Empire of AI Review: OpenAI, Capital, Compute, and the Political Economy of Artificial Intelligence

Books / AI, Power & Political Economy

Karen Hao’s most important move is to treat frontier AI not as a sequence of clever models but as an institution-building project that concentrates capital, compute, data, labor, and agenda-setting power.

Empire of AI is a history of OpenAI, but Karen Hao is after a larger target. She argues that the frontier-AI industry should be understood as a political economy: a system that mobilizes enormous capital, specialized chips, data, energy, water, global labor, and a narrative of technological inevitability. The models are the visible product. The institutional machinery that makes them possible is the deeper subject.

That framing is valuable because AI strategy is often discussed as if intelligence were emerging from software alone. In reality, frontier systems are industrial projects. Training and serving large models require data centers, accelerators, networking equipment, electricity, cooling, land, supply agreements, and financial commitments at a scale that favors a small number of organizations. Hao’s book makes those dependencies part of the story rather than background infrastructure.

The book is also polemical. “Empire” is not a neutral description, and readers should test the analogy rather than inherit it. OpenAI and other AI companies argue that advanced models can create broad economic and scientific benefits; the industry has also developed safety programs, usage policies, and governance mechanisms. A serious review therefore asks where Hao’s evidence is strongest, where the metaphor clarifies incentives, and where competition, regulation, or technological change may complicate the picture.

OpenAI’s mission-capital tension is the book’s central corporate story

OpenAI began as a nonprofit research organization with a public-interest mission and later created structures capable of raising far more capital. That transition is not a footnote. Frontier AI became increasingly dependent on compute, and compute required money on a scale that conventional nonprofit funding could not easily supply. The organization therefore faced a strategic conflict: preserve governance purity or access the resources needed to compete at the frontier.

This is a recognizable organizational problem. Mission-driven companies often believe capital is merely fuel, but financing can alter time horizons, risk tolerance, product pressure, and stakeholder power. The important question is not whether investors are good or bad. It is whether governance explicitly accounts for how dependency on capital changes the feasible set of decisions. Hao’s narrative is strongest when it shows that structure and mission cannot be analyzed separately.

Compute is a strategic input, not just an engineering expense

The rise of large-scale machine learning turned compute into a source of bargaining power. Access to accelerators, data-center capacity, networking, and cloud infrastructure can determine which labs are able to train frontier models and how quickly they can iterate. This creates an industrial concentration mechanism: the cost of entering the frontier is not simply hiring brilliant researchers.

For investors, compute should be analyzed the way a manufacturer analyzes a scarce process node. Who controls supply? How long is capacity committed? What is the depreciation risk if model architecture changes? How much demand comes from training versus inference? Can efficiency gains offset growth in usage? The AI industry’s software margins are increasingly connected to very physical capital-allocation questions.

The “empire” analogy is about resource extraction and agenda setting

Hao links AI development to global flows of labor, data, minerals, water, and energy, arguing that benefits and costs are unevenly distributed. The metaphor directs attention toward people and places that are easy to omit from a product demo: data workers, content moderators, communities near infrastructure, and countries whose material resources support hardware supply chains.

The analogy is most useful when it forces full-cost accounting. A model’s unit economics should not end at API revenue minus cloud expense. A social accounting includes energy infrastructure, water use, labor conditions, data rights, and public subsidies. Not every cost is properly charged to the AI firm, but ignoring externalities because they sit outside the invoice produces a strategically incomplete picture.

Narratives of inevitability can become a competitive weapon

Frontier AI leaders frequently describe development as a race. If transformative AI is inevitable, slowing down can sound equivalent to surrendering leadership to a competitor or geopolitical rival. That framing changes governance because it makes caution appear not merely costly but futile. Hao is skeptical of this inevitability narrative, and that skepticism deserves attention even from readers enthusiastic about the technology.

In strategy, declaring an outcome inevitable can influence the outcome by accelerating investment and discouraging alternatives. The important managerial question is: which parts of the roadmap are technical necessity and which are socially coordinated expectations? Capital markets, talent migration, procurement, and regulation all respond to narratives. AI leaders therefore exercise power not only through models but through the futures they persuade others to treat as unavoidable.

