On April 2, 2026, the New York Times ran a piece called "How A.I. Helped One Man (and His Brother) Build a $1.8 Billion Company." Sam Altman emailed the reporter the same day. Years earlier he had bet his tech-CEO friends that the first one-person billion-dollar company was coming — and he said he'd just won. The company was Medvi, a GLP-1 telehealth reseller founded by Matthew Gallagher. Headcount of two.
Four days later Forrester's J.P. Gownder published the rebuttal: Beware the Magical Two-Person, $1 Billion AI-Driven Startup. The fake doctor profiles in the ads. The FDA Warning Letter for deceptive labeling. The deepfake before-and-after videos. The class action working through California. And then the detail that actually stuck with me: Medvi's chatbot hallucinated so consistently — fabricating drug prices, claiming the company sold products it doesn't — that Gallagher had to route the overflow of failures directly to his personal cellphone. Over a thousand customer calls, to one human, because an AI was making things up.
Altman's proof of concept for the one-person billion-dollar company is a two-person company whose AI hallucinations are being absorbed by the only available human's phone. The story is the thesis. Automation didn't eliminate the work. It concentrated it onto the one person still in the room.
The Math Everyone's Running
Altman's prediction wasn't pulled from thin air. Facebook bought Instagram in 2012 with thirteen employees. It bought WhatsApp in 2014 with fifty-five. The headcount-per-dollar curve has been dropping for a decade, and every step-change in automation flattens it further. AI is the biggest step-change anyone has ever shipped. The extrapolation feels obvious: if a billion-dollar company used to need thirteen people, and AI does more of the work than all prior tools combined, then the number should eventually hit one.
That math is right about production. It's wrong about what it takes to run a company.
Production Scales. Judgment Doesn't.
Every era of automation reveals a new bottleneck underneath the old one. When we automated physical labor, coordination became the bottleneck. When we automated coordination, decision-making became the bottleneck. AI has automated production. What's underneath it — review — is where the work actually piles up now.
Software engineering is already living this. Research shared in early 2026 found that in teams with high AI adoption, average code review time had risen by 199%, median time to first review by 157%, and median time in review by 442%. Surveys from the same quarter found that 38% of developers now say reviewing AI-generated code takes more effort than reviewing human-written code. Cursor — one of the most-used AI coding tools in the world — spent $290 million to acquire Graphite, a code review company. That's not a feature add. That's the market repricing where the bottleneck lives.
And the people doing the actual work are saying it plainly. Simon Willison writes that running four agents at once leaves him "wiped out by 11 a.m." Addy Osmani, on Google's Chrome team, calls code review the new bottleneck. Armin Ronacher, who estimates 90% of his code is now AI-written, puts it this way: the remaining human work is "not reduced volume but concentrated responsibility."
That last phrase is the thesis in seven words.
The Forty-Hour Week Was Always Mostly Padding
Most jobs were never pure judgment. A developer's forty hours used to be maybe ten or fifteen hours of real decision-making, and the rest was keystrokes, lookups, tests passing, context-switching, fiddling. An accountant's forty was maybe eight hours of judgment padded with data entry, reconciliations, reformatting the same report five different ways for five different executives. A lawyer's forty was maybe six or seven hours of argument construction wrapped in document review, citation checking, and formatting.
Most of what filled the week wasn't judgment. It was execution, coordination, and friction. AI is eating the execution, the coordination, and the friction. What's left is the judgment — concentrated, end-to-end, with no padding in between.
The Hours Nobody Measured
So what's the actual ceiling on sustained judgment? Honestly, nobody has measured it directly. But some research puts sustained cognitive work at just two to four focus hours per day. Workplace telemetry keeps landing in the same neighborhood. That works out to somewhere between ten and twenty hours of real judgment in a workweek. Call it twenty on the optimistic end. Not because a study said so — because every adjacent number points at something in that range, and nothing credible points higher.
Medicine has been here before, at the extreme end. When residents were breaking down, regulators capped the workweek — not because they had the outcome data to prove overworked humans made worse decisions, but because it was obvious at the gut level that they did. The gut-level case for AI-era knowledge work is the same. You don't need a study to tell you that a controller making a hundred judgment calls back-to-back is worse at the hundredth than the tenth.
Not One. Not Zero. Somewhere in Between.
The AI-native billion-dollar company of 2030 isn't one person. It also isn't zero. The current record-holders aren't close to one — Cursor is at roughly $1 billion ARR with north of 300 staff, Midjourney is somewhere around $300-500 million with 100-plus. The curve is pointing at dozens, not ones.
Who are those dozens? Mostly reviewers. People making the judgment calls that AI is escalating. People whose signature is the one the system is accountable to. People whose twenty-hour weeks of concentrated review produce the output that a 200-person team used to produce in a forty-hour week of mostly padding.
Some industries are already structured for this. Professions with legally required human sign-off — medicine, accounting, law, engineering — have judgment gates built into the architecture. The CPA signs the opinion. The MD writes the prescription. The PE stamps the drawing. AI doesn't eliminate the signature; it multiplies what gets signed per signature. Those professions fit the many-humans-plus-AI model naturally because the structure is already there. Most other industries will have to invent what medicine and accounting have had for a century: formal review gates with accountable humans on the other side.
This is the counterintuitive upside. Shifting people from production to review is the trade that actually produces multiplicative industry productivity. Twenty hours of concentrated review leveraging AI-produced output beats forty hours of padded execution doing the work by hand, by a lot. The companies that figure this out first staff the review function deeply, cap their people at honest judgment-hours, and out-produce every competitor still running a forty-hour week with half the headcount.
The one-person billion-dollar company isn't coming. Neither is the zero-headcount one. What's coming is the twenty-hour-a-week, thirty-professional, billion-dollar firm — and the ones who design for that first will win the decade.
The phone still rings. Someone still has to pick it up. That person's job was never the one AI was going to eliminate. It was the one AI was going to make denser, scarcer, and more expensive per hour.