Power Concentration in AI

The Accountability Gap

Comparing Nvidia's market capitalization with the economic output of a country can produce a striking headline. At roughly $4.85 trillion, Nvidia's valuation was, for example, larger than Japan's roughly $4.38 trillion annual GDP. But the comparison is a category error. Market capitalization reflects what investors believe a company is worth; gross domestic product measures the value of goods and services produced across an economy over a year. One is a valuation; the other is a flow.

Stripping that comparison away reveals a more useful question. Scale alone is not the problem. The critical issue is how much capital a small number of firms can mobilize, who controls its allocation, and what mechanisms exist for the public to scrutinize decisions whose consequences extend well beyond those firms.

The Real Weight

A better comparison is flow against flow: actual or planned investment against other annual flows of resources. Amazon, Alphabet, Microsoft, and Meta now expect to record roughly $720 billion to $745 billion in 2026 capital expenditures. Much of that spending is being driven by AI and the data centers, chips, servers, networking equipment, and power infrastructure needed to support it.

The scale is extraordinary. The four companies' combined 2026 capital-expenditure plans are roughly four times the federal government's $181 billion FY2026 research-and-development request and amount to about four-fifths of the $891 billion authorized for national defense. These are not perfectly equivalent accounting categories, but they establish the order of magnitude: four corporate management teams are directing annual investment on a scale comparable to major functions of the federal government.

The acceleration is nearly as important as the total. Estimates put the same four companies' 2025 capital investment at roughly $410 billion. Their 2026 plans therefore represent another enormous expansion in infrastructure spending, although differences in company accounting—particularly the treatment of leases—make a precise year-over-year percentage less meaningful than the scale of the change itself.

The financial pressure is already visible. Amazon reported that its trailing-twelve-month free cash flow had fallen to negative $7.6 billion by June 2026. The company attributed the decline primarily to a $66.1 billion year-over-year increase in purchases of property and equipment, net of proceeds and incentives, and said that increase primarily reflected investment in artificial intelligence. Amazon has explicitly described the resulting short-term free-cash-flow pressure as a cost it is willing to bear in pursuit of the longer-term AI opportunity.

None of this means that private infrastructure spending is itself improper, or that governments have historically been the only institutions capable of projects at this scale. The more consequential fact is the concentration of the decision-making. A handful of corporate boards can now determine where hundreds of billions of dollars in computing infrastructure will be built, which technologies will receive that investment, and which bets will shape the physical foundations of the AI economy.

The scale of the capital is one fact. The small number of institutions directing it is another. That leads to the more important question: who actually holds decision-making authority inside those institutions?

The Inner Ring: Control Decoupled from Ownership

The concentration of capital becomes more consequential when economic ownership and decision-making authority do not line up in the ordinary way. Across several leading AI companies, specialized governance structures give particular founders, boards, nonprofits, or trusts powers that exceed what their proportional economic ownership alone would provide.

The mechanisms differ substantially, and the differences matter. Founder super-voting shares are not equivalent to a nonprofit foundation or a mission trust. But they share one structural feature: important governance rights are concentrated in a small group rather than distributed in proportion to ordinary ownership.

The Super-Voting Share: Meta and Alphabet

The clearest examples are Meta and Alphabet.

At Meta, Mark Zuckerberg holds approximately 60.8% of the company's voting power through Class B shares carrying ten votes each, despite holding only about 13.5% of the company's economic interest. That gives him effective majority control over shareholder votes.

The distinction matters in practice. Meta shareholders again considered a proposal in May 2026 to move toward a one-share-one-vote structure. The proposal failed. More important than the result of any single vote is the underlying arithmetic: Zuckerberg's voting block is large enough that, if he opposes such a change, the remaining shareholders cannot enact it through an ordinary shareholder vote. The dual-class structure therefore does not merely give a founder influence; it gives him an effective veto over attempts by ordinary shareholders to dismantle that concentration of voting power.

Alphabet uses a similar mechanism. Larry Page and Sergey Brin together hold approximately 52.7% of Alphabet's voting power through super-voting Class B stock, even though both have stepped back from daily management. Alphabet itself warns investors that this concentration allows the founders to influence the election of directors and major corporate transactions and can determine the outcome of most matters submitted to shareholders.

These structures are not hidden. They are disclosed openly in corporate filings and accepted by investors who buy the stock. The relevant question is therefore not whether they are legitimate under corporate law. It is what their consequences are when companies governed this way make decisions about technologies with effects extending far beyond their shareholders.

