Democratic theory has always assumed that power, to be legitimate, must be accountable to those it affects. Artificial intelligence is in the process of making that assumption look quaint. The technology is already governed — by billionaires, national security agencies, and backroom frameworks that decide who gets access to the most powerful cognitive tools in history. What remains to be seen is whether democracy can reassert itself before the architecture becomes permanent.

AI is already governed. That is the fact that tends to get buried beneath the breathless debate about regulation, safety frameworks, and democratic oversight. The technology is not sitting in a neutral waiting room, patiently holding out for society to decide what to do with it. It is being actively shaped, deployed, and monetised by a remarkably small number of people — people who also happen to have direct lines to heads of state, the ability to manipulate the information ecosystems through which citizens form political opinions, and a financial interest in ensuring that no meaningful democratic check on their power ever materialises. The question was never whether AI would be governed. It is simply who gets to do it.

The answer, right now, is a handful of tech billionaires and the governments that court them. Zuckerberg, Musk, and Altman are not your everyday entrepreneurs. They are political actors of the first order, operating surveillance and communication infrastructure at a scale that no elected government in history has commanded.

Musk reshaped a major global information platform to serve his political agenda, then walked into the American executive branch without a single vote cast in his favour. Altman conducts negotiations directly with heads of state while his company remains fundamentally tied to government contracts for its survival. Zuckerberg, having spent years cultivating the image of a bumbling liberal technocrat, has since declared that Silicon Valley needs more “masculine energy” and “aggression.”

They reflect a coherent worldview, one that sees democratic deliberation as friction, and the infrastructure they control represents a concentration of power that dwarfs anything a conventional political party could dream of assembling.

The most seductive argument against challenging this arrangement is also the most intellectually dishonest one. It runs roughly as follows: restricting AI infrastructure will raise prices, price out poorer users, and entrench a digital divide that benefits the already privileged. Slow AI down, in other words, and you end up hurting the very people you claim to be protecting.

It is a compelling framing, partly because the underlying concern about unequal access is legitimate. AI can provide tools that genuinely help people who cannot afford professional alternatives. Pretending otherwise would be foolish.

Critically, however, access to a tool is an entirely different thing from its governance. A citizen who can afford a ChatGPT subscription has no more say in how that system is trained, what values it encodes, whose data it harvests, or how its outputs shape electoral discourse than someone who cannot.

The poverty premium is real, and it is already worsening, but framing democratic resistance to that consolidation as an attack on poor people’s access is, at best, a category error. At worst, it is a strategy for neutralising opposition by weaponising progressive language against progressive goals. Palantir, a surveillance technology firm with deep ties to the Trump administration, made precisely this argument against data centre regulation.

The access argument has also been overtaken by events in the most revealing way possible. Days ago, the Trump administration signed an executive order establishing a “trusted partners” framework, through which select organisations will receive early access to the most powerful frontier AI models before general release. The criteria for earning that designation remain entirely undefined, left to the NSA and a constellation of national security agencies to determine.

Anthropic has already restricted its most advanced model to an approved group of organisations, personally briefing senior government officials on its capabilities. OpenAI followed, making its most powerful systems available exclusively to vetted government bodies. The architecture of a two-tier AI system — one for the trusted, one for everyone else — is current policy, with the people promising that AI will democratise access simultaneously building the infrastructure to ensure it does not.

This is what the fusion of state and corporate power looks like when it stops bothering to disguise itself. The political class and the tech industry are not allied in any complicated or ambiguous sense. They are, functionally, the same formation. Elected politicians, rather than checking that power, compete for proximity to it.

Democratic governance of AI will therefore develop through pressure rather than permission. The struggle against data centres, although messy and ideologically diverse, represents the most honest version of that pressure currently available. Working-class communities, indigenous groups, farmers, and union members have been showing up across the United States to contest infrastructure projects negotiated behind closed doors, with no public input and almost no transparency.

None of this is sufficient on its own, and treating it as a finish line would be a mistake. These fights matter because they are building the only thing capable of eventually producing genuine democratic governance: organised, distributed power among people who have been deliberately excluded from every room where the decisions are made.

The elite governance of AI is a deliberate architecture, constructed by people with the resources to buy legislators, the access to shape executive policy, and now the formal authority to decide who gets to use the most powerful cognitive tools in human history. Contesting it requires treating it as the political project it is. The algorithm already speaks for all of us. The question is when we start speaking back.