Compute Will Eventually Cost More Than People: The Company Ledger Most Bosses Have Never Run

The ledger nobody runs

At a dinner, a tech lead at one of the big companies told me a Claude Code subscription was too expensive and the company had no budget for it. He runs a 200-person dev team, and not one of them has a seat.

The math is easy. 200 engineers, 400,000 to 500,000 yuan a year in fully loaded cost each, so just under 100 million yuan a year. Kit all of them out with top-tier Claude and you’re at about 1.7 million a year—not even 2% of payroll. A company spends 100 million a year on people and won’t put down 2% of that to give them the best tools.

What everyone else spends

Ramp handles spend for tens of thousands of companies, so it sees real numbers. An average company spends a dozen or so dollars per person per month on AI. The most generous 1% spend 7500 dollars per person per month—more than 50,000 yuan, a gap of over 600x.

The most extreme case is Anthropic, now the highest-revenue AI company in the world. It spends more than twice as much on compute as it does on salaries. An engineer on a 200-odd-thousand-dollar salary gets paired with 500-odd thousand dollars of compute. There, compute has been more expensive than people for a long time.

Headcount is addition, compute is multiplication

Why are these companies willing to spend like that? Because these are two different things. Headcount is addition—you’re buying hours, and one person only has eight of them a day. Compute is multiplication—it’s leverage on your most expensive people. It doesn’t sleep, it copies, and it gets easier the more you use it.

How big is that leverage? I just lived through it. A project that used to take three to five people three or four months—I delivered it alone, with AI, in two days. Across those two days it never slept, and I couldn’t bring myself to either—several threads running at once, and every extra hour of sleep felt like money left on the table. This isn’t a bit faster or a bit slower. It’s a different order of magnitude of output.

Getting results, in three layers

Just buying seats does nothing. Same tool: whoever knows how to use it makes dozens of times their money back, and whoever doesn’t can’t even earn the seat back. How compute gets allocated, who gets it, and how it gets turned into results—someone has to own that. Headcount has HR. Compute, a means of production this big, has nobody in charge at the overwhelming majority of companies. I call that role the Token Resources Consultant.

Getting results comes in three layers. Layer one, hand the repetitive work to AI, efficiency up by 10-plus percent. Layer two, scale the things that actually generate business, take on the orders you used to turn down, revenue up 30 to 50 percent. Layer three is the hardest and the most valuable—finding the new business that only exists because you have AI. There’s nothing to copy at this layer, and a consultant can’t find it for you. Only the CEO getting hands-on themselves will see it.

So when I set up Claude Code for partners and clients, I always have the CEO start using it personally. Explaining it a hundred times is worth less than them running it once themselves.

Tools can’t be hoarded, advantage has to be grown

One entertaining pattern. My videos get forwarded constantly and liked almost never—people quietly send them to themselves, figuring it’d be best if they were the only one who knew. I understand the impulse, but you can’t hoard a tool. A few-hundred-yuan subscription is available to anyone; it’s a matter of six months earlier or six months later. The advantage you can actually hold is what you build with AI inside your own business that nobody can copy. Rather than guarding against other people learning it, go deeper with it yourself.

I believe a company’s compute cost will sooner or later exceed its payroll. That day is not far off.

This episode is an AI digital-avatar demo video (both the likeness and the voice were generated from models trained on my own material). The transcript is the original production script; the timecodes match the finished cut.

[00:00] At a dinner a few days ago, a head of engineering at a big tech company told me Claude Code subscriptions are too expensive, the company has no budget for it, and his 200-person dev team doesn’t have a single seat. I thought that was really strange, because I want to install it for everyone around me—partners, clients, one at a time, every one I meet.

[00:15] So let’s run the numbers. Two hundred engineers, at 400,000 to 500,000 yuan a year all-in per person, is just under 100 million yuan a year. Top-tier Claude for every one of them costs about 1.7 million a year. In other words: you spend 100 million a year on people, and you won’t part with 2% of that to give them the best tools.

[00:29] So how do other companies spend? There’s a US company called Ramp that keeps the books for other companies, so it can see data from tens of thousands of them. The average company spends about ten-odd US dollars per person per month on AI. But the top 1%—the ones most willing to spend—put in $7,500 per person per month, over 50,000 yuan. Same job, kitting people out with AI, and the most generous and the stingiest are more than 600x apart.

[00:46] The most extreme case is Anthropic, now the highest-revenue AI company in the world. What it spends on compute is more than twice its payroll. An engineer on around $200,000 a year gets over $500,000 of compute. Over there, compute has cost more than people for a long time already.

[00:59] So why do these companies dare to spend like that? Because people and compute are fundamentally two different things. People are addition—what you’re buying is hours, and one person has eight hours in a day. Compute is multiplication. It’s leverage on your most expensive people: it doesn’t sleep, it can be copied, and it gets smoother the more you use it.

[01:12] How big can that leverage get? A while back I did a project myself that, in the old days, would have taken three to five people three or four months. Me plus AI delivered it in two days. And what state was I in over those two days? It doesn’t sleep, and I couldn’t bring myself to sleep either. I was getting so much done, with several threads running at once, that one extra hour of sleep felt like money left on the table. This isn’t a bit faster or a bit slower anymore—the sheer output isn’t in the same league.

[01:32] Of course, just buying the subscriptions does nothing. Same tool: someone who knows how to use it gets dozens of times the return, and someone who doesn’t won’t even earn the cost back. So compute needs someone in charge of it—how you hand it out, who gets it, how you get results out of it. Think about it: people have HR. Compute is a huge chunk of what actually produces the work, and at the vast majority of companies nobody owns it right now. I call that role the Token Resources Consultant.

[01:49] So what counts as getting results out of it? The way I see it, there are three layers. The first is the easy one: hand the repetitive work to AI and free up your people. Everybody can do this layer, and it buys you 10-something percent. The second layer: take the things that actually bring in business and scale them up. It used to be more leads than hands, so deals just got dropped on the floor. Now AI writes the proposals and does the follow-ups for you, so you can actually catch them—and revenue can go up 30% to 50%.

[02:09] The third layer is the hardest and the most valuable: going out and finding the businesses that only exist because AI exists. There’s nobody to copy here. But if you find it, that’s where the real money is. And this kind of business is not something a consultant can find for you—only the CEO getting in the game themselves can see it.

[02:21] So when friends and clients around me are willing to use it, and I help them set up Claude Code, I always get the CEO using it first. Explaining it a hundred times is worth less than them running it through once themselves.

[02:29] Which brings me to a funny thing. My videos get a lot of shares and very few likes. Everyone quietly forwards them to themselves, thinking: ideally I’m the only one who knows this, my colleagues and clients had better not pick it up. Wanting to protect your competitive edge—I completely understand that instinct.

[02:43] But think about it: you can’t hide a tool. A subscription costs a few hundred yuan, anyone can buy it, it’s a matter of six months either way. The advantage you can actually hold is using AI inside your own business to build something nobody can copy—that third layer I just talked about. Rather than guarding against other people learning it, go deeper with it yourself. And I’ve never been afraid to teach—the Feynman technique, right? Every time I talk it through with someone, it can spark a deeper understanding.

[03:04] I believe a company’s compute cost will eventually exceed its headcount cost. And that day isn’t far off. That’s it for today. See you next time, bye.