Your next biggest expense may not be payroll

On October 26, Anthropic, the company behind Claude, filed with the SEC to go public. Almost everyone talked about revenue and valuation. I got stuck on something else: a single line, buried in the section where they describe how they use their own product.

It needs a careful read. This is not what they spend: it is what any customer would pay for the same usage at list price. And the payroll it is compared against leaves out stock compensation, which at a company like this is a lot. Even with both caveats, the number is huge.

"Of course, they ARE the AI"

Fair enough. A company that builds AI is going to use a lot of AI. What strikes me is not how much they use it, but the question it leaves open for the rest of us:

In 2 or 3 years, how much will AI weigh on the income statement of a mid-sized company?

And I don't mean just the model. I mean the whole line item:

Today that line item is small for two reasons. First, most companies are only starting to implement AI. Second, and this is the one few people see: the subscription AI almost all of us use is HEAVILY subsidized.

The gym model

Ed Zitron explained it on The Diary of a CEO podcast, citing the research firm SemiAnalysis: on ChatGPT's $200-a-month subscription you can burn up to $14,000 worth of tokens. On Anthropic's, up to $8,000. And even on the $20 plan, up to $400.

"On a $200 a month ChatGPT subscription, you can burn $14,000 worth of tokens. And on Anthropic's, you can burn $8,000 for 200 bucks." Ed Zitron, The Diary of a CEO, minute 12:57

Those are ceilings, not averages: it is what the heaviest users consume. But the mechanism is a gym's. Many people pay and rarely show up, and that covers the few who come every day. 🏋️ For the heaviest users, the subsidy can exceed 90%.

The problem for that model is that agents, running on their own, are going to show up EVERY DAY to consume tokens. And not for an hour: all day.

The same filing shows where this is heading

On Claude Enterprise, a user who only uses the chat spends less than $75 a month on average. The top 5% of people who use agents the most spend about 20 times that: roughly $1,500 a month per person.

And 80% of Anthropic's revenue in the first half of 2026, $13.1 billion, came from charging for tokens consumed, NOT from subscriptions. That is not new: it was already 75% in 2024 and 83% in 2025. The big AI business is already billed by usage.

The "roughly $1,500" is my own math (75 × 20). The filing gives the ratio, not the exact figure.

Where I think this is going: we will pay for AI the way we pay for utilities

My personal view: the gym model will NOT hold. We will end up paying for AI in blocks of millions of tokens, the same way businesses already pay for electricity or cloud computing: by the unit, metered. And when that happens, the shape of company spending is going to change, and A LOT.

This is my opinion, not an Anthropic figure. None of the labs has said its subscriptions lose money. But the usage figures in their own filing point that way.

Should you wait to adopt AI?

NO, I see it the other way around. Today is the cheapest it will ever be to learn and experiment across every area of the business, because the subscription costs far less than it costs the AI companies to provide.

But not to play around. To solve real problems that are worth money:

• Customers leaving for the competition before anyone notices.
• Products out of stock that purchasing catches too late.
• Sales reps who are not delivering, and in which product lines it shows.

Once AI is billed by usage, the company that already knows what to ask it, and with which data, will spend better than the one just getting started.