When the CEO of the world's largest asset manager warns that AI could deepen wealth inequality, it is not speculation. It is positioning. Larry Fink's 2026 annual letter to BlackRock investors laid it out plainly: "There's a real risk artificial intelligence could widen wealth inequality if ownership does not broaden alongside it." He described a world of "K-shaped outcomes," where firms and investors with capital accelerate upward while everyone else stagnates. Coming from someone who manages over $11 trillion in assets, this is not a hot take. It is a risk assessment. But Fink's framing, while correct, only captures half the picture. The ownership gap matters, yes. But there is a second divide forming that is harder to see and harder to fix. It is not about who owns AI companies. It is about who can direct AI, and who gets directed by it.

The orchestrator gap

The real split is not between people who use AI and people who do not. Almost everyone will use AI eventually, the same way almost everyone uses the internet. The split is between people who orchestrate AI and people who are orchestrated by it. Orchestrators set direction. They decide what problems to solve, evaluate whether the output is good, and build systems that compound over time. They treat AI as infrastructure, not a novelty. The orchestrated follow AI-generated task lists. They are evaluated by AI-driven metrics. They interact with AI the way a factory worker interacts with an assembly line, as a component in someone else's system. This distinction matters because it determines who captures value. The orchestrator builds leverage. The orchestrated becomes interchangeable.

Jensen's math

Nvidia CEO Jensen Huang recently put a number on this. Speaking on the All-In Podcast in March 2026, he said he would be "deeply alarmed" if a $500,000 engineer did not consume at least $250,000 worth of AI tokens per year. Asked whether Nvidia is trying to spend $2 billion on tokens across its engineering org, his answer was short: "We're trying to." The logic is straightforward. A highly paid engineer who can direct fleets of AI agents, review their output, and integrate it into production systems could be ten times more productive than one working manually. That is not a productivity gain in the traditional sense. It is a concentration of value into fewer, higher-leverage individuals. Huang even floated the idea of giving engineers an annual token budget as part of their compensation, the same way companies provide equipment or education stipends. "Engineers will soon be asking, 'how many tokens come along with my job?'" he said at GTC 2026. This reframes the economics of knowledge work. The scarce resource is no longer the person's time. It is their ability to deploy compute effectively. And that ability is not evenly distributed.

Intelligence is cheap, infrastructure is not

There is a popular narrative that AI is being democratized. In some ways, it is true. Free tiers exist. Open source models are getting remarkably capable. Anyone with an internet connection can talk to a frontier model. But the gap between free-tier AI and enterprise AI is widening, not shrinking. OpenAI's 2025 State of Enterprise AI report found that frontier workers, those at the 95th percentile of usage, send roughly six times more messages per seat than the median enterprise user. For coding tasks, the gap is seventeen times. Users who engage across seven or more task types report saving five times more hours than those who stick to one or two use cases. This is not because frontier users have access to a secret model. They are using the same systems. The difference is how deeply AI is wired into their daily work. They have the tooling, the workflows, the compute budgets, and the organizational support to make AI a multiplier rather than a toy. Anthropics's research tells a similar story at the macro level. Their AI Usage Index found that for every 1% increase in GDP per capita globally, there is approximately a 0.7% increase in AI usage. Within the United States, the correlation is even steeper: a 1% higher per capita GDP corresponds to a 1.8% increase in AI use. API costs, compute budgets, and tooling access are becoming the new class markers. The raw intelligence is commoditized. The infrastructure to deploy it at scale is not.

The democratization paradox

The democratization narrative is not wrong, it is just incomplete. Yes, a teenager in Lagos can use ChatGPT for free. But that same teenager cannot fine-tune a model on proprietary data, deploy an agent fleet across enterprise workflows, or access the kind of dedicated compute that makes AI a genuine productivity multiplier. The most valuable AI companies are also staying private longer than the tech giants of previous eras, as Fink pointed out. This means everyday investors are locked out of the sector's most explosive growth phase. The wealth generated by AI accrues to a shrinking circle of founders, early employees, and institutional investors. Open source helps. It genuinely does. But it does not close the gap. Running a state-of-the-art open model requires hardware, expertise, and time that most people and most companies simply do not have. The gap between "access to AI" and "ability to leverage AI" is where inequality lives.

The Singapore question

Small, wealthy, tech-literate nations like Singapore might seem well-positioned for this shift. High education levels, strong digital infrastructure, a government that moves quickly on technology policy. On paper, these are the conditions for broad AI adoption. But there is a catch. If the infrastructure layer, the foundational models, the compute clusters, the data pipelines, is controlled primarily by the US and China, then even well-positioned small nations become dependent consumers rather than sovereign builders. The question is not whether your population can use AI. It is whether your institutions can shape how AI is deployed, and who benefits. This is a version of the same orchestrator-versus-orchestrated dynamic, just playing out at the national level.

The knowledge gap is the real gap

You do not need the most expensive tools to benefit from AI. A thoughtful person with a $20/month subscription who understands prompting, chaining, and tool selection can outperform someone with a $200/month enterprise plan who treats AI like a search engine. But that knowledge, knowing which tools exist, how to combine them, when to use which model, how to evaluate output quality, is itself unevenly distributed. It correlates with education, professional networks, and the kind of work environment that encourages experimentation. The frugal optimizer can do a lot with a little. But frugal optimization requires a baseline of knowledge that many people do not have and that no free tier provides.

What this means

AI does not create inequality. It accelerates whatever inequality already exists. If you already have capital, AI helps you deploy it more efficiently. If you already have expertise, AI amplifies it. If you already have access to information networks, AI makes you faster at extracting value from them. Fink is right that broadening ownership matters. But ownership alone is not enough. The deeper challenge is building a world where more people can be orchestrators, where the ability to direct AI is not gated by wealth, geography, or employer. That means investing in AI literacy, not just access. It means rethinking education to focus on judgment, evaluation, and system design rather than rote execution. It means recognizing that the most important skill in an AI-saturated economy is not using AI. It is knowing what to use it for. The gap is real, it is growing, and the window to address it is shorter than most people think.

References