Revenue Per Employee: The AI-Era Metric

In a recent Y Combinator interview, Garry Tan did the math on his own work and got a number that sounds made up. The first time he built Posterous, his 2008 startup, it took six or seven people, about a year and a half, and roughly four million dollars. In January he rebuilt the same product a third time. It took about two hundred dollars of Claude Code usage, five days, and mostly one person directing AI agents. That gap is the whole story of revenue per employee in the AI era. The metric is old. What changed is how far one person can now push it.

Revenue per employee used to be a quiet efficiency line on a board deck. Now it is becoming the number founders get judged on, because it is the cleanest way to see whether AI is producing real output or just adding cost. Tan, the president and CEO of YC, calls his approach tokenmaxxing: spending hard on the best models to do, in his framing, the work of a 400-person engineering team. Here is what that means for the number you should be watching, and how to raise it.

Why revenue per employee is the metric now

Revenue per employee is simple: your annual revenue divided by your headcount. For years a healthy software company sat somewhere around 200,000 to 400,000 dollars per person. That was the band investors used to sanity-check whether you were building a real business or just hiring to look busy.

AI broke the top off that band. A new wave of AI-native companies is posting several million dollars of revenue per head, sometimes on teams of a few dozen people. The pattern is not new, it is just getting common. When Facebook bought WhatsApp in 2014, the team was about 55 people, which on a 19 billion dollar deal is a number no traditional org chart could explain. What used to be a famous outlier is turning into a template.

This is why Sam Altman has said tech founders now keep a betting pool on when the first one-person, billion-dollar company shows up. You do not have to believe the extreme version to take the trend seriously. The direction is clear: the companies winning right now produce more per person than anyone thought possible, and revenue per employee is the metric that makes that visible.

What tokenmaxxing actually looks like

Tan went 13 years without writing code, came back this year, and now estimates he is shipping about 400 times the volume of code he produced the last year he coded regularly. He does it by directing roughly 15 agents at once, dropping more than a dozen pull requests in a couple of days, while running YC as a full-time job.

Tokenmaxxing is his word for the habit underneath that: spend on the best models and burn the tokens, because the output is worth far more than the bill. He compares it to San Francisco rent. It looks expensive until you realize it is more expensive not to pay it. A founder economizing on model spend to save a few hundred dollars a day is, in his view, optimizing the wrong line. The token bill is not overhead, it is the thing producing the work.

He is also honest about the catch. Using these tools at full power, he says, is like driving a Ferrari: exhilarating, and prone to breaking down on the side of the road right when you need it most, so you had better be ready to pop the hood and fix it yourself. The output per person is real, but it does not come from typing one prompt and walking away.

The same signal across our knowledge base

Tan is not an isolated case. The point keeps surfacing across the YC and top-investor conversations we track. Founders building AI-native companies describe getting higher revenue per employee by adding AI instead of headcount, and name token spend, not team size, as the variable that now matters. The framing in one YC talk is blunt: the critical shift is maximizing tokens, not hiring. Another argues that small teams wired with AI into every process can compete with companies many times their size, and that this produces a wave of high-valuation companies with very few employees.

Put those together and the benchmark moves. The question stops being "how many engineers do you have" and becomes "how much can each person, plus their agents, actually ship." That is a revenue-per-employee question, and it is the one your next investor will ask in some form.

How to raise your revenue per employee

The useful part of Tan's account is that the moves are concrete. None of them need a research lab, just a change in how you work.

• Treat token spend like rent, not like a snack budget. Buy the top tier, run the strong models, and stop rationing the one input that produces the work. Economize on the office couch, not the tokens.
• Direct, do not type. Tan's gains come from steering agents, not writing the lines himself. Your job moves to setting the goal, the constraints, and the standard, the same shift behind running a fleet of agents instead of one.
• Boil the ocean. Because the machine does not tire, do the complete version: pull 20 sources instead of one, cross-check them, write the tests you would normally skip. The thoroughness that used to be too expensive is now cheap.
• Test to 80 to 90 percent, then ship. Tan learned the hard way that untested agent output is slop that falls over the moment a real user touches it. As a CISSP, I would underline this: coverage is the control that makes the speed safe, not an afterthought.
• Run the 10x check. For any task, ask what would deliver ten times the value for twice the effort, and aim there instead of at the obvious version.
• Reuse the harness, write the skills. Do not rebuild the agent loop. Put your standards and conventions into plain markdown the agents read on every run, so you are not re-explaining your taste each time.

Where humans still set the ceiling

High revenue per employee is not a story about firing everyone. It is a story about what one motivated person, pointed at a problem they care about, can now produce. Tan is direct that the human still has to supply the agency: the model can do the work, but it cannot decide what is worth doing or what "good" means for your customer. His own setup leans heavily on the agent stopping to ask him questions, because that is where his judgment enters.

That is the real lesson for founders. The metric rewards output per person, and the input that scales it is not headcount, it is a person with taste and direction multiplied by machines. It is the same engine behind the solo founder who now runs like a whole team and the AI-native company built thin on purpose. Revenue per employee is just the scoreboard that shows whether you are playing that game.

What to do this week

• Calculate your current revenue per employee, then set a target. Knowing the number is the first step to managing it.
• Upgrade everyone who builds to the top model tier, and stop treating token spend as a cost to minimize.
• Pick one job you do by hand each week and hand it to an agent you direct, instead of doing it yourself.
• Add the tests you have been skipping. Get the work an agent ships to 80 to 90 percent coverage before it reaches a user.
• Write one skills file: your standards and conventions in markdown, so every agent run starts from your taste.
• Before building anything, run the 10x check and aim for the more ambitious version the machine now makes affordable.

Raising revenue per employee is not about squeezing people. It is about building the company so one person, plus their agents, can do what used to take a department. That operating model is exactly what the AI Operating System for Startups is built to teach.

Sources

• Tokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers (Y Combinator's The Lightcone), the interview this article distills.
• From our video knowledge base: How to Build a Self-Improving Company with AI (higher revenue per employee from AI instead of headcount), How To Build A Company With AI From The Ground Up (maximizing tokens, not hiring), and Everyone's Wrong About AI and Jobs (20VC, high-valuation companies with fewer employees).
• Background on Garry Tan: Wikipedia and TechCrunch on his role and Posterous.
• WhatsApp's headcount at its 2014 acquisition and the one-person billion-dollar-company idea (raised by Sam Altman) are widely reported.