Work and Power: Chips, data centers, energy, and concentration of power

This content reflects the field as of October 2026.

AI products are usually described as running "in the cloud," a phrase that suggests they exist nowhere in particular. In 2026 a growing number of American communities found that the cloud was a building someone proposed to put near them.

That building is a data center: a large structure filled with computers, with the electrical supply and cooling equipment needed to run them day and night. The ones built for AI are among the largest. They need land, a heavy and constant supply of electricity, and in many designs a steady supply of water for cooling.

In August 2026 an episode of The Ezra Klein Show, a podcast from the New York Times, was devoted to the opposition these projects have met (Klein 2026a). The episode's notes describe polling in which a large majority of Americans say they would oppose a data center near where they live. They report that New York's governor ordered a one-year halt to data center construction in the state, and that more than a hundred similar proposals were under consideration elsewhere. The opposition has come from members of both major political parties.

The guest was Jasmine Sun, a journalist who writes a newsletter about the AI industry and had been reporting from the Midwest. Her account, as the episode presents it, is that data centers have become a visible stand-in for how people feel AI is arriving in their lives (Klein 2026a). On this view the arguments at local hearings are partly about the buildings themselves and partly about the technology they serve.

The reasoning behind that account is easy to follow. Nobody can attend a public hearing about a language model. A resident can attend a hearing about a rezoning request, a power line, or a permit to draw water.

The people on each side of these disputes have interests. Developers and the AI companies that rent their space want to build quickly. Local officials weigh tax revenue and construction jobs against strain on the electrical grid and water supply. Residents weigh costs they will bear against benefits that may go elsewhere.

The dispute brings out something that is easy to miss when you type a question into a chatbot. AI depends on physical things, and each of them belongs to someone:

• specialized chips, designed and manufactured by a small number of companies
• data centers, owned mostly by a few very large technology firms
• electricity and water, supplied by utilities and regulated by governments
• land, controlled by local authorities

Once you see that list, questions follow that have nothing to do with how clever a model is. You can ask who owns each of these things, who decides where they go, who pays for the power, and which country makes the chips. People disagree about the answers, and about whether the present arrangement is a problem.

References

• Klein, Ezra, host. 2026a. "The A.I. Revolt Is Here." The Ezra Klein Show, podcast, New York Times, August 4, 2026.
  • Publisher: Apple Podcasts

This content reflects the field as of October 2026.

Introduction

Dozens of AI products compete for your attention, which gives the impression of a crowded field. Behind those products the field narrows quickly, and two competition regulators have mapped how.

This reading describes the links in the chain that produces an AI product, what regulators in the United Kingdom and the United States found about the ties between those links, and why the ties exist.

Four Links

A supply chain is the sequence of suppliers that a product depends on, from raw inputs to the finished product. For AI it has four main links.

1. Chips. Large AI models are trained and run on specialized chips. A few companies design them, such as Nvidia and AMD, and fewer still can manufacture the most advanced ones.
2. Cloud computing. A cloud provider is a company that owns data centers and rents out their computing power over the internet. The largest are Amazon, Microsoft, and Google.
3. Model developers. These companies build the large general-purpose models. They include OpenAI, Anthropic, Google, Meta, and others.
4. Products. Chatbots, writing aids, coding tools, and thousands of other applications are built on those models, often by other companies.

What moves along the chain is computing power: the capacity to run calculations, which for AI comes from specialized chips housed in data centers. Training a large model takes a great deal of it, and so does answering millions of users.

Some companies sit at more than one link. Google designs chips, runs a cloud, develops models, and sells products. Vertical integration is one company's ownership of several links of a supply chain.

What the UK Regulator Mapped

The Competition and Markets Authority is the United Kingdom's competition regulator. In April 2024 it published an update to its review of large general-purpose AI models, which it calls foundation models (Competition and Markets Authority 2024).

It looked at six firms: Google, Amazon, Microsoft, Meta, Apple, and the chip maker Nvidia. It reported "an interconnected web of over 90 partnerships" involving them (Competition and Markets Authority 2024, para. 43). A partnership here is a long-term agreement between two companies to share money, technology, or services, short of one buying the other.

The authority also noted how few companies can do without such ties. Some model developers had formed partnerships with major cloud providers to get computing power, while "only a handful of firms can rely on their own compute resources" (Competition and Markets Authority 2024, para. 32).

