The Commanding Heights of machine labor

On 29 October 2026 Will Manidis and Joe Roberts published The Commanding Heights, a 40,000-word essay on the question everyone building in AI is asking: when intelligence gets cheap and everywhere, who keeps the gains? Not the labs, they argue, but whoever holds the positions that "set the terms upon which that intelligence becomes economic activity." We agree, and would add that one of those terms gets set wherever machine work is bought, judged and paid, which is the layer Taskmarket builds.

TL;DR

• The Commanding Heights argues that the value of machine intelligence will not be captured at the model layer. As intelligence spreads and gets cheaper, the gains settle in durable positions above it.
• Above the model, intelligence turns into revenue one priced, accepted task at a time. A task market is where that happens, which makes it a natural place for value to accrue in the machine economy.
• Three of the essay's predictions (a fourth settlement of the firm, private arbitration and underwriting, and a digital dollar) describe a market Taskmarket already runs.

What the essay argues

Manidis and Roberts start from history. Gutenberg went insolvent after printing his Bible. Pennsylvania's oil was banked in Cleveland and New York. The essay cites the estimate that railroads kept only 7 percent of their social return in 1890. Each time, the inventors were not the ones who kept the gains.

They call today's model builders "Genesis", the industry that designs, trains and serves models, and expect it to keep growing. But they split intelligence itself into tiers, and the broadest one, commodity intelligence, is priced at the cost of compute. In their account, open-weight models from China are a large part of why. The lasting value goes to what they call the Commanding Heights: "scarce, controllable positions of economic power whose holders durably capture the gains of increasingly capable and diffused intelligence."

Scarcity and control matter, but the test they lean on hardest is time. "What characterizes the Commanding Heights is durability," they write, and a fleeting bottleneck does not qualify.

Value moves above the model layer

The phrase that matters in the essay is "becomes economic activity". On its own, intelligence is a model answering a prompt. It becomes economic activity when someone says what they want, someone delivers it, someone checks it and someone gets paid.

If raw intelligence sells at the cost of compute, the margin cannot live in the tokens. It moves up, to the point where a model's output turns into a result a buyer will pay for. That unit is the task: a specified piece of work at a known price, accepted or not. The cheaper and more diffused intelligence gets, the more of the economy runs through that unit, because every agent, model and person selling work has to cross it to get paid.

A Taskmarket task is built around that crossing. The brief, the reward and the settlement terms are written down, and they lock the moment a worker commits. The reward sits in stablecoin escrow before anyone starts. A named evaluator decides whether the work is done, and its record of verdicts, appeals and overturns is public. Payment releases on acceptance, and the whole history lives onchain. Once the models are interchangeable, that layer, where the work is priced, judged and settled, is where the machine economy's value collects.

The fourth settlement of the firm

The essay's chapter on the firm starts where our own essay does, with Ronald Coase: companies exist because using the market is expensive. Manidis and Roberts trace three settlements of the firm, owner-operators, then professional managers, then shareholder discipline, and argue that intelligence is creating a fourth. "The role of management itself will be replaced by the fusion of operators and capital." More bluntly: "Labor leaves the firm."

That is the zero-employee company: a few operators with capital, directing work done by models and by people outside the firm. The essay leaves open where that labor goes once it leaves. We think it goes to a market with escrow, verdicts and appeals built in, so a company of one can buy work from anyone and trust what comes back. The five-step playbook shows the setup.

Judges, underwriters and accountable humans

The essay predicts that private arbitration forums will take over work that courts used to do, and imagines "a machine intelligent Lloyd's of London": "a market of underwriters pricing risk and refusing whatever they cannot profitably write."

Something like it already runs on Taskmarket. Evaluators publish a fee, a window and an optional stake, and build a public track record with every verdict. Dispute resolvers hear appeals. The contract enforces the settlement terms, and timeouts mean no party can stall a task forever. Judging machine work has become a job with a fee, a stake and a reputation to lose, and that is the start of an underwriting market for machine labor.

The essay also expects that, in licensed work, "a liable human must remain responsible for the output." A task market already has a seat for that person. The evaluator is a named wallet, and it can belong to a licensed professional who signs off on work the agents did.

Paid in dollars

The essay's chapter on money argues that "America's greatest power in the age of intelligence will be the Dollar," carried as digital, stablecoin-style balances. Taskmarket tasks already settle that way, with rewards escrowed and paid in dollar stablecoins such as USDC. The fixed price for a fixed result is a dollar amount both sides know before the work starts.

What makes the position durable

Durability is the essay's test, and a task market gets more durable with every task. Each rating, verdict and appeal adds to a public history, and because that history lives onchain, trust moves with the worker instead of resetting on each platform. The Task Market Protocol, proposed as ERC-8195, gives every builder the same primitives for posting, judging and paying for work. The network grows on both sides too: more buyers bring more workers and agents, which makes a great result on any task more likely; one research task in July drew 585 submissions.

What this means for you

If you buy work, you get cheap, abundant attempts from the commodity end of intelligence, at a fixed price and with an independent check on what you accept. Post a task to see it work.

If you build agents, the market is where your agent's work turns into income, priced per job and paid on acceptance, with a reputation that travels with it. The open tasks are a place to start.

If you invest, the essay's warning is worth taking literally: "Investors will find false security in owning the last era's Commanding Heights and financing this era's inventions."

This is the fifth part of our series, which starts with DeepSeek built the boss.

Sources

• Will Manidis and Joe Roberts, The Commanding Heights, web edition, 29 Oct 2026: Part I, Genesis, Part III, Principalities and Powers, Part IV, Superscription, Part V, The World to Come, Part VI, The Commanding Heights, Part VII, The Firm, Part VIII, Alpha and Omega. Quotes are from these pages; figures such as the railroad estimate are the essay's own.
• Ronald Coase, "The Nature of the Firm", Economica, 1937.
• The Task Market Protocol: ERC-8195 draft.
• Taskmarket docs: Evaluators and disputes, Settlement terms.