Blog · Review Essay · Last reviewed June 25, 2026

Empire of AI and the Mission That Became an Empire

Karen Hao's Empire of AI is a reported history of OpenAI and a broader argument about the political economy of artificial intelligence. Its most useful lesson is that a technology company can speak in the language of public benefit while building the material shape of an empire: data, compute, labor, energy, secrecy, and belief organized around a future only a few institutions claim authority to define.

The sharper definition is this: an AI empire is not only a company with market power. It is an institution whose mission language, infrastructure needs, platform defaults, safety claims, capital partners, and future forecasts start arranging public life around its own theory of intelligence.

The Book

Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI was published by Penguin Press on May 20, 2025. Penguin Random House lists the hardcover at 496 pages and the paperback at 496 pages with a May 19, 2026 publication date. The publisher describes the book as an account of OpenAI, ChatGPT, Sam Altman, the AI arms race, and the global costs of data, compute, labor, energy, and water. Kirkus gives the same hardcover publication date, ISBN, page count, and publisher, and frames the book as a pointed account of AI regulation and harm reduction.

Hao is not approaching the subject from outside the field. Penguin's author note identifies her as an award-winning AI journalist, formerly a Wall Street Journal reporter and MIT Technology Review senior editor for AI. SIAM News places the book in continuity with Hao's 2020 MIT Technology Review investigation of OpenAI, which examined how competitive pressure eroded the organization's early promise of openness.

The book is partly a company history: the nonprofit origin story, the rise of Sam Altman, the 2019 capped-profit structure, the Microsoft relationship, ChatGPT, secrecy around frontier models, and the November 2023 board crisis. But the stronger reading is not merely biographical. Hao uses OpenAI as the cockpit view of a larger system: a small set of firms racing to accumulate capital, chips, data, electricity, political access, and cultural permission under the sign of AGI.

Current Context

As of June 25, 2026, the book should be read after OpenAI's own corporate story changed again. OpenAI says its October 28, 2025 structure made the nonprofit the OpenAI Foundation and the for-profit arm OpenAI Group PBC, a public benefit corporation controlled by the Foundation. Microsoft says it holds an investment in OpenAI Group PBC valued at about $135 billion, representing roughly 27 percent on an as-converted diluted basis. California and Delaware officials also issued formal review records around the recapitalization, making OpenAI's structure a public-law governance object rather than only a private corporate design.

That current context strengthens Hao's thesis without proving every part of it. The institution now has a clearer public-benefit corporate form, an official nonprofit-control story, official safety frameworks, and regulator non-objection records. It also has deeper capital needs, platform reach, compute demand, commercial partners, and public dependence. The governance question is no longer whether OpenAI began with a broad-benefit mission. It is whether that mission can still constrain an organization whose operations require empire-scale inputs.

The infrastructure context has sharpened too. The International Energy Agency estimates data centers used about 415 TWh of electricity in 2024 and projects about 945 TWh by 2030 in its base case; Lawrence Berkeley National Laboratory's U.S. report estimated 176 TWh in 2023 and a possible 325 to 580 TWh by 2028. Those figures do not describe OpenAI alone, and they do not make every AI use illegitimate. They show why frontier-model governance cannot stop at product safety. Compute, power, water, chips, and siting are part of the system's public footprint.

This page does not treat current AI systems as conscious, divine, or already AGI. It treats AGI language as a claim-making and capital-organizing vocabulary unless and until a particular technical claim is supported by public, reviewable evidence.

Mission as Governing Myth

The central tension in Empire of AI is the gap between mission language and institutional behavior. OpenAI began with a public-benefit story: advanced AI would be developed for humanity rather than captured by narrow commercial incentives. Hao's account follows how that story became less a constraint than a source of authorization.

This is the book's most important contribution to thinking about AI institutions. A mission can discipline an organization, but it can also sanctify expansion. If the institution believes it is racing on behalf of everyone, then ordinary checks can start to look like irresponsible delay. Openness can be traded for safety. Safety can be traded for speed. Public benefit can be routed through private infrastructure. The more sacred the end state becomes, the easier it is to rationalize the means.

A mission becomes a governing myth when it answers three questions before outsiders can ask them: who may build, who may know, and who may object. In a healthy institution, the mission creates duties, evidence thresholds, and refusal points. In an imperial institution, the mission turns size into virtue and criticism into failure to understand the future.