What the US Regulator Found Inside Three Partnerships

The Federal Trade Commission, a United States regulator, used its legal powers to obtain the terms of three partnerships. Its staff reported in January 2025 on Microsoft with OpenAI, Amazon with Anthropic, and Google with Anthropic. The publicly reported investments at the time were about 13.75 billion dollars, 8 billion dollars, and 2.55 billion dollars (Federal Trade Commission 2025).

The staff found common features.

The third row describes money that travels in a circle. A cloud provider invests in a developer, and the developer pays much of it back for computing.

The report states that it doesn't assess whether anyone has acted illegally (Federal Trade Commission 2025).

Why the Partnerships Exist

Both sides have ordinary commercial reasons. A model developer needs computing power on a scale that costs billions of dollars, and the chips are in short supply. The UK authority noted that the availability of AI chips "remains limited" (Competition and Markets Authority 2024, para. 32). A partnership gets a developer the capacity it couldn't otherwise buy.

A cloud provider gets a large, committed customer, a share in the developer's success, and leading models to offer to its own customers.

The UK authority acknowledged the benefits. It wrote that such partnerships "may be an essential ingredient for the success of independent developers" (Competition and Markets Authority 2024, para. 44). Its concerns about them are a separate matter from the map itself.

Conclusion

AI products rest on a chain of chips, cloud computing, and model developers, with a small number of firms at each link and some firms at several. As of the regulators' reports in 2024 and 2025, model developers and cloud providers were tied together by many partnerships that trade investment and computing power for ownership stakes, revenue, and spending commitments. The existence and terms of these partnerships are documented. Their effect on competition is disputed.

Key Terms

• Supply chain: The sequence of suppliers that a product depends on, from raw inputs to the finished product.
• Cloud provider: A company that owns data centers and rents out their computing power over the internet.
• Computing power: The capacity to run calculations, which for AI comes from specialized chips housed in data centers.
• Vertical integration: One company's ownership of several links of a supply chain.
• Partnership: A long-term agreement between two companies to share money, technology, or services, short of one buying the other.

References

• Competition and Markets Authority. 2024. AI Foundation Models: Update Paper. London: CMA, April 11, 2024.
  • Free: GOV.UK
• Federal Trade Commission. 2025. Partnerships Between Cloud Service Providers and AI Developers. FTC staff report, January 2025.
  • Free: Federal Trade Commission

This content reflects the field as of October 2026.

Introduction

You'll hear that AI is draining the electrical grid, and you'll hear that its energy use is a rounding error. Both claims can be supported with real figures, because they answer different questions.

This reading gives the main figures on data center electricity and water, separates what has been measured from what is projected, and explains how the same numbers support two framings.

What Has Been Measured

A data center is a building that houses large numbers of computers, with the power and cooling they need. Data centers run AI models, and they also run websites, video streaming, and business software. Figures for data centers therefore cover more than AI.

The International Energy Agency, an intergovernmental body that tracks the world's energy systems, published a report on energy and AI in 2025. It measures electricity in the terawatt-hour, a unit of electrical energy equal to a billion kilowatt-hours, used for the consumption of countries and industries.

The agency estimates that data centers consumed about 415 terawatt-hours in 2024, around 1.5 percent of the world's electricity. That consumption has grown about 12 percent a year since 2017, more than four times as fast as electricity use overall (International Energy Agency 2025).

Even this "measured" figure is an estimate. Companies don't publish complete data, so the agency builds its total from the information that is available.

What Is Projected

A projection is an estimate of a future quantity, calculated from present data and stated assumptions. The agency projects that data center consumption will "more than double to around 945 TWh by 2030," which it notes is slightly more than Japan's total electricity use today (International Energy Agency 2025).

That number comes from the agency's central scenario. It publishes others, with faster and slower growth, and by 2035 they range from 700 to 1,700 terawatt-hours. The outcome depends on how much AI is used, how efficient chips become, and how quickly power plants and grid connections can be built.

Where the Demand Falls

Electricity demand is the amount of electricity that users draw from the grid. Data center demand is concentrated in a few places. The United States accounted for 45 percent of the 2024 total, China for 25 percent, and Europe for 15 percent (International Energy Agency 2025).

