Anthropic is spending $500,000 per employee on AI. Not on salaries — on top of salaries. I came across that number recently and couldn’t stop thinking about what it means for the rest of us building budgets and headcount plans.
I’ve talked before about how hard AI costs are to pin down — token pricing keeps shifting, companies are moving from seat-based plans to usage-based billing, and the actual value you get back is hard to trace to a dollar figure. That’s already a messy environment. Then I read a piece that put real numbers behind just how wide the AI spending gap has gotten across companies, and it reframed how I think about budgeting for this entirely.
The Spread
Here’s the spread. The median company spends about $137 a year per employee on AI — roughly $10 a month. That’s most companies. The top 1% of software companies, though, are spending around $89,000 a year per employee, about 40% of a $220,000 senior engineer’s salary, just in token costs. And then there’s Anthropic — the AI-native frontier — spending over half a million dollars a year per employee, 2.3 times its average payroll.
Sit with that range for a second. $10 a month at the median. $89,000 a year for the top 1%. $500,000-plus a year at the frontier. That’s not a normal spread for any line item I’ve budgeted before.
I think about it like compensation bands, except instead of stretching across a 20-year career, it’s compressing into a 3-year window. The AI-native companies are essentially giving each employee the equivalent of a partner’s salary in tokens. The top 1% are giving them an associate’s salary. And the median company’s $10 a month is like handing someone half a Netflix subscription.
What’s Driving It
What’s driving this is a shift from seat-based pricing to usage-based, token-driven pricing. We saw it first with Anthropic and OpenAI enterprise deals, then GitHub did the same thing. Every time a vendor makes this switch there’s backlash, and then nothing changes — usage-based billing is where this is all heading, and it means spend scales with how much your team actually uses AI, not with how many seats you’ve licensed.
So where does this go? I see three scenarios. In the bear case, technology gets more token-efficient and usage growth flattens out — token deflation wins. In the base case — and this is closer to my own view — usage keeps climbing steadily; I don’t think we’ll ever use fewer tokens tomorrow than we did today. And in the bull case, frontier models get good enough to act as genuine replacements rather than add-ons — instead of hiring someone for $200,000, you replace the role with $150,000 of AI.
I’m not betting on the bull case playing out at scale anytime soon. But I’d be cautious betting against any token growth at all, too.
What I’d Actually Budget
Here’s what I’d actually do with this if I were building a budget right now. I wouldn’t model myself going as extreme as Anthropic — spending more on AI than on payroll is still a low-probability outcome for most companies. But I would start modeling something like 10% of employee salary as an AI line item, layered on top of comp. If you’re paying someone $100,000 a year, budget for roughly $10,000 in AI tooling and token spend on top of that, and build your headcount plans around that number, not around zero.
Because here’s my real prediction: a few years from now, looking back and saying you only spent $10 a month per person on AI is going to look ridiculous. I could easily see that closer to $1,000 a month per person, if not more.
Key Takeaways
- The AI spend gap between companies is enormous: $10/month at the median, $89,000/year at the top 1% of software companies, $500,000+/year at AI-native frontier firms.
- The shift from seat-based to usage-based, token-driven pricing is permanent — plan your forecasting models around usage, not headcount.
- Consider modeling AI spend as roughly 10% of an employee’s salary as a starting budget assumption, layered on top of comp.
- Whichever scenario plays out — bear, base, or bull — today’s $10/month median is likely to look low within a few years.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: AI Spend Per Employee
Leave a Reply