Who Pays for AI's Power? OpenAI's Community Compact and Fei-Fei Li's Human-Centered AI

AI infrastructure is often discussed as if its most important questions begin and end with chips, models, and benchmark scores. OpenAI’s November 5 announcement puts another set of questions in view: who gets a say when a compute campus is planned, who pays for the electricity and grid connections, and how will the operator show that it kept its promises? The announcement cannot settle those questions by itself, but it makes them harder to treat as side issues.

That shift connects naturally to The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI by Dr. Fei-Fei Li. The book is a first-person account of a scientist’s life and work, published by Flatiron Books. Its central relevance here is not a prediction about a particular data center. It is a reminder that AI is a human undertaking, shaped by choices about what we build, whose needs count, and how institutions explain the consequences. The publisher’s page for the exact hardcover edition lists ISBN 9781250897930.

What did OpenAI announce about the communities around AI campuses?

In its November 5, 2026 post, OpenAI announced a National Community Compact and a Compute Campus Policy Blueprint. The company says that every future compute campus it owns and operates will have a local compact. It says residents will be engaged early enough to shape project plans, receive a local point of contact, and be able to review a public annual report on performance.

The company also describes several commitments on costs and resources. It says it will pay the full cost of electricity and infrastructure needed to serve its campuses, provide financial assurances before major spending begins, develop water plans, prioritize closed-loop or low-water cooling, and publish site-level water and electricity use. The announcement also describes publishing construction plans and independent noise studies before operations begin and annually afterward. OpenAI says it will not use nondisclosure agreements with public officials, apart from narrow protections for legitimate confidential matters. Those are claims about the company’s stated policy, not evidence that every future project has already met those standards.

That distinction matters. A policy document is an invitation to evaluate future conduct, not a substitute for it. A community can still ask how the commitments will be enforced, who verifies the numbers, what happens if a project changes, and whether residents can influence decisions before contracts and site plans are effectively settled. OpenAI’s post says the blueprint proposes policies and the compact sets expectations for campuses the company owns and operates. It should not be read as a blanket guarantee about every data center built by every company.

Why is the electricity question becoming impossible to ignore?

Data centers are not a niche part of the energy system anymore. The International Energy Agency’s Energy and AI analysis estimates that data centers used about 415 terawatt-hours of electricity worldwide in 2024, around 1.5% of global electricity consumption. The IEA says this demand grew at an average rate of 12% a year over the preceding five years. In its base case, it projects data-center electricity use to reach about 945 TWh in 2030, just under 3% of global consumption.

Forecasts are not certainties, and the IEA’s number is a scenario rather than a promise about what will happen. But even a projection carries a practical signal: electricity supply, connections, cooling, and land-use decisions will increasingly shape where computing can be built and how quickly. AI is one driver among several, and the exact amount attributable to AI depends on how workloads and equipment evolve. The useful point for local planning is not that one forecast is guaranteed. It is that a large, fast-changing load deserves explicit assumptions and transparent updates.

The U.S. picture shows why communities may pay attention before a campus opens. The Department of Energy’s Electricity Demand Growth Resource Hub, citing Lawrence Berkeley National Laboratory’s 2024 U.S. Data Center Energy Usage Report, says data centers consumed about 4.4% of total U.S. electricity in 2023. It reports the LBNL estimate that data centers could account for approximately 6.7% to 12% of U.S. electricity use by 2028. The range is wide because the forecast depends on uncertain growth and efficiency assumptions. It is a warning about planning uncertainty, not a precise prediction for any one town or facility.

That is why an energy commitment should be tested at the local level. A national percentage cannot tell a neighbor whether a particular substation upgrade is being paid for by the project, whether a new line is needed, how the campus will use water during a hot period, or whether noise will be measured after construction. Those questions require site-specific documents and an accountable process. The better the information, the less room there is for residents and operators to argue from incompatible assumptions.

How does Fei-Fei Li’s book help frame the human side?

Macmillan describes The Worlds I See as a memoir and account of AI’s development from the perspective of a scientist, with a particular emphasis on the human story behind the technology. That makes it a useful companion to a discussion of infrastructure, even though the book is not a handbook on data-center permitting. Infrastructure can make AI sound abstract: megawatts, servers, capital investment, cooling systems. A human-centered lens restores the people who live near projects, the workers who build and maintain them, and the users who rely on the tools those campuses support.

Fei-Fei Li’s book also offers a counterweight to a common mistake in technology debates: treating technical capability as if it automatically determines social value. A model may become more capable while organizations still have to decide how to use it, what risks to accept, and who should share in the gains. The community compact raises a parallel question about physical infrastructure: even if more computing capacity can support useful services, what would make the way it is built fair and trustworthy?

