Power Concentration in AI
The Accountability Gap
Commentators frequently contrast Nvidia’s $4.85 trillion market capitalization with Japan’s $4.38 trillion economy, offering the juxtaposition as proof that technology firms have eclipsed sovereign states. This comparison is a category error. Market capitalization reflects speculative expectation; gross domestic product measures annual economic output. One captures a fluid snapshot of investor optimism; the other reflects a nation’s concrete productivity.
Stripping this analytical error away reveals a different, more staggering reality. Scale alone is not the problem. The critical issue lies in how few hands direct this volume of capital, and how little say the public has in its governance. A rigorous assessment of this power requires a precise unit of measurement.
The Real Weight
The honest unit of measurement is flow against flow: actual dollars spent against actual dollars spent. In 2026, the combined annual AI infrastructure spending of just four tech giants—Microsoft, Alphabet, Meta, and Amazon—is projected to reach between $650 billion and $725 billion. This is deployed capital, not Wall Street sentiment. This single-sector corporate outlay already dwarfs the entire US federal research and development budget, and it equals roughly three-quarters of the nation’s entire defense budget.
The speed of this spending reveals the sheer fury of the buildout. These same four companies spent roughly $410 billion on capital investments in 2025. Surging toward $725 billion the following year represents a year-over-year spike of more than 70%. This is a private industrial mobilization—not a gradual corporate expansion, but a near-doubling of an already historic spend in less than twelve months.
This single technological gamble is consuming the deepest reserves of wealth on earth. Consider Amazon: after spending $50.7 billion on property and equipment for AI, its annual spare cash plummeted from $38 billion to $11 billion. Morgan Stanley and Bank of America now project that Amazon will run between $17 billion and $28 billion in the red for 2026. A trillion-dollar digital empire is burning through its own historic revenue to fund a single infrastructure project. This is a state-scale commitment made by a boardroom.
In any prior era, an infrastructure project mobilizing $700 billion a year would have run through public budgets, scrutinized by voters and debated by lawmakers. Today, that world-altering allocation of resources is masterminded by a handful of executives. The scale of the money is historic; the smallness of the group directing it is the real crisis. This raises the immediate question of who, exactly, granted them the authority to aim state-scale capital.
The Inner Ring: Control Decoupled from Ownership
At Meta and Alphabet, two or three individuals hold voting control over some of the largest corporations on earth. This is not an accident of the stock market; it is the product of deliberate legal design. Across the artificial intelligence frontier, tech giants are using specialized legal tools to sever the right to make decisions from the actual ownership of the company.
The methods vary, but the result is identical: immense capital flows into these organizations, but control stays locked in the boardroom.
The Super-Voting Share: Meta and Alphabet
The cleanest tool is the “super-voting” share. At Meta, Mark Zuckerberg controls roughly 61% of the voting power despite owning only a 14% financial stake in the company. He achieves this through special Class B shares that carry ten votes each, compared to the single vote carried by regular shares.
This is not a theoretical quirk; it is a shield against outsiders. At Meta’s May 2026 Annual General Meeting, independent investors put forward ten separate proposals—including a push to establish a fair “one-share-one-vote” rule. Every single one was rejected. Zuckerberg simply outvoted them. Even when a massive majority of regular shareholders demands a check on power, the structure makes it impossible.
Google’s parent company, Alphabet, uses the same blueprint. Founders Larry Page and Sergey Brin retain 52.7% of the voting power while owning just 11.5% of the company. They command majority control long after stepping back from daily operations, and the rules cannot be changed without their explicit permission.
Private Absolute Control: xAI
This concentration can be pushed even further by avoiding the stock market entirely. Following their merger, xAI operates as a private wing inside SpaceX. While outside money might find its way into this ecosystem, it carries virtually no say. Control here is not a mathematical majority won through shares; it is total, private, and entirely insulated from public scrutiny.
The Nonprofit Shield: OpenAI
Control can also be detached from ownership through clever corporate conversions. Consider OpenAI’s restructured corporate layout. The nonprofit OpenAI Foundation technically sits on top, appointing the board of the OpenAI Group—a “public-benefit corporation” designed to balance profit with societal good. However, the nonprofit holds only about 26% of the equity, while Microsoft holds 27%.
In reality, a public-benefit corporation’s obligation to “do good” is entirely up to its leaders. As Harvard Law’s Roberto Tallarita and other corporate governance experts have shown, these structures have no legal teeth; outside observers have no way to enforce the company’s grand promises. It is a different mechanism that arrives at the same destination: control retained by a select few, disconnected from the underlying money.