Scale creates capability, but it can also narrow the research agenda

Large labs can run experiments that smaller organizations cannot afford. That can produce genuine scientific progress. Yet when frontier performance becomes strongly associated with scaling data and compute, research incentives may converge around the approaches that well-capitalized institutions are positioned to pursue. Alternative architectures, smaller models, or slower lines of inquiry can become less prestigious even before the evidence is conclusive.

This is a portfolio problem. A healthy innovation ecosystem benefits from multiple technical hypotheses and organizational forms. If capital concentration causes the entire field to optimize the same benchmark using similar methods, systemic model risk increases. Investors and policymakers should care about diversity of approach for the same reason portfolio managers care about correlated positions.

Labor remains inside the machine

AI marketing often emphasizes automation, but modern model development has depended on human labor for annotation, ranking, evaluation, safety classification, data curation, and content moderation. Some of that work has been performed through global outsourcing networks. Hao’s reporting makes these workers part of the production function rather than invisible support.

The strategic implication is broader than labor ethics. Human-feedback quality affects model behavior. Turnover, low context, psychological burden, unclear instructions, and vendor incentives can become product-quality variables. Companies that treat this work as an interchangeable low-cost service may be underestimating its role in the reliability of the final system.

Energy and water turn AI into local infrastructure politics

As data-center demand rises, AI firms increasingly interact with utilities, grid planners, local governments, and communities. The International Energy Agency has highlighted data centers as a growing source of electricity demand. The exact share attributable to AI is uncertain and technology efficiency can change rapidly, but the direction matters: model economics are becoming entangled with power-system economics.

That creates constraints unfamiliar to pure software strategy. A promising site can be limited by interconnection queues, transformer supply, water availability, permitting, or community resistance. The most successful AI organizations may need infrastructure-development competence alongside research talent. The sector is becoming simultaneously more digital and more industrial.

Governance instability is not a side story

OpenAI’s 2023 leadership crisis exposed the difficulty of governing an organization that combined a nonprofit mission, a capped-profit history, major commercial partnerships, valuable employees, and a rapidly scaling consumer product. The episode demonstrated that formal authority and practical power can diverge sharply. A board can possess legal authority while employees, investors, partners, and customers hold forms of leverage that make certain decisions hard to sustain.

This is a case study in stakeholder architecture. Governance should be evaluated not only by who has a vote but by who can credibly impose costs after a vote. If a structure requires all stakeholders to behave cooperatively during the moment when interests diverge, the structure has not resolved the conflict; it has postponed it.

Hao is strongest on institutions and more debatable on analogy

The reporting base is extensive, and the book benefits from Hao’s years covering AI. The institutional details—organizational evolution, labor, infrastructure, financing, and internal ideology—give readers a richer model than a founder-centric history. The “empire” frame helps connect these parts by emphasizing concentration and externalities.

The risk is that empire becomes too elastic. Powerful firms are not states; cloud contracts are not territorial rule; users can switch products more easily than subjects can exit political authority. Analogies are analytical tools, not conclusions. Readers should keep the factual reporting even where they reject the metaphor, because the underlying questions about concentration, dependency, and accountability remain important.

The investor lesson: AI economics must be read below gross margin

AI companies can generate software-like revenue while carrying infrastructure-like capital intensity through partners, long-term contracts, or direct investment. Gross margin alone may obscure the durability of compute costs, model-training cycles, customer concentration, and the bargaining power of chip and cloud suppliers. Frontier competition can also create a treadmill in which today’s expensive capability becomes tomorrow’s baseline feature.

A stronger diligence framework separates model advantage, distribution advantage, data advantage, compute access, and organizational learning. If one layer commoditizes, where does pricing power remain? Hao’s institutional perspective is useful because it discourages treating “AI” as one moat. The stack contains different assets with different half-lives.

The operator lesson: externalities eventually become constraints

Companies often treat externalities as public-relations issues until they become regulatory, contractual, or infrastructure constraints. Data provenance becomes litigation. Energy demand becomes interconnection delay. labor practices become vendor risk. Safety failures become enterprise-customer objections. The timing is uncertain, but the pattern is familiar across industries.

An AI operator should therefore map externalities before regulation maps them on the company’s behalf. Which costs are not priced today? Which affected stakeholders lack a contract but can influence permits, policy, or reputation? Which assumptions depend on public tolerance? This is not anti-growth. It is a way to preserve growth by identifying the constraints that scale will make visible.

