Many workforce decisions being made now come down to what it costs to run AI versus what it costs for a human to do the same job. Companies are restructuring around that math.
Ramp, which tracks corporate spending across more than 70,000 businesses, found that the top 1% of companies — the ones it calls “AI-pilled” — are now spending roughly $7,500 per employee per month on AI tools and compute. For comparison, the average U.S. software engineer costs about $16,000 a month, so the heaviest AI spenders are still paying less than half of one engineer’s salary to put AI in an employee’s hands.
That spend among the AI-pilled companies has grown more than twelvefold since mid-2023, from under $600 per employee to nearly $7,500, with each year rising faster than the last. That kind of growth won’t stay below an engineer’s salary for long.
It isn’t just the early adopters, either. The median company spends only about $11 per employee, but it’s on the same curve, just earlier.
The rapid growth in AI spending is already straining some budgets. Mercor is now spending more on AI than on employee salaries. So are divisions inside Nvidia. Uber capped employee AI spending in April after burning through its entire 2026 coding budget in four months.
OpenAI’s audited pre-IPO financials leaked this month, and the numbers are an eye-opener. The company pulled in about $13 billion in revenue in 2025 against $34 billion in costs — an operating loss of roughly $21 billion in a single year. The product is being sold for far less than it costs to build and deliver.
It’s the Silicon Valley playbook: launch below cost, deliver so much value at that price that adoption goes vertical, then raise the price once everyone’s hooked. Remember when an Uber across Manhattan was cheaper than a yellow cab? When Netflix was $7.99 a month? When Google Workspace was simply free? None of those prices were meant to last. The services were priced below cost to win market share, and once they hit mass adoption, the price climbed steadily as the focus shifted to profit.
We are squarely in that phase with AI. AI models read and write text in units called tokens — very roughly, a token is a word or a piece of one — and every provider charges by the token. The OpenAI financials show that the tokens we’re using today are priced below what they actually cost to develop and deliver. That can’t last forever, and when it ends, the bill will rise.
The falling cost of tokens won’t save anyone here, either. The price of a given level of AI capability has fallen dramatically in the last two years. Yet the work has grown faster than the price has fallen: the newest models use far more tokens for each inference, and agentic systems that chain together many steps burn even more. We have access to cheaper tokens, but need more of them, so the total bill keeps climbing even as the per-token price collapses.
Organizations are reshaping their workforces today around AI costs that are inherently unstable — cutting roles and holding headcount flat on the strength of AI pricing the providers are losing billions to sustain.
How many of the workforce bets being placed today are built on pricing that can’t last? What happens when the prices go up?