The Mission Trust: Anthropic
We must apply the exact same critical lens to the tech world’s more idealistic structures. Anthropic relies on a “Long-Term Benefit Trust” to oversee its operations. This trust holds a special class of stock with zero economic value, and through it, five unelected trustees appoint three of the five seats on the corporate board. No investor—not even corporate giants like Amazon or Google—can override them.
This is the exact same species of structural insulation as Meta’s dual-class shares. Justifying the structure with the language of public safety rather than a founder’s personal vision does not change the mechanics. We cannot mistake a noble stated purpose for a democratic structure. A trust that cannot be overridden by its financial backers is, simply put, power insulated from capital.
Four different corporate instruments, all engineered toward a single reality: a tiny handful of people hold decisive authority over world-altering technologies, immune to the influence of the capital that funds them.
The urgent question that follows is whether these separate centers of power are truly independent of one another.
They are not.
The Interlock: Apparent Diversity Collapses
The idea that separate centers of power control AI requires a closer look. The reality is that the handful of people running these firms do not operate in isolation. The companies themselves are hardwired together. What looks from the outside like a fierce battlefield of independent rivals is, structurally, an overlapping web of alliances among the exact same actors.
The Tech Layer Cake
The entanglement begins with the physical infrastructure. Artificial intelligence is built like a skyscraper. At the foundation are the specialized computer chips; in the middle is the massive cloud data center infrastructure that houses them; and at the top are the labs training the actual models.
Instead of operating independently, each layer is financially fused to the others through massive equity stakes and multi-decade commitments. The labs do not simply buy parts from suppliers—they are corporate roommates.
Trace these hidden dependencies, and the illusion of rivalry dissolves:
OpenAI is heavily anchored to Microsoft, which owns a 27% stake and supplies its vital cloud capacity, alongside computing alliances with Oracle and SoftBank.
Anthropic runs on Amazon’s infrastructure fueled by investments scaling toward $25 billion, while simultaneously leaning on a reported $40 billion backing from Google.
xAI relies entirely on Nvidia as both a primary investor and its sole chip supplier, all while being tucked safely inside SpaceX.
Google and Meta do not need outside cloud partners because they are the cloud.
Ultimately, the question of who builds the models and the question of who owns the physical computing power resolve to the exact same short list of names.
The Circular Cash Machine
The clearest example of this closed loop happens inside a single financial alliance. The chip giant Nvidia is a major investor in a specialized cloud company called CoreWeave—and is also CoreWeave’s primary chip supplier. Under their multi-billion-dollar contract, Nvidia has actually promised to buy up any computing capacity that CoreWeave fails to sell. In essence, the chipmaker is acting as an insurance policy, guaranteeing demand for the very data centers its own chips fill.
To close the loop entirely, Microsoft alone accounts for 67% of CoreWeave’s revenue. Capital, hardware, and customers simply spin in a perfect circle.
Government regulators have already pointed out this pattern. In a formal industry investigation, the Federal Trade Commission documented this exact type of circular spending. They described a bizarre ecosystem where an investment returns directly to the investor as cloud revenue—which is then booked as “growth” to justify the next massive round of fundraising. The government described these cloud giants as occupying a triple position: acting as investor, supplier, and customer to the exact same firms all at once.
The Illusion of Choice
Count the brand names, and the market looks crowded. Count the actual decision-makers, and the crowd collapses into a handful. A dozen distinct storefronts resolve into a few tight clusters of capital and control, each bound to the others by financial stakes that make their fortunes mutual.
The diversity is real on the surface, but largely absent where the real choices are made. The same few people who control each firm are tied directly to the same few people who control the rest.
If this is the hidden structure of who decides, a final question remains:
Who has to live with the consequences of these decisions—and were we ever given a say in making them?
The Outer Ring: Consequences Without Consent
The International Energy Agency estimates that data centers consume about 1.5% of the world’s electricity today, a figure projected to hit 3% by 2030. On paper, that global growth looks modest—smaller, even, than the projected energy demand for electric vehicles or air conditioning.
But global averages lie. They conceal exactly where the heavy machinery actually lands. Nearly half of all US data center capacity is packed into just five regional clusters. In Virginia, these facilities already devour a staggering 26% of the entire state’s electricity. The global total is a useless metric. The real issue is who pays to keep the lights on when a power load is this concentrated.