Founder Control: SpaceX and xAI

SpaceX and xAI illustrate a different form of concentration.

xAI became a wholly owned subsidiary of SpaceX following the companies' merger. SpaceX then entered the public markets in June 2026 with a dual-class share structure designed to preserve overwhelming voting control for Elon Musk. Class B shares carry ten votes each, and the structure gives their holders separate rights over the election of a majority of the board. SpaceX consequently operates as a controlled company even after opening its shares to public investors.

This is not the same ownership-control separation found at Meta or OpenAI. Musk also possesses a very large economic interest in SpaceX. The significance lies instead in the degree to which both economic ownership and formal voting authority are concentrated in the same individual, while xAI sits within the corporate structure he controls.

The distinction is useful. Concentrated AI governance does not arise from one legal template. Sometimes voting rights greatly exceed economic ownership. Sometimes ownership and voting control are both concentrated. The common feature is the small number of people ultimately empowered to determine corporate direction.

The Nonprofit Controller: OpenAI

OpenAI presents a different model again.

Following its 2025 recapitalization, the commercial entity, OpenAI Group, became a Delaware public benefit corporation. Microsoft holds roughly 27% of its equity. The nonprofit OpenAI Foundation holds approximately 26%. Yet equity percentages do not determine control: through special governance rights held solely by the Foundation, the Foundation appoints every member of OpenAI Group's board and can replace those directors.

This is an unusually explicit separation of economic ownership from corporate control. A shareholder holding a larger economic stake than the Foundation does not thereby acquire greater governance authority.

The public-benefit-corporation element should not be dismissed as legally meaningless. Delaware law requires directors of a public benefit corporation to balance shareholders' financial interests, the interests of people materially affected by the corporation's conduct, and the public benefit specified in its charter. OpenAI therefore operates under legal obligations that differ from those of a conventional corporation.

Those obligations nevertheless have limits as a mechanism of public accountability. Delaware law does not give every person affected by the corporation a direct fiduciary claim against its directors simply because that person's interests are implicated. And OpenAI's special governance rights remain vested in the Foundation's board rather than in the broader public whose interests the mission is intended to serve.

That distinction matters. Mission-oriented governance may provide a genuine check on ordinary shareholder pressure without thereby becoming democratic governance. The Foundation may be designed to act for a public purpose, but the public does not appoint its directors.

The Mission Trust: Anthropic

Anthropic offers perhaps the clearest attempt to build an independent mission constraint directly into corporate governance.

Anthropic is a public benefit corporation, but it also created a Long-Term Benefit Trust. The Trust holds a special class of stock giving it authority to appoint and remove a portion of Anthropic's board, with that authority designed from the beginning to phase in until Trust-appointed directors constituted a majority. The trustees are intended to be financially disinterested: they hold no Anthropic equity and do not participate in the company's profits.

That planned transition was not merely theoretical. In April 2026, Anthropic announced that an additional Trust-appointed director had brought LTBT-appointed directors to a majority of the board at that time. Board membership has changed since then, so the precise current seat count is better treated as a live governance indicator than frozen into this essay. The underlying governance authority remains part of Anthropic's corporate structure.

The Trust is also not absolutely immune from shareholders. Anthropic's original description of the structure includes failsafe provisions allowing sufficiently large shareholder supermajorities, under specified circumstances, to modify the arrangement without trustee consent. That qualification matters because the Trust is intended to be durable, not literally untouchable.

It would therefore be a mistake to describe Anthropic's Trust as simply another version of Zuckerberg's super-voting shares. The similarity is structural but limited. Both mechanisms separate some governance authority from proportional economic ownership. But Meta concentrates that authority in its founder; Anthropic deliberately assigns authority to financially disinterested outsiders in an attempt to counterbalance investor and management incentives.

The normative question is consequently more difficult than whether the structures are the same. They are not. The question is whether assigning extraordinary authority to a small group becomes equivalent to public accountability merely because that group is given a public-benefit mandate. It does not. A mission trust may provide an important internal check while still leaving consequential decisions in the hands of people whom the broader public neither selects nor removes.

Across these companies, then, there is no single mechanism and no single motive. Founder shares entrench founders. Controlled-company structures combine ownership with voting authority. Nonprofit governance and mission trusts deliberately constrain ordinary investor control. Some of these arrangements may provide valuable protections against short-term financial pressure.

But they produce a common structural fact: enormous economic investment does not translate straightforwardly into proportional governance authority, and public importance does not translate straightforwardly into public control. Decisions over some of the world's most consequential AI systems remain concentrated among relatively small groups operating through governance arrangements largely designed inside the companies themselves.