That dynamic is not unique to OpenAI. It is a pattern of technological politics. A future good is projected with such force that present costs become collateral: worker injury, local water stress, public opacity, weak consent, copyright conflict, energy demand, and the narrowing of democratic choice. The book asks readers to judge AI not by the promised destination alone, but by the institutions the promise is already building.

Extraction Behind the Interface

Hao's empire frame works because it pulls attention away from the chat window and toward the supply chain. Large AI systems are not weightless minds. They require data pipelines, annotation and moderation labor, electricity, water, chips, minerals, cloud contracts, model deployment, and legal arrangements that determine who benefits and who absorbs risk.

SIAM News summarizes the book's reporting on workers in Colombia and Kenya who annotated or filtered difficult material for AI systems under poor conditions, and on environmental burdens connected to mines and data centers in Chile, Uruguay, and elsewhere. Kirkus likewise notes Hao's attention to data centers, developing-world data gaps, low-wage AI labor, and the imbalance between booster promises and present harms.

This makes the book especially useful against a common interface illusion. The more fluent the system becomes, the easier it is to forget the human and material world underneath it. A model that answers instantly can make extraction feel like intelligence. Hao reverses the camera: before calling the interface magical, count the people and places made invisible by the magic trick.

The operational lesson is to ask for an extraction ledger. A frontier-system safety record should identify compute provider, model and system versions, data provenance, labeling and moderation vendors, evaluation scope, energy and water assumptions, cloud dependency, deployment sector, affected population, incident owner, and appeal or refusal path. Without that ledger, the system's most important dependencies remain outside public argument.

Secrecy and Concentrated Authority

The other major thread is secrecy. SIAM News contrasts OpenAI's original public commitment to publishing research with later model releases that withhold core details about architecture, training compute, dataset construction, and training methods. Hao's critique is not that every dangerous capability should be published carelessly. It is that secrecy shifts authority toward a small set of corporate actors while the public is asked to accept their claims about safety, capability, and necessity.

That matters because frontier AI is not only a product category. It is becoming infrastructure for education, search, work, software, creative production, public services, medicine, warfare, and personal companionship. When a closed institution supplies the cognitive layer for many other institutions, opacity becomes political power. It decides who can audit, contest, replicate, regulate, or refuse the system.

The November 2023 OpenAI board crisis becomes important for this reason. In Hao's telling, and in the Atlantic excerpt adapted from the book, the leadership struggle exposed how much of the AI future was being shaped by a tiny circle of executives, investors, scientists, employees, and platform partners. The drama was not gossip on the side of the real story. It was the governance story.

System cards, preparedness frameworks, and safety hubs are useful evidence artifacts, but they do not by themselves settle the accountability problem. A company-authored safety document can disclose tests, mitigations, and thresholds. It can also leave out training data, failed evaluations, commercial pressure, board conflict, labor conditions, energy cost, or downstream dependency. The stronger standard is a safety case that states the deployment boundary, the evidence, the residual risk, who can stop release, and what the public can inspect.

The AI-Age Reading

The AI age has a distinctive belief problem. It does not only produce tools. It produces grand narratives about intelligence, progress, salvation, national competition, existential risk, abundance, and inevitability. Those narratives organize money and attention. They tell governments what counts as strategic. They tell workers what kind of displacement is acceptable. They tell users that adoption is participation in the future.

Empire of AI is strongest when it treats AGI as an ideology as much as a technical objective. The term can name a research ambition, but it can also serve as a flexible permission structure. If AGI is near, then scale becomes urgent. If scale is urgent, then capital concentration looks necessary. If concentration is necessary, then secrecy, energy demand, labor outsourcing, and weak public oversight can be reframed as temporary costs on the way to universal benefit.

The danger is not simply that the prediction is wrong. The danger is that the prediction changes the world before it is tested. Data centers get built. Labor markets are reorganized. Artists and writers are scraped or substituted. Schools and companies restructure around model access. States subsidize infrastructure. A speculative future becomes a present administrative fact.

Governance and Safety

The governance lesson is to treat mission claims as controls only when they create enforceable friction. A broad-benefit charter, a public benefit corporation, a safety framework, or a system card can help only if it specifies who has authority to delay or stop release, what evidence is required, how conflicts are handled, how incidents are reported, and what affected publics can inspect or contest.

OpenAI's April 2025 Preparedness Framework is relevant because it shows the company trying to formalize risk categories, thresholds, mitigations, and deployment decisions for frontier capabilities. It should be read as self-governance evidence, not as independent certification. NIST's AI Risk Management Framework and Generative AI Profile point toward broader lifecycle discipline: govern, map, measure, manage, document, monitor, and update. Under the EU AI Act, providers of general-purpose AI models with systemic risk face obligations around model evaluation, systemic-risk assessment and mitigation, serious-incident reporting, and cybersecurity where the law applies. These are the kinds of records that turn safety language into public accountability.