Within countries it is more concentrated still. The agency reports that a typical AI-focused data center uses as much electricity as 100,000 households, and that the largest ones under construction will use 20 times as much. In the United States, it projects that data centers will account for nearly half of all growth in electricity demand between now and 2030.

The 2026 AI Index, an annual report from Stanford University, gives another measure. It puts the power capacity of the world's AI data centers at 29.6 gigawatts, which it compares with the peak demand of New York State (Stanford HAI 2026). Capacity is the most power that can be drawn at one moment, where terawatt-hours measure energy used over a year.

Water

Water use is the water a data center consumes, mainly for cooling its equipment. Chips turn nearly all the electricity they use into heat, and many cooling systems remove that heat by evaporating water.

Water figures are weaker than electricity figures. Companies report little, and use varies with climate and cooling design. The 2026 AI Index includes an estimate of the yearly water used in running one widely used model, which it compares with the drinking-water needs of a large population (Stanford HAI 2026). It is an inferred estimate, built from assumptions about where and how the model runs.

Water matters most locally. A quantity that is trivial for a country can be significant for a town in a dry region.

Two Framings of the Same Figures

Neither framing misstates the numbers. They differ in what they treat as the relevant unit: the planet or the county. Companies that build data centers benefit when the global framing prevails, and opponents of a particular site benefit from the local one.

Two further points are disputed. One is whether more efficient chips will slow the growth. The other is who pays for new power plants and lines: the companies, or all customers of the local utility.

Conclusion

As of the 2025 report, data centers used an estimated 1.5 percent of the world's electricity, a share that is growing fast and is projected to more than double by 2030 in the central scenario. The demand is concentrated in a few countries and in particular localities, so local effects are far larger than the global share suggests. Present consumption is an estimate from incomplete data, the 2030 figure is a projection, and water figures are the least certain of all.

Key Terms

• Data center: A building that houses large numbers of computers, with the power and cooling they need.
• Terawatt-hour: A unit of electrical energy equal to a billion kilowatt-hours, used for the consumption of countries and industries.
• Projection: An estimate of a future quantity, calculated from present data and stated assumptions.
• Electricity demand: The amount of electricity that users draw from the grid.
• Water use: The water a data center consumes, mainly for cooling its equipment.

References

• International Energy Agency. 2025. Energy and AI. Paris: IEA.
  • Free: International Energy Agency
• Stanford Institute for Human-Centered Artificial Intelligence. 2026. The 2026 AI Index Report. Stanford University, April 2026.
  • Free: Stanford HAI

This content reflects the field as of October 2026.

Introduction

A common claim in debates about AI is that a handful of giant companies will end up controlling it. Regulators in two countries have examined that possibility and published their concerns, and the companies involved have a reply.

This reading sets out what the regulators warned about, the counter-position, and which parts are established and which are disputed. It stops at the concern. What governments might do about it is a separate subject.

The Terms of the Concern

Competition is rivalry among companies for customers, which pushes them to lower prices and improve products. Market power is a company's ability to raise prices or limit choice without losing its customers to rivals. An incumbent is a company that already holds a strong position in a market.

Competition regulators exist to keep markets open to rivals. Their worry about AI is that incumbents from earlier technology markets, such as search, cloud computing, and phone software, could carry their positions into this one.

The UK Authority's Three Risks

In April 2024 the Competition and Markets Authority, the United Kingdom's competition regulator, named three linked risks (Competition and Markets Authority 2024, para. 29).

1. Control of critical inputs. A critical input is something a product can't be made without. For AI models the authority lists computing power, data, and expertise. Firms that control these "may restrict access to them to shield themselves from competition."
2. Use of existing market positions. Incumbents "could exploit their positions in consumer or business facing markets to distort choice." If a chatbot comes built into the phone, the search engine, or the office software you already use, you may never compare it with alternatives.
3. Partnerships. Agreements involving the main firms "could reinforce or extend existing positions of market power." A large firm that can't or doesn't buy a rival might get a similar result by investing in it.

The US Report's Concerns

Staff of the United States Federal Trade Commission studied three partnerships between cloud providers and AI developers and reported in January 2025. They identified three things to watch (Federal Trade Commission 2025).

