12, 2026.
• European Union, Regulation (EU) 2023/2854, the Data Act, official legal text, especially Chapters VI and VIII on switching and interoperability, reviewed August 12, 2026.
• European Commission, Data Act explained, applicability, cloud-contract, interface, export, functional-equivalence, and switching-charge timelines, reviewed August 12, 2026.
• European Union, Regulation (EU) 2024/1689, the Artificial Intelligence Act, Chapter V general-purpose AI duties, Article 53 open-source scope, systemic-risk obligations, and legacy-model deadline, reviewed August 12, 2026.
• European Commission, General-purpose AI obligations under the AI Act, documentation, open-source exemptions, copyright, training-content-summary, systemic-risk, incident-reporting, and cybersecurity duties, reviewed August 12, 2026.
• European Commission, Guidelines for providers of general-purpose AI models, application, enforcement, and legacy-model timelines and the guidelines' nonbinding status, reviewed August 12, 2026.
• European Commission, General-Purpose AI Code of Practice, voluntary compliance role and transparency, copyright, safety, and security chapters, reviewed August 12, 2026.
• European Union, Directive (EU) 2024/2831 on platform work, algorithmic-management protections and December 2, 2026 transposition deadline, reviewed August 12, 2026.
• UK Competition and Markets Authority, Cloud services market investigation, July 31, 2025 final outcome and recommendation, reviewed August 12, 2026.
• UK Competition and Markets Authority, "CMA announces package of actions on business software and cloud services", March 31, 2026 egress-fee and interoperability steps and continuing review, reviewed August 12, 2026.
• UK Competition and Markets Authority, Microsoft's business software ecosystem, open strategic-market-status investigation and indicative timetable, reviewed August 12, 2026.

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

• Amazon, Cloud Empires by Vili Lehdonvirta, affiliate listing, reviewed August 12, 2026.

Return to Blog · Return to Books