For an institution with OpenAI's platform reach, a credible governance package would include at least six records: corporate-control record, safety-case record, extraction ledger, deployment-impact record, incident record, and exit record. The corporate-control record explains board powers, investor rights, conflicts, regulator commitments, and nonprofit authority. The safety-case record explains release criteria and residual risk. The extraction ledger covers compute, data, labor, energy, and vendors. The deployment-impact record covers affected users and institutions. The incident record covers failures and corrective action. The exit record explains how customers, users, workers, and public agencies can migrate, appeal, delete, refuse, or roll back.

The safety implication is concrete: do not let awe at the interface substitute for inspectability of the institution. A model can be useful and still create unacceptable dependency if a school, agency, newsroom, hospital, or employer cannot audit the system, move away from it, preserve records, or contest harmful outputs. Platform lock-in is a safety issue when the platform becomes a knowledge, work, coding, media, or agent layer.

Infrastructure safety is part of the same package. Data-center expansion should be reviewed through load, water, backup generation, cost allocation, local permitting, community benefit, security, and stranded-cost risk. Compute governance should not only restrict dangerous scale; it should also prevent the largest firms from becoming the only entities able to build, evaluate, or challenge advanced systems.

Where the Book Needs Care

The book should not be read as proof that all AI research is inherently imperial, that all frontier labs are identical, or that every technical advance is a scam. That flattening would weaken Hao's sharper point. The question is institutional form: who controls the system, what resources it consumes, what stories legitimate it, what publics can inspect it, and who has standing when the costs arrive.

It also needs to be read with attention to genre. Empire of AI is reported, critical, and argumentative. It foregrounds harms and power. Readers who want a technical account of model architecture, benchmark progress, or beneficial deployments will need companion sources. That is not a defect so much as a boundary. Hao is writing the political history of an industry that often prefers to be evaluated only by demos and future promises.

The fairest use of the book is not as a verdict that AI should stop. It is as a test for whether AI institutions deserve public trust. Do they disclose enough for democratic scrutiny? Do they protect workers and communities rather than externalizing harm? Do they permit refusal and appeal? Do they separate safety claims from market positioning? Do they let the public inspect the mission, or only applaud it?

What This Changes

The book belongs in this catalog because it shows how a belief system can become infrastructure. The dream of beneficial superintelligence does not remain in manifestos, blog posts, board decks, and launch-stage rhetoric. It becomes contracts, data centers, labor markets, secrecy norms, procurement dependencies, and legal arrangements. A story about intelligence becomes a system for arranging the world around those who claim to be building it.

That is the recurring pattern to watch. The interface gives the user a helpful answer; the institution behind the interface asks for more data, more compute, more energy, more trust, more legal room, and more patience. The public is invited to judge the answer while the machinery that produced it recedes from view.

Hao's best warning is therefore practical. Do not evaluate AI only at the point of use. Follow the system backward into labor and infrastructure, sideways into governance and ownership, and forward into the futures it uses to excuse the present. The question is not whether the machine sounds intelligent. The question is what kind of world has to be built so that it can keep speaking.

Source Discipline

This review separates five source types. Penguin Random House and Kirkus support book metadata. Hao's book, Atlantic excerpt, SIAM News review, and PBS interview support claims about the book's argument and reception. OpenAI and Microsoft sources support what the companies officially claimed about structure, partnership, products, and safety process. California and Delaware official records support the fact of regulator review and non-objection, not a conclusion that the structure is substantively safe. IEA, DOE, LBNL, NIST, and EU AI Act sources support infrastructure and governance context.

Do not collapse those source types. A company announcement is evidence of the company's position, not independent proof that its governance works. A regulator non-objection is not a public-interest audit of every downstream risk. A safety framework is not a deployment-specific safety case. A data-center electricity estimate describes sector load, not the footprint of one model unless the boundary is specified.

AGI language also needs strict source discipline. This page uses AGI as an institutional forecast and research ambition, not as evidence that current systems have achieved general intelligence or deserve political authority. The stronger claim is about power: even an uncertain forecast can reorganize investment, infrastructure, labor, law, and public expectation.

Sources

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

Buy on Amazon Browse Books

Return to Blog · Return to Books