The next question is whether these centers of authority operate independently. They compete intensely—but competition does not eliminate the financial and infrastructural dependencies connecting them.

The Interlock: Competition and Dependency

The companies developing frontier AI compete intensely. OpenAI, Anthropic, Google, Meta, xAI, Microsoft, and others compete for users, enterprise contracts, researchers, chips, power, data-center capacity, and investment. Their rivalry is real.

But competition does not imply independence.

The frontier AI industry is developing through a network in which the same small group of firms repeatedly appears in different roles: investor, cloud provider, chip supplier, model developer, customer, and distribution platform. A company may compete with another at one layer of the system while depending on it—or on one of the same few infrastructure providers—at another.

The result is not the disappearance of competition. It is competition embedded within a highly concentrated and interdependent stack.

Investment Becomes Infrastructure

The Federal Trade Commission examined three of the most important relationships in this system: Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic. Its 2025 staff report documented arrangements involving more than $20 billion in cumulative investment, along with substantial non-monetary exchanges.

The partnerships went far beyond ordinary minority investments. According to the FTC, their terms included equity and revenue-sharing rights, billions of dollars in cloud-computing commitments, discounted compute, consultation and exclusivity provisions, access to technical and business information, and opportunities to integrate the AI developers' models into their partners' products. The agency warned that such arrangements could raise switching costs, affect access to scarce computing resources, and give the cloud providers access to information unavailable to competitors.

This creates an unusual relationship between financing and infrastructure. The companies providing capital to frontier-model developers are often also selling them the computing services on which the models are trained and deployed.

That does not mean the investment simply “comes back” to the investor in some closed accounting loop. The FTC did not make that claim, and the economics are more complicated. But it does mean that investment, infrastructure purchasing, strategic access, and product distribution can become contractually intertwined in ways that make the relationship substantially deeper than a conventional financial stake.

OpenAI and Microsoft: Dependency With Increasing Flexibility

The Microsoft–OpenAI relationship illustrates both the depth of these dependencies and their ability to change.

Microsoft remains a major OpenAI shareholder and, under the companies' April 2026 amended agreement, remains OpenAI's primary cloud partner. OpenAI products are scheduled to launch first on Azure unless Microsoft cannot or chooses not to provide the required capability. Microsoft also retains a license to OpenAI's model and product intellectual property through 2032.

But the relationship is becoming less exclusive.

The 2026 agreement made Microsoft's IP license non-exclusive and allows OpenAI to serve its products to customers across any cloud provider. OpenAI has also been expanding compute relationships outside Azure.

That is important counterevidence to any simple lock-in story. OpenAI remains deeply connected to Microsoft, but it has simultaneously acquired greater infrastructure flexibility.

The pattern is therefore not permanent dependence on one provider. It is a frontier developer attempting to diversify while remaining embedded in relationships with some of the world's largest technology and infrastructure companies.

Anthropic: Diversification Across the Same Concentrated Layer

Anthropic makes that pattern even clearer.

Amazon remains Anthropic's primary cloud provider and training partner. In April 2026, Anthropic announced an agreement committing more than $100 billion over ten years to AWS technologies and securing up to five gigawatts of additional compute capacity. Anthropic said it was already using more than one million Amazon Trainium2 chips to train and serve Claude.

Yet Anthropic is not simply an Amazon-dependent company.

It has also announced a major infrastructure agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity beginning in 2027. Anthropic says it trains and operates Claude across Amazon Trainium, Google TPUs, and Nvidia GPUs. Claude is also available to customers through AWS, Google Cloud, and Microsoft Azure.

This is genuine diversification. It reduces dependence on any single chip architecture or cloud provider.

But it also reveals the structural feature that matters here: diversification at the frontier often means distributing dependence across the same small set of enormous infrastructure providers.

Anthropic can reduce its exposure to Amazon by using Google and Nvidia. OpenAI can reduce its exposure to Microsoft by purchasing compute elsewhere. Those are meaningful changes. Yet they do not necessarily make the infrastructure layer itself much less concentrated.

The Chokepoints Beneath the Models

Three layers illustrate the point particularly clearly.

In semiconductor fabrication, TSMC accounted for approximately 72% of global foundry revenue in the first quarter of 2026, up from 70.4% the previous quarter. Samsung, the second-largest foundry, accounted for only 6.5%. The figure covers the broader foundry market rather than AI chips alone, but AI GPU and xPU demand was an important contributor to TSMC's growth.

