What Do Companies with $10 Million in Revenue per Head Get Right? The Same Five Things, and I Do All of Them

The companies that create the most wealth per head all get the same five things right. Last episode I said we already have the world’s top AI in our hands; to make more money, the first step is figuring out which ways of making money work best. So yesterday I spun up 23 AI agents to look into it, going through cases from more than 100 companies and individuals — AI companies, but also businesses selling chocolate and gold. The ones that made money fastest rode a wave or had resources nobody else had, which ordinary people can hardly copy. But among service companies, there is a group that never took investment and has the highest per-head figures. Their playbooks are strikingly similar and come down to five things — which are also why my own business has been growing lately.

One: sell services, not tools. Tools earn software money; services earn labor money. Sequoia Capital said in March this year that for every dollar companies spend on software, they spend six on services. Selling tools means racing the large models: each time model capability improves, a tool’s existing features may lose their value. Service companies turn their service process into tools and templates, so every model improvement makes the work faster and cheaper.

Surge AI makes training data for large-model companies. It delivers finished data and built its own quality management system to manage outside annotators. With fewer than 100 people, its revenue per head exceeds $10 million a year — revenue, not profit. It has been profitable since day one and has never taken a cent of investment. Across more than 1,000 privately held software companies, by contrast, median annual revenue per head is only around $140,000. It amounts to selling AI as if it were a person and charging a service fee; a higher value anchor brings higher profit.

Two: value-based pricing — pricing on the gains a client gets, or even pricing and delivering on results. Under hourly billing, the faster the work, the less the pay; under gain-based pricing, the less time spent, the more profit remains. One step further is pay-only-if-it-works: the client doesn’t have to put money down and gamble on the outcome, and the close rate goes way up. McKinsey now bills roughly a quarter of its fees on results. Surge AI’s newly launched ready-made datasets work the same way: large-model companies it trusts can train on them first and pay once they see results. Hourly-billing firms, meanwhile, use AI to become more efficient and end up logging fewer hours. The precondition is that what the result is, and what counts as working, gets spelled out up front.

Three: serve only a few clients with high ability to pay. The more a client’s time is worth and the bigger their business, the bigger the gains the same service can bring, so the price can be higher; fewer clients also means deeper service for each one. Two former SoftBank fund managers do AI training for big Wall Street institutions, with clients including Citi and Bank of America. They charge $25,000 a day, are booked out two months, and almost every client comes back. Small clients are the opposite: under value-based pricing, a small business brings small gains, and the profit can’t go up.

Four: owned traffic — winning business through one’s own content and referrals from existing clients rather than paying for traffic. Paid traffic costs money all over again for every new client; content, once published, keeps getting seen, and existing clients keep referring new ones. Gamma, which uses AI to make PPTs, raised only $23 million before reaching $100 million in revenue, and more than half of its new paying users came through word of mouth. With about 50 people, its revenue per head is around $2 million a year, and it has been profitable for two years straight. By comparison, an App that counts calories from a photo makes $1.4 million in gross profit a month but spends $770,000 on advertising and marketing alone, eating up more than half of its gross profit; with 30 people, its profit per head is only around $110,000 a year.

Five: few people, no idle staff. All work is outsourced to partners rather than kept in a fixed in-house team, so there is no fixed cost of a whole team to carry when there are no projects. For the same profit, fewer people means higher profit per head. Submagic, which does AI subtitle editing, has 10 people and no sales team, with revenue per head of over $800,000 a year. PhotoAI, which does AI portraits, is just its founder — one person — and going by this March’s numbers, makes over $900,000 a year in profit.

Looking back at my own work, I’m doing all five of these things, and I’ve more or less reached a state of project freedom: I take on the clients I want and turn down the ones I don’t, charge less for projects I like, and my time is entirely my own.

EP0114 · 2026-10-06 · 04:17. The picture uses a digital human, and the voice uses my cloned voice. This is an English translation of the approved Chinese script used to create the narration.

The companies that create the most wealth per head all get the same five things right.