Democracy Behind NDAs
You can only fight a cost if you know you are paying it. Most of the time, the public is left completely in the dark. In 25 of the 31 Virginia communities hosting these facilities, developers secured approvals behind strict nondisclosure agreements. These NDAs go far beyond typical corporate privacy. They deliberately hide the true scale of the project, and they conceal the tech giant’s identity from the very residents who will live next door to the facility. What the secrecy protects isn’t proprietary tech—it is the decision itself.
This boardroom secrecy is the industry norm, not the exception. In March 2026, Senator Richard Blumenthal launched a 50-state investigation into these exact corporate tactics. A parallel analysis by NBC News of 30 massive national data center projects revealed that a majority relied on this exact same wall of silence. In case after case, developers routed their deals through anonymous shell companies, keeping local communities completely unaware until the contracts were signed and the land was sold.
The Cost Mismatch
Naturally, public anger is rising fastest where transparency is lowest. A recent Washington Post and Schar School poll found that 59% of Virginia voters are now uncomfortable with having a data center built near their homes—a massive spike from just 24% a few years ago. The gap between public anxiety and closed-door deals is widening by the day.
This creates a dangerous structural mismatch. Local town and county boards hold the sole legal authority to approve these massive projects, but they don’t have the tools or the scope to evaluate statewide grid failures or regional water use. A tiny local government gets an immediate windfall of tax revenue for its local budget, while the broader regional grid forces everyday citizens to shoulder the massive infrastructure costs on their monthly utility bills.
The body that profits most from signing the deed understands the downstream costs least. A government entity too small to weigh the true cost signs the approval, and the public learns the truth long after the concrete is poured.
This is the accountability gap operating at the level of a single county. And this exact same pattern of exclusion repeats at every single level of human oversight.
The Gap: What No Institution Is Sized to Close
The secrecy at the county level is just the smallest visible instance of a total systemic failure. Zoom out to the national institutions that are supposed to catch what local governments miss, and the exact same mismatch appears. A democratic society traditionally relies on regulatory agencies and elections to check concentrated corporate power. Today, those channels are failing.
The failure exists across all three avenues of public accountability:
The Regulatory Apparatus is Obsolete: Our antitrust tools were built for an era of railroads and oil monopolies. When the UK Competition and Markets Authority examined the Microsoft-OpenAI partnership in 2025, they found heavy influence—but because Microsoft didn’t formally buy OpenAI outright, regulators had no legal authority to intervene. Former EU competition commissioner Margrethe Vestager had warned that deals structured this way can disguise control and evade scrutiny. The FTC has documented this circular spending, but observation is not enforcement. Our legal instruments are completely blind to the modern architecture of influence.
The Legislature is Paralyzed: Congress remains entirely deadlocked. When a proposed state-level moratorium died in July 2025, and subsequent federal frameworks collapsed, the executive branch stepped into the vacuum with Executive Order 14365. Accountability is now dictated entirely by executive fiat and endless litigation, aligning the government with specific industry camps. The legislative deadlock is, in itself, the accountability gap.
The Electoral System is Priced Out: The electoral channel has been hijacked at a scale the public cannot match. AI super-PACs have poured hundreds of millions into the current election cycle. The industry doesn’t even uniformly oppose regulation—Anthropic recently funded a pro-regulation campaign with $20 million. But that is the deeper problem, not a defense. Anthropic spending $20 million in “safety” language is the exact same maneuver as rival PACs spending $140 million in “innovation” language. The public is simply forced to watch a proxy war fought between near-trillion-dollar entities.
We are left with a structural void. In the inner ring, legal loopholes and financial interlocks collapse a supposedly diverse market into a handful of decision centers. In the outer ring, the consequences of those decisions are dumped on the public.
No existing institution closes the loop between them. The regulatory bodies cannot reach the mechanisms of control, the legislature is paralyzed, and the electoral arena is bought by the very actors it is meant to bind.
No single villain planned this failure. This architecture of global consequence is the accidental result of a thousand separate, selfish choices. Executives chose super-voting shares; developers chose non-disclosure agreements. The result is a society whose oversight mechanisms no longer reach the power they were built to check.
Whether a new institution capable of checking this power ever forms is currently being decided by the very people it would govern.
Control remains insulated from ownership. Consequence remains insulated from consent. And the gap between them remains insulated from every channel built to close it.
The figures in this essay are tracked and updated quarterly here: AI Power — Live Figures.