At the cloud layer, Amazon, Microsoft, and Google together accounted for 63% of worldwide cloud-infrastructure-services spending in Q1 2026: 28%, 21%, and 14%, respectively. Their combined share was also 63% in the preceding quarter. The concentration did not increase, but neither did it meaningfully dissipate as the market expanded.

At the accelerator layer, Nvidia remains a pivotal supplier to many of the companies building the systems that ultimately compete with one another. Nvidia's Data Center business generated $75.2 billion in its first fiscal quarter of 2027. Approximately half of that revenue came from the “hyperscale” category, which Nvidia defines to include public clouds and the world's largest consumer-internet companies.

That last relationship runs in both directions.

The hyperscalers depend heavily on Nvidia's accelerators and networking technology. Nvidia, meanwhile, derives roughly half of its Data Center revenue from the hyperscale customer category.

This is not simply a supplier exercising power over customers. It is mutual dependence between highly concentrated layers of the same system.

The major cloud companies are responding by developing their own accelerators—Google TPUs, Amazon Trainium, and other custom silicon—while AMD and additional competitors challenge Nvidia. Those developments could weaken today's chokepoints. They are therefore important indicators to track rather than reasons to assume that the present structure is permanent.

CoreWeave: Concentration Can Weaken

CoreWeave offers another useful qualification.

The company became an important provider of GPU-intensive cloud infrastructure while depending extraordinarily heavily on a small number of customers. Microsoft accounted for approximately 67% of CoreWeave's 2025 revenue.

But that concentration has begun to decline.

In the first quarter of 2026, CoreWeave reported that its largest customer accounted for 45% of revenue and its second-largest for another 20%. The filing does not identify those customers and explicitly warns that the anonymized customer labels are not necessarily consistent between reporting periods. It would therefore be incorrect to claim that Microsoft itself remained at any particular percentage.

The relevant finding is more nuanced: CoreWeave remained highly concentrated—two customers still generated 65% of its revenue—but dependence on a single customer had fallen substantially.

That movement matters because it runs against the strongest version of the concentration thesis. The system is capable of diversifying.

A useful account of AI power should register that evidence rather than treating every change as confirmation of increasing concentration.

Switching Costs and Strategic Information

The remaining concern is not merely how many providers exist, but how easily firms can move between them.

The FTC found that frontier-AI partnerships included large committed cloud expenditures and contractual provisions that could restrict or complicate the use of alternative providers. It also identified technical barriers to migrating between specialized cloud services and different AI chips.

Those switching costs matter because infrastructure choices are cumulative. Models are optimized for particular hardware; engineers build expertise around specific software ecosystems; data pipelines and deployment systems are constructed around particular clouds; and large compute commitments can extend for years.

The same partnerships can also give infrastructure providers access to unusually sensitive information. The FTC reported access to model-development methods, confidential chip co-design plans, financial information, customer usage, and revenue data. Some of those providers are simultaneously developing AI systems that compete with the developers they support.

Again, that does not make the companies one organization. It means the boundary between competitor, supplier, investor, and strategic partner is unusually porous.

Competition Inside a Concentrated Stack

The distinction matters.

If OpenAI and Anthropic compete fiercely, that competition is real. If Amazon develops Trainium partly to reduce dependence on Nvidia, that is real competitive pressure. If CoreWeave diversifies its customer base, concentration genuinely declines. If OpenAI acquires the ability to use additional cloud providers, its dependence on Microsoft weakens.

None of those developments contradicts the broader structural concern.

The concern is that much of this competition still occurs through infrastructure controlled by a relatively small number of firms. Frontier developers can switch among clouds, but the global cloud market is dominated by three providers. They can diversify accelerator architectures, but leading-edge fabrication remains extraordinarily concentrated. Cloud providers can develop their own chips, but those chips still depend on scarce advanced manufacturing capacity.

The system therefore contains two realities at once:

vigorous competition among firms and persistent concentration in the infrastructure on which that competition depends.

That is a more complicated picture than a cartel or a unified corporate bloc. It is also more consequential than an ordinary competitive market in which participants can readily substitute suppliers and exit relationships.

The question is not whether the major AI companies secretly constitute a single organization. They do not.

The question is how much independence competition can provide when competitors repeatedly depend on the same limited set of capital providers, chip manufacturers, cloud platforms, and physical infrastructure—and when many of those providers are themselves competing participants in the AI market.

Those dependencies become especially important when the physical costs of AI development leave the balance sheets of the companies and arrive in particular communities.

The Outer Ring: Footprint and Political Accountability

The concentration described so far takes place inside companies and infrastructure markets. Its consequences do not.

