The New Standard Oil: How AI Is Concentrating Wealth Like the Robber Barons

In 1870, John D. Rockefeller controlled roughly four percent of American oil refining capacity. By 1880, he controlled ninety percent. The mechanism of his ascent was not primarily technological superiority — Standard Oil’s refineries were efficient, but not astonishingly so. It was control of infrastructure. Rockefeller negotiated secret rebates with the railroads that shipped his oil, meaning he paid less per barrel than his competitors, and competitors’ payments partially subsidized his own costs. He controlled pipeline networks that competing refiners depended on. He used predatory pricing — temporarily selling below cost in specific markets — to force smaller competitors to either sell to him or go bankrupt.

The result was a monopoly so complete that Congress was eventually forced to create entirely new categories of law to address it. The Sherman Antitrust Act of 1890, and the subsequent breakup of Standard Oil in 1911, were not abstract exercises in economic theory. They were responses to a specific, observed, destructive concentration of economic power in a technology that had become essential to industrial civilization.

The AI industry is not Standard Oil. But the economic forces creating winner-take-all dynamics in AI rhyme with the forces that created Standard Oil’s dominance with uncomfortable precision.

The first and most important structural similarity is the infrastructure control dynamic. Standard Oil’s dominance was not primarily about having better oil; it was about controlling the physical infrastructure through which oil moved. The railroads and pipelines were the chokepoints, and Rockefeller controlled them.

In the AI era, the equivalent infrastructure is compute — the data centers, the GPU clusters, the networking, the power connections. Training frontier AI models requires capital investment on a scale that only the largest technology companies and, increasingly, sovereign wealth funds and governments can contemplate. A single training run for a competitive frontier model can cost hundreds of millions of dollars in compute alone. Deploying that model at scale requires data center infrastructure worth billions more. The companies that already control cloud computing infrastructure — Microsoft Azure, Amazon Web Services, Google Cloud — have an enormous structural advantage because they own the substrate on which AI runs.

This is not a transient advantage that will erode as the technology matures. If anything, it compounds. The companies with the most compute can train the best models. The best models attract the most users. More users generate more data, which can improve subsequent models. The revenue from users funds more compute investment. The cycle reinforces itself in a way that makes the early leaders progressively harder to challenge.

The second structural similarity is the role of data as a proprietary input analogous to Rockefeller’s access to cheap railroad capacity. Language models improve with more and better data, and the companies with the largest existing user bases have access to data that no competitor can replicate. Google has decades of search queries, email content, and document editing patterns. Meta has detailed information about human social connections, communication patterns, and content preferences for billions of people. Microsoft has access to corporate document creation and communication patterns through Office. Amazon has detailed purchasing and logistics data.

A new AI company — even one with excellent researchers and access to capital — cannot acquire this data by building better technology. The data is a consequence of existing at scale, of having already won a previous competitive battle. It is proprietary in the deepest sense: not secret, necessarily, but simply unavailable to anyone who was not already large enough to collect it. This creates an asymmetry between incumbents and new entrants that becomes more pronounced as AI systems improve, because better AI extracts more value from the same data, which means the incumbents’ existing data stocks become more valuable over time.

The third structural similarity is the network effects that make AI services self-reinforcing. Standard Oil’s pipeline network became more valuable as more producers connected to it, and switching to an alternative network became progressively harder because the alternative had fewer connections. Modern AI services exhibit similar dynamics, though through different mechanisms. A coding assistant that has been trained on and used by millions of developers is better at helping with real-world codebases than one trained in isolation. An AI model embedded in a productivity suite that employees use daily generates usage data that improves it for those specific productivity tasks. The improvement from use creates lock-in, and lock-in creates durable competitive advantage.

The antitrust challenge posed by these dynamics is genuine, and the existing legal frameworks are poorly equipped to address it. Standard Oil’s monopoly was legally addressable because it depended on explicit contractual arrangements — the railroad rebates, the pipeline agreements — that could be identified, enjoined, and restructured through the judicial system. The competitive advantages of today’s AI incumbents are not primarily contractual. They emerge from the size of training datasets, the scale of compute infrastructure, and the self-reinforcing dynamics of network effects. You cannot break these up in the way that the Supreme Court broke up Standard Oil into thirty-four companies, because the competitive advantage is embedded in accumulated assets and relationships that do not divide cleanly.

What would breaking up Big AI actually require? This is a question that antitrust theorists have been grappling with, and the answers are not encouraging in their simplicity. Data portability requirements — forcing AI companies to give users their data in transferable formats — help at the margins but do not address the core problem, which is that the value of data comes from its volume and diversity, not just from any individual user’s contribution. Interoperability requirements — forcing AI services to work with each other’s interfaces — could reduce switching costs, but the competitive advantage of better models trained on more data remains.

The most aggressive proposals involve either structural separation — prohibiting companies that control AI infrastructure from also operating AI services, in the same way that laws have historically separated banking from commerce — or treating certain AI capabilities as public utilities, requiring non-discriminatory access at regulated prices. Both approaches face serious practical and legal obstacles, and both represent more interventionist uses of regulatory power than Western governments have been willing to contemplate in the technology sector.

There is a version of the Standard Oil analogy that should be taken more seriously than it typically is: the eventual outcome. Standard Oil’s dominance lasted roughly four decades before being broken up by government action. The breakup was not a free-market outcome — oil markets were not trending naturally toward competition. It was a political decision, made by a government that decided the concentration of economic and political power in a single company had become incompatible with democratic governance.

The question for AI is whether something analogous will happen, and if so, when. The political conditions for aggressive antitrust action are building slowly. The public discourse about tech monopoly power has shifted dramatically in the past decade. Regulatory agencies in the US and Europe have become more willing to challenge large tech mergers and to use existing antitrust tools more aggressively. Politicians across the political spectrum have found common ground in tech skepticism, though for very different reasons.

But political will is different from legal capacity. The antitrust tools available to regulators were designed for an industrial economy. Applying them to AI requires either creative legal interpretation — which courts may or may not accept — or new legislation — which is slow and subject to the same regulatory capture dynamics that have slowed AI regulation broadly. Rockefeller’s monopoly took forty years and an entirely new body of law to address. There is no particular reason to expect the AI monopoly problem to resolve faster.

The pattern repeats. The technology changes. The economic logic of winner-take-all infrastructure control is remarkably stable across a century of technological revolution. That is the most important thing the Standard Oil comparison tells us: not that the outcome is inevitable, but that the forces are structural. Resisting them requires deliberate, sustained political effort — and knowing from history what the alternative looks like.