The strategic question after Empire of AI is who gets to set the objective function

A model is optimized against chosen objectives; an industry is too. Capital providers reward some outcomes, governments reward others, researchers value benchmarks, users reward convenience, and communities care about local effects. No single metric captures the system. Hao’s deepest challenge is that a small set of institutions may have disproportionate influence over which objectives receive resources.

For leaders, the response should be procedural as much as moral. Make tradeoffs legible. Separate safety claims from marketing. Publish evidence where feasible. Build channels for external challenge. Track resource intensity and labor dependencies. The goal is not to eliminate power—scale always creates power—but to make power more accountable to evidence and consequences.

A practical metric: compute-adjusted learning

AI labs often celebrate benchmark gains without disclosing how much additional compute, data, or engineering effort produced them. Track capability improvement relative to resource growth. If each increment requires disproportionate capital, the business may be scaling revenue and strategic dependency at the same time. Efficiency is not only a technical virtue; it is a bargaining-power variable.

Why distribution can matter more than the best benchmark

A frontier model can lead on a benchmark and still lose economically if another company owns the customer relationship, developer ecosystem, operating system, or enterprise procurement channel. Hao’s institutional framing helps separate technical leadership from market power. Investors should ask which layer captures the surplus when model quality converges.

Safety governance needs independent clocks

Commercial teams operate against launch windows and competitive milestones. Safety work often needs time to investigate low-frequency, high-severity behavior. If both functions share one deadline, the faster clock wins. Organizations can reduce this bias by defining safety gates that cannot be compressed solely because a competitor announced first.

A frontier-AI power map for investors and operators

• Capital dependence: Map who finances training, infrastructure, and losses, and what strategic leverage those relationships create.
• Compute concentration: Identify cloud, chip, networking, and data-center dependencies that can constrain growth or bargaining power.
• Data provenance: Track where training and evaluation data come from and which legal or reputational assumptions support their use.
• Human labor: Include annotation, moderation, red-teaming, and evaluation work in the production model.
• Energy exposure: Model power availability, interconnection, price volatility, and local infrastructure constraints.
• Governance authority: Distinguish formal control from practical leverage held by employees, partners, investors, and customers.
• Safety incentives: Ask whether release pressure and safety review operate on compatible timelines.
• Narrative power: Monitor claims of inevitability or emergency that can compress strategic deliberation.
• Externalities: Identify costs likely to migrate from society into regulation, litigation, procurement standards, or infrastructure pricing.
• Moat decomposition: Separate model quality from distribution, brand, data, compute access, developer ecosystem, and organizational learning.

Questions readers are likely to ask

What is Empire of AI about?

Karen Hao uses OpenAI and the broader frontier-AI industry to examine the concentration of capital, compute, labor, data, infrastructure, and decision-making power behind modern artificial intelligence.

Is the book anti-AI?

It is strongly critical of industry power structures, but its central subject is how AI is developed and governed rather than an argument that the technology has no value.

Why does Hao use the word “empire”?

The analogy highlights concentrated power, resource extraction, global labor, and uneven distribution of costs and benefits. Readers should treat it as a lens to test, not a literal equivalence.

Who should read the book?

AI founders, investors, policy professionals, enterprise buyers, infrastructure leaders, and readers who want an institutional rather than purely technical history of the AI boom.

Verdict: essential institutional context for an industry that prefers technical narratives

Empire of AI is valuable because it makes the frontier look physical, financial, and political. Models do not float above the economy. They are produced by organizations with governance structures, suppliers, labor systems, infrastructure demands, investors, and stories about the future. Hao forces those layers into the same frame.

Readers can dispute the empire analogy and still benefit from the reporting. The durable strategic lesson is that scale changes the nature of a technology company. Once a firm becomes dependent on scarce infrastructure and capable of shaping markets, policy, and public expectations, technical excellence is no longer enough. Institutional design becomes part of the product.