AI systems ultimately require physical facilities: data centers, transmission lines, power generation, cooling systems, land, and increasingly large quantities of electricity. The resulting costs and benefits are distributed geographically, often far from the corporate boards deciding where and how rapidly to build.

That creates a different kind of accountability problem. The question is no longer simply who controls an AI company. It is how decisions made by a relatively small number of firms interact with electric grids, local governments, communities, regulators, and ultimately voters.

A Global Footprint With Local Concentration

Globally, data centers consumed approximately 485 terawatt-hours of electricity in 2025, about 1.5% of worldwide electricity demand. The International Energy Agency expects that consumption to roughly double to around 950 terawatt-hours by 2030, approaching 3% of global electricity demand. AI-focused data centers are expected to grow substantially faster than the sector overall.

Those global percentages are useful context. They show that data centers remain a relatively modest share of worldwide electricity use even during a period of extraordinary expansion.

But the global average obscures something equally important: data centers are not distributed evenly across the electrical system.

Northern Virginia illustrates the difference. Dominion Energy reports that data centers accounted for 28% of Virginia Power's electricity sales in 2025, up from 26% in 2024 and 24% in 2023. The company describes their concentration, particularly in Loudoun County, as a unique challenge requiring significant investment in generation and transmission infrastructure.

That is a much more precise claim than saying data centers consume 28% of “Virginia's electricity.” Virginia Power is a particular regulated utility, not the entire state. But within that system, more than one-quarter of electricity sales already go to one rapidly growing class of customer.

This distinction between global scale and local concentration is central to understanding AI's physical footprint. A technology can consume a manageable percentage of electricity globally while creating very large infrastructure requirements in particular regions.

Who Pays for the Buildout?

Rapid electricity-demand growth requires investment: new generation, transmission lines, substations, and other infrastructure. Dominion's own filings explicitly connect its investment plans to increasing demand from larger data-center customers.

The difficult political question is how those costs should be allocated.

It would be too simple to say that ordinary ratepayers necessarily pay for the AI buildout. Utilities, regulators, data-center operators, and legislatures can structure rates, contracts, deposits, reimbursements, and infrastructure requirements in different ways. Dominion, for example, has imposed deposits and reimbursement provisions intended to protect against projects entering its queue and later being canceled.

But the allocation problem is real. Large new loads can require system-wide investments whose costs and risks must ultimately be assigned somewhere. That makes decisions about data-center expansion not merely private transactions between technology companies and utilities, but questions of public utility regulation and infrastructure policy.

Transparency Before Approval

Those questions become harder when local governments negotiate with developers under confidentiality agreements.

Researchers at the University of Mary Washington submitted public-records requests to every Virginia locality they could identify with an existing, approved, or proposed data-center project. They found that 25 of 31 localities had at least one nondisclosure agreement with a data-center company.

That finding should be stated carefully. It does not mean that 25 localities secretly approved data centers under NDAs. The study established the prevalence of confidentiality agreements, not that every project approval occurred without public scrutiny.

But the contents of the agreements still raise a legitimate accountability question. The researchers reported NDAs covering not only proprietary technology but broader business plans and other nonpublic information, and argued that such restrictions can limit the information available for public debate while projects are being considered.

Confidentiality during economic-development negotiations is not unique to data centers, and companies have legitimate reasons to protect commercially sensitive information. The problem arises when the information withheld is also material to a community's ability to evaluate the scale, resource requirements, or consequences of a project on which local officials will eventually vote.

The relevant question is therefore not whether all NDAs are illegitimate. It is whether the boundary between legitimate commercial confidentiality and information necessary for democratic decision-making is being drawn in the right place.

The Public Response Is Changing

Virginia also provides evidence that public attitudes toward data-center development have changed sharply.

In a 2023 Washington Post–Schar School poll, 69% of Virginia voters said they would be comfortable with a data center being built in their community. By spring 2026, that figure had fallen to 35%. In the same 2026 survey, 59% said data centers were having a negative effect on the local environment.

That shift does not establish why opinion changed, and it would be a mistake to attribute it simply to secrecy or electricity prices. Voters may be responding to multiple concerns at once: energy demand, environmental effects, land use, noise, utility bills, tax incentives, employment expectations, or broader unease about AI.

But the political change is no longer hypothetical. Data-center projects have become contested issues in local and state politics, and some major proposals have recently been rejected or abandoned after sustained opposition.

Public opposition therefore represents an actual countervailing force. Corporate capital and infrastructure demand can be enormous, but local political institutions and organized residents can still alter outcomes.

That evidence matters because an account of concentrated power should not assume that concentrated power is absolute power.