Last episode I said that we already have the world’s top AI in our hands. So if we want to make more money, we first have to figure out which ways of making money work best. That’s why yesterday I spun up 23 AI agents to look into it. They went through cases from more than 100 companies and individuals. Some were AI companies, and some sold chocolate, or gold. What I found is that the ones that made money fastest either rode a wave at the right moment, or had resources nobody else had. That’s very hard for ordinary people to copy. But among service companies, there’s a group that never took investment and has the highest per-head figures. They all play it in a very similar way, and it comes down to five things. These five things are actually also why my own business has been growing lately.

Number one: sell services, not tools. A lot of business owners think AI transformation means going out and buying tools, or building a tool to sell. But tools earn software money; services earn labor money. Sequoia Capital said it back in March this year: for every dollar companies spend on software, they spend six dollars on services. Now, if you sell tools, you’re racing the large models. Every time model capability improves, the features your tool used to offer may stop being worth anything. Selling services is different. What the client buys is the finished thing. Service companies turn their service process into tools and templates. Every time the models get better, the work gets done faster, and it costs less, too. There’s a company in the US called Surge AI that makes training data for large-model companies. What it delivers is finished data. It built its own quality management system to manage outside annotators. The company has fewer than 100 people, and revenue per head is over $10 million a year. That’s revenue, not profit. It’s been profitable since day one, and it has never taken a single cent of investment. And ordinary software companies? Someone looked at more than 1,000 privately held software companies. Median annual revenue per head was only around $140,000. It’s basically selling AI as if it were a person, and charging a service fee. When the value anchor is higher, the profit is higher too.

Number two: value-based pricing. That means pricing on the gains the client gets, or even pricing on results and delivering on results. Why does this matter? If you charge by the hour, the faster you work, the less you actually get paid. Pricing on gains flips that around. For the same job, the less time you spend, the more profit is left over. Take it one step further, and it’s pay only if it works. The client doesn’t have to put money down and gamble on the outcome. Hand the certainty to the client, and the close rate goes way up. You might not know this, but even McKinsey now bills roughly a quarter of its fees based on results. The company I just mentioned does this too. It recently launched a set of ready-made datasets. Large-model companies it trusts can take them and train first, and pay once they see results. And the companies that bill by the hour? They use AI to become more efficient too, and end up having to log fewer hours. Of course, what the result is, and how far it has to go to count as working, has to be spelled out up front, right?

Number three: serve only a few clients with high ability to pay. The more a client’s time is worth, the bigger their business, and the bigger the gains the same service can bring them. So the price can be higher. With fewer clients, you can also serve each one more deeply. There are two former SoftBank fund managers who do AI training specifically for big Wall Street institutions. Their clients include Citi and Bank of America. They charge $25,000 a day, and they’re booked out two months. Almost every client comes back to them again. And serving small clients? By the logic of value-based pricing, a small client’s business brings small gains, so the profit just can’t go up.

Number four: owned traffic. That means getting business through your own content and through referrals from existing clients, not paying for traffic. With paid traffic, every new client means paying all over again. But content, once it’s out there, keeps getting seen. And existing clients keep referring new ones, so the acquisition cost is saved. There’s a company that uses AI to make PPTs, called Gamma. It took very little investment. Before reaching $100 million in revenue, it raised only $23 million. More than half of its new paying users came through word of mouth. The company has about 50 people, with revenue per head of around $2 million a year. It’s been profitable for two years straight. For comparison, there’s an App that counts calories from a photo. It makes $1.4 million in gross profit a month, and spends $770,000 on advertising and marketing alone. That one item eats up more than half of its gross profit. The company has 30 people, which works out to profit per head of only around $110,000 a year.

Number five: few people, no idle staff. All the work is outsourced to partners. No fixed team kept in-house. When there’s a project, you partner up. When there isn’t, you don’t have to carry the fixed cost of a whole team. With the same profit, the fewer people in the company, the higher the profit per head, naturally. There’s a company doing AI subtitle editing called Submagic. Ten people, no sales team. Revenue per head is over $800,000 a year. And there’s PhotoAI, which does AI portraits — it’s just the founder, one person. Going by this March’s numbers, it makes over $900,000 a year in profit.

So I looked back at my own work, and found I’m actually doing all five of these things. These days, I’ve more or less reached a state of project freedom. Clients I want to take on, I take on. Clients I don’t, I don’t. Projects I like, I charge a bit less for. My time is entirely my own.

That’s it for today. See you tomorrow, bye.