• Access to computing power and talent. The partnerships could tie up computing capacity and engineers that rival developers need.
• Switching costs. A switching cost is the expense and effort a customer faces in moving from one supplier to another. Contract terms and technical dependence could make it hard for a developer to change cloud providers.
• Sensitive information. A cloud provider learns things about its partner's models and customers that other firms can't see.

The report states that it makes no assessment of whether anyone has acted illegally. Two of the five commissioners issued statements dissenting from parts of it.

The Counter-Position

The companies concerned, and some economists, read the same facts differently. Each of the following arguments is one they would accept as theirs, and the UK authority's own paper records several of them.

• Scale is necessary. Building frontier models costs billions of dollars. The authority wrote that the largest firms "can contribute a huge wealth of resources and expertise" (Competition and Markets Authority 2024, para. 5), and that partnerships "may be an essential ingredient for the success of independent developers" (para. 44).
• The field is active. The authority counted more than 120 models released in the six months to March 2024, and more than 330 in total. It described developers "competing intensely" (para. 27).
• Open models widen access. Some developers publish their models for anyone to run and adapt. The authority called these "an important force for competition and innovation" (para. 13).

The companies have an obvious stake in this reading. Regulators have one too: an agency that identifies a risk strengthens the case for its own powers.

What Is Established and What Is Disputed

The regulators' wording follows this division. They say "may" and "could" about harm and "we have identified" about the partnerships. Neither report finds that any company broke the law.

Both sides are forecasting. The regulators forecast that control of inputs will harden into lasting market power. The companies forecast that new entrants and open models will prevent that. The market is only a few years old, and evidence that would settle it, such as sustained high prices or rivals shut out of computing power, hasn't been reported by either regulator.

Conclusion

Competition authorities in the United Kingdom and the United States have described how control of critical inputs, existing market positions, and partnerships could entrench a few firms in AI. As of October 2026 the partnerships and their terms are documented, and harm to competition is a stated risk that these reports don't claim to have found. The counter-position holds that scale is needed and that the field remains open.

Key Terms

• Competition: Rivalry among companies for customers, which pushes them to lower prices and improve products.
• Market power: A company's ability to raise prices or limit choice without losing its customers to rivals.
• Incumbent: A company that already holds a strong position in a market.
• Critical input: Something a product can't be made without.
• Switching cost: The expense and effort a customer faces in moving from one supplier to another.

References

• Competition and Markets Authority. 2024. AI Foundation Models: Update Paper. London: CMA, April 11, 2024.
  • Free: GOV.UK
• Federal Trade Commission. 2025. Partnerships Between Cloud Service Providers and AI Developers. FTC staff report, January 2025.
  • Free: Federal Trade Commission

This content reflects the field as of October 2026.

Introduction

News about AI and China often turns on a product most people never see: the specialized chip. The United States has restricted sales of these chips to China, loosened the restriction, and argued with itself about which was the mistake.

This reading explains why chips became the focus, how US policy has shifted, and the two arguments behind the shifts. It describes policy as it stood on particular dates and takes no view on what is wise.

Why Chips

An advanced chip, in this context, is a specialized computer chip powerful enough to train or run large AI models. Training a leading model takes many thousands of them.

Governments focus on chips for a practical reason. Data and software can be copied and sent anywhere. Chips are physical objects. A few companies design the most capable ones, even fewer can manufacture them, and each shipment crosses a border where it can be counted and stopped.

This places chips at the center of strategic competition: rivalry between countries for economic and military advantage. The United States and China each treat leadership in AI as a matter of national security.

A Policy That Has Shifted

An export control is a government rule that restricts the sale of certain goods or technology to other countries. In the United States, the Bureau of Industry and Security, part of the Department of Commerce, writes and enforces these rules.

The United States has restricted exports of advanced chips to China since 2022. The policy has changed several times since.

In rescinding the tiered rule, the bureau said it would have "stifled American innovation" and harmed relations with dozens of countries by "downgrading them to second-tier status" (Bureau of Industry and Security 2025). The same announcement tightened controls aimed at China.

The January 2026 change went in the other direction. License review is the process by which a government decides whether to permit a controlled export. Under the revised policy, applications to export certain chips to China are reviewed case by case if conditions are met. The announcement names one chip model from Nvidia and one from AMD, along with similar products. The exporter must show that the sale won't reduce the supply of chips available to US customers, that the Chinese buyer has compliance procedures in place, and that the product has been independently tested in the United States (Bureau of Industry and Security 2026).