Political Money Moves in Both Directions

The struggle over AI policy is also moving into elections.

Here too, the picture is more complicated than a unified technology industry spending money to prevent regulation.

Leading the Future, a political network backed by figures associated with OpenAI, Andreessen Horowitz, Palantir, and others, has raised roughly $140 million for the 2026 election cycle. It generally favors candidates supportive of rapid AI development and a lighter regulatory approach.

At the same time, a rival political network centered on Public First Action and affiliated organizations has assembled roughly $80 million to support a more safety-oriented approach to AI policy. Anthropic has contributed $40 million to Public First Action during 2026, although Anthropic says those funds cannot be used directly to support or oppose individual candidates.

These figures require a distinction that is easily lost in political reporting: money raised is not the same as money spent. Political organizations operate through different legal entities, including super PACs and nonprofit advocacy organizations, with different disclosure and spending rules. Aggregating every announced commitment into a single number can therefore give a misleading impression of how much money has actually entered elections.

The more defensible conclusion is broader.

AI-linked wealth is financing competing visions of how AI should be governed.

That is evidence against the idea of a politically unified AI industry. Anthropic and OpenAI-associated political networks, for example, are spending substantial resources on opposing sides of important regulatory questions.

But the competition creates another accountability issue. Increasingly, the contest over the rules governing AI is itself being financed by actors with enormous economic stakes in the outcome.

This does not mean campaign spending purchases election results, nor that voters and elected officials lack independent agency. Political money does not translate mechanically into political control.

It does mean that access to the political arena is highly unequal. Organizations able to commit tens or hundreds of millions of dollars can fund advertising, research, advocacy, candidate support, lobbying, and sustained political organization at a scale unavailable to ordinary citizens.

The result is a peculiar form of pluralism: competing private centers of wealth may check one another, but the range and intensity of political advocacy can still be disproportionately shaped by those wealthy enough to finance it.

The Accountability Problem Changes at the Boundary

The physical and political footprint of AI therefore produces a different form of concentration from the corporate structures discussed earlier.

Electricity consumption is not governance control. An NDA is not market concentration. Campaign spending is not ownership. They should not be collapsed into a single metric.

What connects them is the movement of consequences across institutional boundaries.

A decision to build computing infrastructure may originate inside a technology company, require chips produced through a concentrated semiconductor supply chain, depend on one of a handful of cloud or infrastructure providers, generate electricity demand managed by regulated utilities, require land-use decisions by local governments, and ultimately provoke action by voters, legislators, courts, and regulators.

No single institution governs that entire chain.

Nor are public institutions powerless. Utility regulators can regulate rates. Local governments can approve or reject projects. Courts can overturn approvals. State legislatures can establish environmental, energy, and land-use rules. Congress and federal agencies can regulate competition, securities, interstate commerce, energy, and other dimensions of the system. Voters can remove officials who make decisions they oppose.

The problem is not the absence of institutions.

It is that authority is divided among many institutions while the underlying technological and economic system crosses all of their jurisdictions.

That is the accountability gap.

The Gap: What No Institution Is Sized to Close

The problem described here is not that government has disappeared.

Regulators investigate AI partnerships. Competition authorities scrutinize cloud markets. Utility commissions decide how infrastructure costs are allocated. Local governments approve or reject data centers. Courts review those decisions. States enact AI laws. Congress debates national standards. Elections determine who appoints regulators and writes the rules.

The problem is that each institution sees a different piece of the system.

Corporate law governs voting rights and fiduciary duties. Competition law asks whether markets remain competitive. Securities law governs disclosure to investors. Utility regulation governs electricity rates and infrastructure. Local land-use law governs where facilities can be built. Campaign-finance law governs particular forms of political spending. None of these institutions was designed to ask how concentrated corporate control, cloud infrastructure, semiconductor supply chains, electricity demand, local permitting, and political influence interact as one system.

That is the accountability gap.

Regulation Can Reach the Pieces

The Microsoft–OpenAI relationship provides a useful example of both the reach and the limits of existing regulation.

In 2025, the United Kingdom's Competition and Markets Authority examined whether Microsoft's relationship with OpenAI created a merger situation that fell within the agency's jurisdiction. The CMA concluded that Microsoft exercised material influence over OpenAI's commercial policy. But it also concluded that Microsoft did not exercise de facto control and that there had therefore been no qualifying change of control under the relevant merger provisions. Without that jurisdictional threshold being met, the CMA could not review the partnership as a merger. The agency explicitly cautioned that this did not amount to a finding that the relationship raised no competition concerns.

That distinction is important.

