The Question

A vast server farm stretching to the horizon, lit in blue, with a single worker in the foreground dwarfed by the scale

In 1800, if you owned land, you owned the economy. The farm, the rent, the food supply — it all flowed from the ground beneath your feet. By 1900, the game had changed. Factories and railways were the new land. The Carnegies and Rockefellers understood this before almost anyone else, and they got fabulously wealthy while millions of workers who powered those machines stayed poor. The pattern is not coincidence. It is a structural feature of how economies work.

We are at the beginning of a third version of this story. The means of production in the AI era are not acres or assembly lines. They are data — the vast archives of human text, images, and behaviour used to train AI systems — and compute, the specialised computer chips and data centres needed to run them. Both are already extraordinarily concentrated. And the gap between those who own them and those who do not is growing faster than anything the industrial era produced.

What the Evidence Shows

The numbers are stark. As of 2026, three companies — Microsoft, Google, and Amazon — control roughly 65% of global cloud computing infrastructure. Nvidia supplies approximately 80% of the specialised AI chips (called GPUs) that power the training of large AI models. OpenAI, Meta, and Google DeepMind hold the largest proprietary training datasets and the most capable AI models. The entire AI value chain, from chip to model to cloud deployment, flows through a handful of corporations whose combined market capitalisation exceeds the GDP of most countries.

What does this mean in practice? It means that when AI automates a task — writing, coding, customer service, logistics planning, medical imaging, legal research — the productivity gains do not flow to the workers whose jobs were displaced, nor to the customers who use the service. They flow as profit to the shareholders of the companies that own the AI. In a traditional factory, workers could unionise, strike, and negotiate a share of productivity gains. AI does not unionise. The gains accumulate silently at the top of the ownership structure.

"We may be approaching a moment where the economy divides not between capital and labour, but between those who own AI and those who are replaced by it."

— IMF — "AI and the Future of Work" — International Monetary Fund, 2024

Early data supports this concern. Between 2020 and 2025, the combined wealth of the owners of major AI infrastructure companies grew by an estimated $4 trillion. Over the same period, wage growth for the bottom 60% of earners in the United States, adjusted for inflation, was negative. The two trends are not unrelated.

"Every era has its means of production. Whoever owns them owns the future — and right now, almost no one owns AI."

Why This Is Happening

AI has extreme economies of scale. The cost of training a frontier AI model is measured in hundreds of millions of dollars. Only the largest companies and the most well-funded startups can afford to build at the frontier. Once built, however, a model can serve millions of users at near-zero marginal cost. This creates winner-take-most dynamics even more extreme than those in social media or search — markets that already produced some of the most concentrated wealth in history.

Data is the new land — and it was enclosed before most people noticed. The training data that powers today's most capable AI systems was harvested, largely without compensation, from the internet — from books, articles, social media posts, code repositories, and images created by billions of people over decades. That data is now locked inside proprietary models owned by private companies. The people who created it received nothing. The companies that captured it are now worth trillions.

Governments have been slow to respond. Antitrust law was designed for an era of physical monopoly — Standard Oil controlling pipelines, AT&T controlling telephone lines. The mechanisms for breaking up a data monopoly or taxing a model's productive output do not yet exist in most legal systems. By the time legislation catches up, the concentration may be irreversible in practical terms.

What Could Happen

AI Ownership Concentrates, Gap Widens Sharply Most likely

By 2032, a handful of AI infrastructure owners capture a growing share of global economic output as AI automates more labour and the productivity gains accumulate at the top. The gap between AI owners and non-owners becomes the defining economic divide of the era — larger, in relative terms, than anything the industrial revolution produced. Most people are users of AI, not owners of it, and the economic terms of that arrangement are set entirely by those who own it.

Public Ownership Models Emerge Possible

Democratic governments — pushed by a combination of public anger and economic evidence — begin treating AI infrastructure as a public utility. A European AI commons, publicly funded national compute resources, and mandatory data-sharing regimes begin to distribute AI access more broadly. Sovereign wealth funds and national AI programmes in countries like Singapore, the UAE, and France start to build competitive infrastructure outside the US-China duopoly. Access widens, concentration slows. The gap still grows, but more slowly.

Open-Source AI Democratises Access Less likely

Capable open-source AI models — already developing rapidly through projects like Meta's Llama family — become powerful enough to match proprietary systems for most commercial applications. The cost of running AI falls to the point where small businesses, individuals, and governments in developing countries can deploy genuinely capable AI without depending on the major platforms. The infrastructure gap persists but its economic consequences are limited because the models themselves are freely available. This is possible; whether it happens at sufficient scale by 2032 is the key uncertainty.

Our Assessment
We assign 82% probability — highly likely that by 2032, ownership of AI infrastructure will be the defining economic divide, creating a wealth gap larger than the industrial era's gap between landowners and labourers. The structural case is overwhelming: extreme capital requirements create barriers to entry that almost no individual or small organisation can clear; data network effects entrench incumbents; and the legal and regulatory frameworks that might redistribute the gains do not yet exist. The open-source scenario is the most credible check on this outcome — but open-source models still require compute to run, and compute remains concentrated. The gap is not inevitable in its scale, but its existence is. The question before democratic societies is whether to treat AI infrastructure as they eventually treated electricity and roads — as something too important to be entirely privately owned.

What Can We Do

Community members gathered around shared computers in a public library, learning to use AI tools together

The concentration of AI ownership is not a force of nature. It is the result of policy choices — some made deliberately, many made by default. It can be shaped differently.

Understand what you are actually giving away. Every time you use a free AI tool, your interactions are typically used to improve the model — for free, at scale, for the benefit of the company that owns it. Knowing this does not mean you should stop, but it means you should push for reciprocity: data rights legislation, compensation frameworks, and transparency about how your inputs improve their products.

Support public AI infrastructure investment. The argument for public compute — government-funded AI infrastructure accessible to universities, small businesses, and public institutions — is the same argument that justified public roads, public libraries, and public broadband. When you vote, this is a legitimate item to weigh. Politicians who treat AI infrastructure as a pure private-sector matter are making a structural choice about who benefits from the technology.

Pay attention to antitrust cases involving AI. The legal battles over whether Microsoft's investment in OpenAI, Google's dominance in AI search, and Amazon's control of cloud AI infrastructure constitute monopolistic behaviour will shape the competitive landscape for decades. These cases are dry and technical. Their outcomes are not.

Invest in AI literacy, not just AI tools. The people best positioned to benefit from AI are those who understand it well enough to deploy it strategically — not just use it passively. AI literacy is becoming a form of economic capital. Pursuing it, and advocating for its inclusion in schools and workplace training, is one of the highest-leverage actions available to individuals.

Back open-source AI projects. Meta's open-sourcing of its Llama models, Mistral AI's open releases, and the broader open-source AI community represent the best near-term check on proprietary concentration. Using, promoting, and where possible funding these projects is not just a technical preference — it is a structural one with economic consequences.

Sources
  • IMF — "AI and the Future of Work" — International Monetary Fund, 2024
  • Synergy Research Group — Cloud Infrastructure Market Share — 2025
  • Oxfam International — "Inequality Inc." — Oxfam, 2024
  • OECD — "Artificial Intelligence in Society" — OECD, 2023
  • Acemoglu D. — "The Simple Macroeconomics of AI" — NBER Working Paper, 2024
  • Forecast The World Research Desk — 800+ data sources