The Two Arguments

Both arguments are about national security, and both have been made by US officials.

The case for restriction. Advanced chips are the input a rival can't easily make for itself. Denying them slows that rival's progress in AI, including its military uses. In May 2025 Jeffrey Kessler, the under secretary of commerce who heads the bureau, described the aim as "keeping the technology out of the hands of our adversaries" (Bureau of Industry and Security 2025).

The case for controlled sales. If a rival can't buy US chips, it has every reason to build its own industry, and US companies lose the income that funds their next generation of products. Selling somewhat older chips keeps the rival dependent on US technology. In January 2026 the same official said that permitting sales "under controlled conditions will strengthen the American technology ecosystem" (Bureau of Industry and Security 2026).

Each argument rests on a forecast that can't yet be checked. The first assumes that restriction will hold a rival back for long enough to matter. The second assumes that dependence will last and that the chips sold won't close the gap.

Who Benefits

Chip makers gain directly from permission to sell, since China is a large market. Their public statements in favor of sales should be read with that in mind. Companies that compete with Chinese AI developers gain from restriction.

Both governments describe their positions as matters of security. China's government objects to the controls and has invested in a domestic chip industry. The two sources cited here are US government press releases, which give that government's own account of its policy.

Conclusion

Advanced chips are the part of AI that governments can most easily control, and US policy on selling them to China has moved from restriction, to a tiered worldwide rule that was withdrawn, to case-by-case approval of some sales. As of October 2026 the policy in force permits certain chips to be exported to China under conditions. The disagreement between restricting and selling turns on forecasts about how a rival will respond, and the policy may change again.

Key Terms

• Advanced chip: A specialized computer chip powerful enough to train or run large AI models.
• Strategic competition: Rivalry between countries for economic and military advantage.
• Export control: A government rule that restricts the sale of certain goods or technology to other countries.
• License review: The process by which a government decides whether to permit a controlled export.

References

• Bureau of Industry and Security. 2025. "Department of Commerce Announces Rescission of Biden-Era Artificial Intelligence Diffusion Rule, Strengthens Chip-Related Export Controls." Press release, May 13, 2025.
  • Free: Bureau of Industry and Security
• Bureau of Industry and Security. 2026. "Department of Commerce Revises License Review Policy for Semiconductors Exported to China." Press release, January 13, 2026.
  • Free: Bureau of Industry and Security

Introduction

When a regulator publishes concerns about an industry, headlines tend to report that the regulator "found" a problem. Regulators choose their verbs with care, and the difference between "could" and "has" is the difference between a warning and a finding.

This reading takes one paragraph from the United Kingdom's competition regulator and reads it word by word to separate what the regulator observed from what it fears.

Locating the Passage

The document is AI Foundation Models: Update Paper, published on April 11, 2024 by the Competition and Markets Authority, the UK's competition regulator. It is 24 pages long and free on the GOV.UK website, listed in the References. "Foundation models" is the authority's term for large general-purpose AI models, which the paper shortens to "FMs."

The passage is paragraph 29, under the heading "We see three key risks for fair, open and effective competition." The paper numbers its paragraphs, and this reading cites them by number. Paragraphs 31 to 46 explain each risk in turn. All quotations are from this paper (Competition and Markets Authority 2024).

Paragraph 29 opens: "Based on our work to date, we see the following three key interlinked risks to fair, open and effective competition." Three numbered sentences follow.

Walking Through the Passage

Step 1: Read the first risk and identify the critical inputs

The first risk reads: "Firms that control critical inputs for developing FMs may restrict access to them to shield themselves from competition."

A critical input is something a product can't be made without. Paragraph 31 names three: "compute, data or expertise." Compute is the industry's word for computing power, meaning the specialized chips and data centers that models are trained and run on.

The same paragraph gives two motives a firm might have for restricting access. One is to stop others from building models that would compete with its own. The other is to protect its position in a neighboring market, such as search or office software, from rivals who would use AI to challenge it.

Paragraph 32 supplies the facts behind the concern. Some developers have formed partnerships with major cloud providers to secure computing power, and "only a handful of firms can rely on their own compute resources." The supply of AI chips "remains limited." Those are observations. The restriction of access is the feared consequence.