The CMA was not blind to the relationship. It investigated it in detail and identified significant influence. The limitation arose because merger law asks a particular legal question: had control changed in the way required to trigger merger jurisdiction?

Other regulators ask different questions. The U.S. Federal Trade Commission used its information-gathering authority to investigate Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic. Its report documented equity and revenue-sharing rights, cloud-spending commitments, switching costs, access to sensitive information, and other features that could affect competition.

The FTC could observe relationships that did not necessarily constitute conventional acquisitions.

The same pattern appears elsewhere. Competition authorities have investigated barriers to switching and multi-cloud use, while newer digital-market regimes are giving regulators tools beyond conventional merger law.

These examples make the institutional problem more precise.

Existing regulatory tools are not useless. Some are adapting. New powers can be created. Agencies can investigate contractual arrangements that fall outside conventional merger analysis.

But the legal system still divides the underlying structure into separate questions: Is this a merger? Is this conduct anticompetitive? Does this firm possess strategic market power? Are customers able to switch providers?

Those are necessary questions. They are not the same as asking whether the combined concentration of capital, governance authority, compute infrastructure, and physical resources creates a broader distribution-of-power problem.

Federalism Creates Another Layer

AI governance is also divided vertically between federal, state, and local authority.

Congress demonstrated in July 2025 that this allocation of authority remains unsettled. A proposal that would have imposed a ten-year moratorium on many state AI regulations was removed from federal legislation by a 99–1 Senate vote.

That was not the collapse of a state-level moratorium. It was the rejection of a proposed federal restriction on state regulation.

The underlying dispute did not disappear.

In December 2025, Executive Order 14365 established an administration policy favoring a more uniform national AI framework and directed the Justice Department to create a task force to challenge certain state AI laws. The order also called for recommendations for federal legislation that could preempt conflicting state rules while treating some areas of state authority differently.

Meanwhile, members of Congress have continued to advance competing approaches to the division of federal and state authority.

This is not legislative paralysis in the literal sense. It is an unresolved constitutional and political question about which level of government should govern which parts of AI.

And that question matters because the system itself operates at several levels simultaneously.

A frontier model may be developed by a company operating globally, trained on chips fabricated overseas, deployed through interstate cloud infrastructure, powered by facilities requiring state-regulated electricity, and housed in data centers permitted by a county government.

There is no obvious jurisdictional boundary around “AI power” because the phenomenon does not respect the boundaries through which government is organized.

Fragmentation Is Not Powerlessness

It would be a mistake to infer from this fragmentation that corporations can simply do whatever they want.

The evidence in this essay points in the opposite direction as well.

Local opposition has stopped or altered data-center projects. Regulators have investigated cloud concentration and AI partnerships. Companies are responding to competitive pressure by diversifying cloud providers and developing alternative chips. CoreWeave's dependence on a single customer has declined. Congress overwhelmingly rejected one proposed restriction on state AI regulation. Public officials, courts, shareholders, consumers, competitors, and voters remain capable of constraining corporate behavior.

Power is concentrated, but it is not absolute.

That distinction matters because the question is not whether private companies have replaced governments.

They have not.

The question is whether existing mechanisms of accountability are well matched to the form that private power is taking.

A county government can decide whether a data center belongs on a particular parcel of land. It cannot redesign the cloud market.

A utility commission can decide how grid costs are allocated. It cannot alter Meta's voting structure.

An antitrust agency can investigate switching costs or market power. It does not appoint the board of the OpenAI Foundation.

Corporate shareholders can vote on directors and resolutions, except where dual-class structures sharply limit their influence. They do not represent residents living beside new transmission infrastructure.

Each institution may function exactly as designed while still leaving important questions outside its field of view.

The Problem Is the Connections Between the Pieces

This is why the relationships documented throughout this essay matter.

A hyperscaler is not merely a technology company. It may simultaneously be a cloud provider, AI developer, investor, infrastructure purchaser, political donor, and customer of the semiconductor supply chain.

A frontier AI laboratory may be simultaneously an independent company, a recipient of strategic investment, a customer of one or several cloud providers, a partner in chip development, and a participant in policy debates about the rules governing all of those relationships.

A data-center proposal may appear locally as a land-use decision while being economically connected to global competition for compute capacity and physically connected to regional electricity planning.

Viewed separately, every relationship can look ordinary.

Companies invest in suppliers. Customers sign long-term contracts. Founders retain voting control. Utilities build infrastructure. Local governments compete for development. Political organizations raise money.

The structural concern emerges from their combination.