For a reader choosing the book, The Worlds I See is most relevant as a personal account of scientific work and the people behind technological change. The publisher identifies Dr. Fei-Fei Li as the author and Flatiron Books as the imprint on the hardcover edition. Read it alongside current reporting and policy documents rather than asking one memoir to answer every question about the energy system.

What does trustworthy infrastructure planning look like in practice?

OpenAI’s announcement offers a useful checklist for readers, local officials, and businesses evaluating AI infrastructure plans. The items below translate the company’s published commitments into questions that can be checked against a project record:

• Was community input early enough to matter? Look for evidence that feedback could change a site plan, construction schedule, or mitigation measure, rather than arriving only after the core decisions were made.
• Are the costs clearly assigned? Ask which party pays for electricity, substations, transmission upgrades, water infrastructure, and other services. A promise to pay should be specific enough to distinguish project costs from costs borne by other customers.
• Can the public inspect the assumptions? A credible plan should identify expected power demand, water use, cooling approach, and the method used to calculate them. Where forecasts change, the updated assumptions should be visible.
• Is there a local point of contact and a path to resolution? Residents need a way to raise concerns during construction and operation, along with a process for receiving an answer and tracking follow-up.
• Are measurements independent and repeated? Noise, electricity, and water should not be treated as one-time disclosures. A recurring report can reveal whether actual operations match initial estimates.

This checklist is not a substitute for legal review or local permitting. It is a way to make broad promises concrete enough to discuss. A well-run process does not mean every resident will agree with every project. It means the key trade-offs are visible, the people affected can participate before it is too late, and the commitments can be checked after construction.

Does AI adoption tell us whether the benefits are being shared?

Adoption statistics explain why companies and governments are building quickly, but they do not answer whether the gains are reaching workers, customers, or the places hosting infrastructure. Stanford HAI’s 2025 AI Index economy chapter reports that 78% of surveyed organizations said they used AI in at least one business function in 2024, compared with 55% in 2023. The same chapter says 71% reported using generative AI in at least one business function in 2024, up from 33% in 2023.

Those figures measure organizational uptake, not the quality of implementation or the distribution of benefits. The chapter also notes that most companies reporting financial effects from AI within a business function estimated those benefits at low levels. That is a useful reminder for executives: adoption is an input, not a finish line. An organization may deploy AI widely while still needing to redesign workflows, train employees, test outputs, and decide how productivity improvements will be shared.

The same principle applies to infrastructure. A data center can be framed as an investment in a future technology economy, but that framing should not make the local costs invisible. If operators want broad support, they need to show how the project handles power, water, noise, emergency planning, and community concerns. If a company wants the public to believe that AI’s benefits are broadly shared, it must make the operational details available to the people who experience its footprint.

What questions should leaders ask before expanding AI?

For a business leader, the lesson is not simply to slow down. It is to connect an AI investment to a clear use case and a defensible operating plan. Before expanding AI workloads or buying services that depend on new compute capacity, leaders can ask:

• What decision or customer problem will this AI system improve, and how will we know?
• Which parts of the work still require human judgment, review, and responsibility?
• What data, energy, and infrastructure assumptions sit behind the promised benefit?
• What happens if efficiency gains or reliability fall short of the business case?
• How will employees, customers, and communities be informed about changes that affect them?

These questions connect model deployment to governance. They also help prevent executives from confusing a compelling demonstration with a durable operational result. If an AI system is useful, the case for it should survive careful measurement. If its benefits depend on hidden costs or unsupported assumptions, leaders need to know that before they scale it.

For individual readers, the same habit begins with curiosity. Read the announcement, find the policy or report it points to, and separate a company’s promise from independently verified performance. Then map who is affected, what evidence is available, and what is still unknown. That is a small but practical way to keep public conversations about AI grounded in people and consequences instead of slogans.

Key Takeaways

• OpenAI’s November 5, 2026 announcement sets out commitments for future compute campuses the company owns and operates; it is a policy statement to evaluate over time, not proof of completed outcomes.
• The IEA and Lawrence Berkeley National Laboratory describe rapidly growing data-center electricity demand, while emphasizing forecasts and ranges that carry uncertainty.
• Fei-Fei Li’s The Worlds I See provides a human-centered companion to the technical debate, but local planning still requires site-specific evidence.
• Trust depends on early participation, clear cost responsibility, public measurements, and a way to address problems after construction.
• AI adoption is rising, but adoption alone does not demonstrate that benefits are substantial, fairly shared, or worth every local cost.