Intelligence Has Become Infrastructure. The Question Now Is Who Gets to Own It

Every era inherits infrastructure whose ownership settles arguments about power for decades, from railway gauges to electrical grids to undersea cables. Artificial intelligence took this position in under a decade, and its terms of ownership are being fixed now, largely out of public view.

Public debate keeps circling back to humans against machines, obscuring the division that will shape the 2030s. Intelligence held by a handful of operators behaves very differently from intelligence anyone can obtain, and the gap shows up in economics, politics and knowledge.

Growth Has Stopped Needing Crowds

For most of industrial history, ambition required crowds, since a manufacturer doubling output hired more hands, a railway needed surveyors and navvies by the thousand, and production stayed tethered to participation.

Something in that arrangement has come loose, because building anything formidable now takes a small group of exceptional engineers alongside enormous capital and computation, and the ratio keeps tilting toward the second.

Lovable reached $400 million in annual recurring revenue with 146 employees, and Midjourney earns around $200 million a year with a team of about eleven people. Set that beside the $660 billion to $690 billion in capital expenditure five American technology companies have committed for 2026, most of it for chips, data centres and networking.

Labour has slipped down the list of binding constraints as hardware has climbed to the top, which reprices the whole stack for anyone allocating capital. Which side of that dependency is a portfolio on? Returns gather around accelerators, power contracts and land near substations, and almost every other business now runs on terms set by a few suppliers.

Concentration Without the Crowd

If a group can raise its capability without employing or trading with a large population, it has thin reasons to bring that population along. Wealthy actors of earlier centuries needed workers, customers and soldiers, and that dependency produced wages, schools, hospitals and political leverage.

Remove the dependency and the incentives underneath change quietly. Picture a thousand people, or a million on the outside, holding capital, models and robotics. They could pursue enormous projects, from orbital manufacturing to Mars, with no functional need for the billions outside the room.

Whatever still calls for a crowd increasingly gets answered by machinery instead, because a fleet of robots costs less than a workforce once a task is specified precisely enough. Every job that once needed many hands is priced against that substitute, and the promise that robotics and AI will handle our problems assumes somebody owns the robots.

Such clusters would drift from everyone else without dramatic rupture, and artificial general intelligence is not a prerequisite, since today’s capabilities, extended and cheapened, would do.

Access to the strongest models may end up a sharper source of inequality than money, because capital compounds at a few percent a year, whereas an advantage in model quality compounds through every piece of research, code and product it touches.

Influence That No Market Price Captures

Concentrated intelligence shapes what people believe as much as it shapes markets. Weekly use of AI chatbots for news climbed from 7% to 10% globally in a single year, and more than half of 18-to-24-year-olds now name social media, video networks or chatbots as their main way of learning what happened.

Search engines wielded editorial power through ranking, but they point outward to many resources, and a reader still meets a source and judges it. An assistant hands over a conclusion in a calm, authoritative voice, and arguing with it takes effort few readers will spend.

When a limited number of providers decide what a model says, an agenda can be embedded in a form close to undetectable from outside. Growing capability makes detection harder still, since a stronger model produces more persuasive reasoning and more plausible mistakes. How would anyone outside the company notice?

Scepticism has not vanished, in fairness, since trust in AI-delivered news sits at 20% globally, below the 37% recorded for news overall, and audiences stay wary even as their reliance deepens.

Open Weights Close Half the Gap

Published models genuinely improve the picture, and the newest are strong. Moonshot AI released the weights of Kimi K3 at 2.8 trillion parameters, bringing open releases closer to the frontier than at any point since DeepSeek R1.

Serving one of those models is another matter for anyone outside a data centre. A mixture-of-experts system on the scale of K3’s scale needs a cluster of accelerators kept resident and warm, which puts self-hosting beyond most universities, startups and researchers.

Most people consequently reach open models through a hosted endpoint owned by a company with capital and GPUs. The company resells access, and customers cannot confirm that the weights served match the weights published, or that nothing sits between prompt and answer. What has been opened up in that arrangement, beyond the licence itself?

Ownership Begins at the Hardware Layer

Licence terms alone fail to distribute power, because hardware decides who can exercise the rights a licence grants. Owning therefore begins where the licence ends, at the machine itself, since whoever runs the model is the only party able to test it, trace an answer back to the weights that produced it, and correct what comes out wrong.

Each of those depends on access to enough computation at a price ordinary institutions can absorb, so a public ecosystem needs open weights and open compute together. Either one alone recreates the concentration it sets out to dissolve.

A Public Compute Layer Is Economically Plausible

Capable hardware already sits idle in enormous quantities across universities, regional data centres, former mining operations and private hands, and pooling it has been a hard engineering problem.

Covenant-72B, a 72-billion-parameter model, was pre-trained across more than seventy independent peers over commodity internet links, with no whitelist, no central cluster and a blockchain mechanism validating contributions as they arrived. Verification sits at the heart of that design, since participants nobody has vetted must still prove the work they submit is real.

Efficiency remains the honest objection, because a swarm on ordinary connections still trails a tightly coupled cluster, latency penalises the largest runs, and decentralised systems retain a habit of re-concentrating wherever capital gathers.

Distance to this point is closing quickly, though, since the Covenant run held 94.5% hardware utilisation under real networking constraints and produced a model competitive with centrally trained peers of similar size. Spare capacity costs less than new build, which leaves the economics resting on engineering maturity more than on arithmetic.

Decisions Made by Default

So who decides which future arrives, given that nobody will ever vote on the question? Outcomes will accumulate from procurement choices, licences, research funding and regulation made over the next few years, most individually unremarkable.

One road leads to a few providers operating the substrate of economic and intellectual life, answerable to shareholders and whichever governments they depend on. Another leads toward something closer to a utility, with open models running on compute that many parties own and anyone can audit. And what keeps the second road open?

Money and procurement, mostly, since every contract signed for hosted intelligence and every grant directed at shared infrastructure nudges the balance. Both remain reachable from 2026, and the window for choosing deliberately is measured in years – not many of them.