NORBERT BIEDRZYCKI
4 min readBusinessRead this in Polish

The token bill is the new payroll

For decades, people were the largest cost line in a services or technology company. Budgets were planned in headcount, margins were calculated from rates, and company value was estimated, among other things, from revenue per employee. That model is starting to crumble, and faster than finance departments can keep up.

Three signals from one month

At the end of October three stories appeared that, taken together, show the direction.

Jacob Sansbury, founder of the ten-person startup Spawn, wrote that his company spent $3.1 million on tokens in a single month, roughly thirty times its entire payroll. That is a founder’s claim, nobody has verified it independently, and Sansbury says nothing about revenue. But the ratio alone speaks volumes.

A few days later Fortune described how Cisco CEO Jeetu Patel gave a twenty-seven-year-old engineer unlimited access to tokens and a goal of rewriting 98% of the company’s legacy code in 24 months, with four weeks to prove it can be done at all. The team is to consist of two people. Nobody disclosed the cost.

Finally, Anthropic published a report on how AI has changed work inside the company itself. Agent working hours there exceed human working hours six to one, and revenue is growing much faster than headcount, faster even than in the previous generation of technology companies at the same stage.

The common denominator is simple: the cost of labour is moving from salaries to compute.

What it changes in the P&L

From the perspective of a board and a private equity fund, I see three changes worth naming plainly.

The cost structure changes. Salaries are a largely fixed and predictable cost. The token bill is variable, grows in jumps and is often hard to allocate. In many companies nobody can say today whether it is a cost of services sold or an overhead. And the gross margin a fund shows the next buyer depends on that answer.

Revenue per employee breaks down as a metric. If ten people produce what once required a hundred, revenue per employee looks superb. Except it no longer says much about efficiency. We need a new measure, something like the cost of executing a process, where people and tokens sit on the same line.

Supplier concentration risk appears. In June 2026, access to the Fable models was suspended for almost three weeks to comply with US export rules. Sansbury claims work at Spawn dropped to zero at the time. However accurate that account is, it shows something missing from most due diligence reports: an AI-native company is only as independent as its technology stack.

What I see in portfolio companies

Working with clients and fund portfolio companies, I see something that rarely makes it into presentations. When token bills started to grow, companies had to rank use cases by return on investment. In some of them it turned out that for a specific task it was cheaper and more reliable to hire an experienced developer than to keep an agent burning tokens without a clear result.

This is not an argument against AI. It is an argument for doing the maths. The period in which token consumption was treated as proof of modernity is ending. The market is starting to ask not how many tokens a company consumes, but how much value it extracts from them.

A lesson for Poland

For the past twenty years our development model has rested on relatively cheap, well-educated labour. Shared service centres, software houses, IT outsourcing. Western companies paid for hours of work in Kraków, Wrocław and Łódź.

If the model becomes a company in which a few people spend a multiple of their own salaries on tokens, that difference does not go to a team in Poland. It goes to the account of one of a handful of AI labs in the United States. That is a question not only for individual companies, but for an entire economy that has spent years selling hours.

Questions for due diligence

When I assess a company today in which AI is meant to play a significant role in the operating model, I ask three questions.

How much does it cost to run the key process, counting people and tokens together? Not how much the licences cost, but how much the process costs end to end.

What happens if the main model provider raises prices by half tomorrow, or suspends access? How long would it take to switch to another model, and has anyone ever tested it?

Who on the board is responsible for the token budget? If the answer is “the IT department”, the company still treats AI as a tool rather than as a workforce.

Why at Sophona we think about the process, not the chat

That is why we design AI Operators at Sophona around specific processes rather than around access to a chat. An Operator has an assigned scope of work, so its cost can be calculated just like the cost of the team that used to do that work. For a CFO, that is the difference between “we spend a lot on AI” and “we know what closing the month costs us”.

For years salaries were the most predictable line in the budget. The token bill has yet to become one. Companies that learn to plan, control and diversify it before their competitors will build an advantage that no subscription can buy.

Read next

  1. AI won’t run out of customers. It will run out of power and time to repay its debt

    Many people read the sell-off in data center stocks as the first crack in the AI bubble. I look at the same numbers and see something else. The demand is there, and it is substantial. The trouble lies elsewhere: in the construction itself, in access

  2. The risk in AI isn't demand. It's the capital structure.LinkedIn

    That isn't the profile of a tech company. It's the profile of a leveraged buyout, except the asset isn't a factory but graphics processors. That is what the balance sheet of CoreWeave, one of the largest neoclouds, looks like today.

  3. The AI bubble is the wrong questionLinkedIn

    Microsoft has just pulled out of a deal to build a data centre in Loughton, near London. Not for lack of demand, but for lack of power: the grid connection there won't be ready until the 2030s. And it's not the only case.