No single actor needed to design this system as a whole. The governance structures described here were deliberate. The investments were deliberate. The cloud contracts were deliberate. The NDAs were deliberate. The political expenditures were deliberate.

What was not centrally designed was the aggregate architecture produced when those decisions accumulated across institutions.

That distinction is important. The system is not the product of a conspiracy, nor simply the accidental result of selfish choices. It is an emergent structure created by many actors pursuing different objectives under different rules.

What Accountability Would Have to See

A serious response to concentrated AI power therefore does not necessarily require one new super-regulator with authority over everything.

There are good reasons to divide governmental power among institutions with specialized expertise. Antitrust agencies should not decide local zoning disputes, and county boards should not regulate semiconductor competition.

The harder problem is coordination and visibility.

An adequate accountability system would need to be able to see relationships that cross institutional boundaries: when infrastructure concentration reinforces corporate dependence; when governance structures insulate decisions from ordinary shareholders; when local infrastructure choices accumulate into regional energy constraints; when strategic investments create both financial ties and switching costs; and when the private actors most affected by regulation become major participants in financing the political debate over that regulation.

The missing capacity is therefore not simply more regulation.

It is the capacity to understand how the pieces interact and to determine when individually lawful arrangements collectively produce concentrations of power that existing institutions should address.

That is a different question from whether AI itself is good or bad, whether a particular company behaves responsibly, or whether any single corporate governance structure is legitimate.

It is a question of institutional design.

The Accountability Gap

The pattern that emerges is more complicated than the one I originally set out to describe.

Competition among AI companies is real. Infrastructure dependencies can weaken. Mission-oriented governance structures can provide genuine checks on investors. Regulators can intervene. Communities can reject projects. Political money flows toward competing visions of AI policy rather than a single unified industry position.

Those qualifications matter.

They do not eliminate the underlying concentration.

A small number of firms still mobilize extraordinary quantities of capital. Important governance rights remain concentrated among a small number of founders, boards, foundations, and trustees. The infrastructure underlying frontier AI remains concentrated at several critical layers. The physical consequences of that development are geographically concentrated. And the political contest over how the system should be governed increasingly includes organizations funded by actors with enormous economic stakes in its future.

The resulting problem is not that democratic institutions have ceased to exist.

It is that the structure of private power increasingly crosses the boundaries among them.

That is the accountability gap: not an absence of authority, but a mismatch between a system whose components are increasingly interconnected and public institutions whose authority remains divided among separate domains.

Whether that mismatch requires new institutions, stronger coordination among existing ones, different corporate-governance rules, new competition policy, greater transparency, or some combination of these remains an open question.

But diagnosing the problem accurately comes first.

The indicators discussed in this essay are tracked on the accompanying Power Concentration in AI — Live Indicators page and updated when new comparable observations become available. The purpose of that tracker is not to assume that every measure will move toward greater concentration. It is to record whether the structural conditions described here—capital mobilization, governance control, infrastructure concentration and dependency, and the resulting physical and political footprint—strengthen, weaken, or persist over time.

Reasoning. The central claim of this essay is not that AI companies form a single bloc, that competition among them is illusory, or that public institutions are powerless. It is that unusually large concentrations of capital, governance authority, and critical infrastructure increasingly intersect across institutional boundaries that are governed separately. The evidence I place the most weight on is the scale of capital mobilization by a small number of firms; concentrated voting and governance rights; concentration at key infrastructure layers such as cloud services, advanced chip fabrication, and AI accelerators; and the fragmented way corporate, competition, utility, land-use, and political institutions oversee those different pieces. Evidence running in the other direction matters as well: firms compete intensely, infrastructure dependencies can weaken, customers and suppliers can diversify, communities can block projects, and regulators can intervene. I therefore regard the argument as a claim about structural concentration and institutional mismatch, not absolute corporate control. The accompanying Live Indicators page is intended to make that claim testable over time by recording evidence whether it strengthens, weakens, or persists.

Note on method and AI use. The argument and final judgments are mine. I used ChatGPT as a research and editorial assistant during revision: to audit factual claims, locate and compare source material, identify possible errors and overstatements, surface counterevidence, test the logic of the argument, and draft or revise passages for my review. I decided which claims and revisions to accept and take responsibility for the resulting argument. Sources were checked against primary company, regulatory, government, and legal materials where available; market-wide estimates, polling, and other claims that could not be established from primary sources rely on the research or reporting cited. Confidence is therefore not identical across every figure: company filings and legal documents generally provide firmer support than analyst market-share estimates, political-spending compilations, or broader causal interpretations. I have tried to identify those distinctions rather than present all evidence as equally certain.