A decision lab for applying the book

Decision lab 1: Empire of AI tests capital dependence against a current operating choice. Empire of AI starts with evidence already visible. Empire of AI then names evidence still missing. Empire of AI assigns one owner to downside risk. Empire of AI sets a deadline for uncertainty reduction. Empire of AI chooses one reversible experiment next. Empire of AI uses that experiment to create information cheaply. Empire of AI avoids committing to the full-risk path too early. Empire of AI asks which observation would falsify the current thesis. Empire of AI identifies who can surface unwelcome evidence. Empire of AI separates factual change from status pressure. Empire of AI separates incentive change from political pressure. Empire of AI records the interval from signal to decision. Empire of AI records the interval from decision to measurable result. Empire of AI compares those intervals with the speed of the underlying risk. Empire of AI turns capital dependence into an observable management practice. Empire of AI is useful only when this practice changes a real decision. Empire of AI should leave a written trace: evidence, Empire of AI, owner, deadline, reversible move, Empire of AI, falsifier, result. Empire of AI can then be revisited after the outcome of Empire of AI to distinguish luck from judgment about Empire of AI and learning from narrative.

Decision lab 2: Empire of AI tests compute concentration against a current operating choice. Empire of AI starts with evidence already visible. Empire of AI then names evidence still missing. Empire of AI assigns one owner to downside risk. Empire of AI sets a deadline for uncertainty reduction. Empire of AI chooses one reversible experiment next. Empire of AI uses that experiment to create information cheaply. Empire of AI avoids committing to the full-risk path too early. Empire of AI asks which observation would falsify the current thesis. Empire of AI identifies who can surface unwelcome evidence. Empire of AI separates factual change from status pressure. Empire of AI separates incentive change from political pressure. Empire of AI records the interval from signal to decision. Empire of AI records the interval from decision to measurable result. Empire of AI compares those intervals with the speed of the underlying risk. Empire of AI turns compute concentration into an observable management practice. Empire of AI is useful only when this practice changes a real decision. Empire of AI should leave a written trace: evidence, Empire of AI, owner, deadline, reversible move, Empire of AI, falsifier, result. Empire of AI can then be revisited after the outcome of Empire of AI to distinguish luck from judgment about Empire of AI and learning from narrative.

Decision lab 3: Empire of AI tests data provenance against a current operating choice. Empire of AI starts with evidence already visible. Empire of AI then names evidence still missing. Empire of AI assigns one owner to downside risk. Empire of AI sets a deadline for uncertainty reduction. Empire of AI chooses one reversible experiment next. Empire of AI uses that experiment to create information cheaply. Empire of AI avoids committing to the full-risk path too early. Empire of AI asks which observation would falsify the current thesis. Empire of AI identifies who can surface unwelcome evidence. Empire of AI separates factual change from status pressure. Empire of AI separates incentive change from political pressure. Empire of AI records the interval from signal to decision. Empire of AI records the interval from decision to measurable result. Empire of AI compares those intervals with the speed of the underlying risk. Empire of AI turns data provenance into an observable management practice. Empire of AI is useful only when this practice changes a real decision. Empire of AI should leave a written trace: evidence, Empire of AI, owner, deadline, reversible move, Empire of AI, falsifier, result. Empire of AI can then be revisited after the outcome of Empire of AI to distinguish luck from judgment about Empire of AI and learning from narrative.

Decision lab 4: Empire of AI tests human labor against a current operating choice. Empire of AI starts with evidence already visible. Empire of AI then names evidence still missing. Empire of AI assigns one owner to downside risk. Empire of AI sets a deadline for uncertainty reduction. Empire of AI chooses one reversible experiment next. Empire of AI uses that experiment to create information cheaply. Empire of AI avoids committing to the full-risk path too early. Empire of AI asks which observation would falsify the current thesis. Empire of AI identifies who can surface unwelcome evidence. Empire of AI separates factual change from status pressure. Empire of AI separates incentive change from political pressure. Empire of AI records the interval from signal to decision. Empire of AI records the interval from decision to measurable result. Empire of AI compares those intervals with the speed of the underlying risk. Empire of AI turns human labor into an observable management practice. Empire of AI is useful only when this practice changes a real decision. Empire of AI should leave a written trace: evidence, Empire of AI, owner, deadline, reversible move, Empire of AI, falsifier, result. Empire of AI can then be revisited after the outcome of Empire of AI to distinguish luck from judgment about Empire of AI and learning from narrative.

Sources and further reading

• Penguin Random House — Empire of AI
• Science News — Empire of AI review
• MIT Technology Review — Karen Hao archive
• The Atlantic — Karen Hao author page
• Stanford AI Index 2026
• International Energy Agency — Energy and AI
• OpenAI — Our structure
• Financial Times — 2025 Business Book longlist

Editorial review prepared for Millionaire Lesson. Book-specific claims are distinguished from broader analytical inference.