Step 2: Read the second risk and identify whose positions in which markets

The second risk reads: "Powerful incumbents could exploit their positions in consumer or business facing markets to distort choice in FM services and restrict competition in FM deployment."

An incumbent is a company that already holds a strong position. The positions in question aren't in AI. Paragraph 36 lists them: "mobile and other devices, search engines, or productivity software." These are markets where a few firms already serve most users.

The worry is about the route by which AI reaches you. Paragraph 37 gives examples of what the authority had seen: Microsoft putting its own models and those of its partner OpenAI into its office software, operating system, and search engine, and Google using its own model in search. The paragraph adds that such integration "can bring benefits – such as innovation and efficiencies."

The risk is that a firm controlling the route gives its own model, or its partner's, an advantage that has nothing to do with quality. Paragraph 38 lists the means: "pre-installation, technical bundling, accessibility, integration, and compatibility."

Step 3: Read the third risk and note the word "partnerships"

The third risk reads: "Partnerships involving key players could reinforce or extend existing positions of market power through the value chain."

The word is "partnerships," and it was chosen over "mergers." When one company buys another, competition regulators can review the purchase. A partnership, such as a large investment paired with a supply agreement, may achieve some of the same effects without a purchase. Paragraph 45 says so directly: "not all such partnerships and investments will fall within the scope of merger control rules and some may have been structured to seek to avoid them."

Paragraph 42 names the firms. The authority groups Google, Apple, Microsoft, Meta, and Amazon under an abbreviation and adds Nvidia, "the leading supplier of AI accelerator chips." Paragraph 43 gives the count: "We have identified an interconnected web of over 90 partnerships" involving those six.

Paragraph 44 then gives the other side. Such partnerships "may be an essential ingredient for the success of independent developers," and the authority understands "that they can potentially bring pro-competitive benefits."

Step 4: Mark the modal verbs

Go back to paragraph 29 and look at the verb in each sentence.

• Risk 1: firms "may restrict access."
• Risk 2: incumbents "could exploit their positions."
• Risk 3: partnerships "could reinforce or extend" market power.

"May" and "could" are modal verbs. They say that something is possible and stop short of saying that it happened. The lead-in sentence does the same work with the word "risks" and the phrase "based on our work to date."

The pattern continues in the explanatory paragraphs. Paragraph 31 says firms "could restrict access." Paragraph 39 says exclusive access "could entrench" positions. Paragraph 45 says the authority is "vigilant against the possibility" that incumbents "may try" to use partnerships to quash threats.

Step 5: Separate what the authority found from what it warns might happen

Now sort the paper's statements by their verbs.

Found, in the present or past tense:

• "We have identified" more than 90 partnerships (para. 43).
• "We have seen" incumbents rapidly integrating models into their existing products (para. 37).
• "We have seen" developers forming partnerships with cloud providers to get computing power, and chip supply "remains limited" (para. 32).

Warned of, with "may" or "could":

• that access to inputs will be restricted
• that consumer choice will be distorted
• that market power will be reinforced or extended

The paper contains no statement that any firm has restricted access, distorted choice, or broken competition law. In paragraphs 35 and 41 the authority says it is "yet to take any provisional decisions" on which areas to investigate.

Key Considerations

A competition authority's work often runs in this order: it maps a market, states risks, and only later, after an investigation with evidence and a right of reply, reaches findings. An update paper sits at the second stage.

The common mistake is to report a stated risk as a finding of wrongdoing, as in "UK regulator finds tech giants are stifling AI competition." The paper says no such thing. It says they could.

The opposite mistake is to treat "could" as meaning "nothing to see." The authority considered these risks serious enough to publish, and the facts it did establish, the partnerships and the integration, are the conditions under which the risks would arise.

The authority has a stake of its own. The paper refers to new legal powers then before Parliament, and a regulator describing risks is also making the case for having the tools to address them. The paper dates from April 2024, and the market and the authority's powers have changed since.

Summary

Paragraph 29 states three risks, each with a modal verb, resting on facts the authority did establish. For each risk, the table sets out what would have to happen for it to become real and what evidence would show that it had.

References

• Competition and Markets Authority. 2024. AI Foundation Models: Update Paper. London: CMA, April 11, 2024.
  • Free: GOV.UK