AI is more expensive than humans
“For my team, the cost of compute is far beyond the costs of the employees” - Quote by Bryan Catanzaro, vice president of applied deep learning at Nvidia [1]
I have to say, that’s a ridiculous thing to say. Or is it?
The narrative we’ve been fed for years is that AI will make human workers too expensive to keep around. Turns out, at some organizations, it’s already the other way around.
How does this even happen?
Think about what it takes for compute costs to overtake salary costs. You need a lot of agents running, a lot of tokens being consumed, and probably not a lot of awareness about either.
A few questions immediately come to mind: Do the employees know they should spend tokens carefully? How many agents are running in parallel, at any given time? And who, exactly, is keeping track?
Most developers I know, including myself, work with one or two subscriptions. A Claude Pro plan here, a Cursor subscription there. That’s it. That’s the whole AI budget. And honestly, for the amount of coding I do day-to-day, it’s more than enough.
Obviously Nvidia is not most teams.
The problem of budgeting
Fees for AI software have already increased by 20% to 37% over the past year. That trend is not slowing down. The more useful these tools become, the more you use them, and the more you use them, the more you pay.
Nobody is budgeting for this properly yet. Most teams are still treating AI costs the same way they treated cloud costs ten years ago. As something IT worries about, not something to actively optimize.
What this means for the “AI replaces humans” story
Here’s the irony: if AI is already more expensive than the people it’s supposed to replace, then the economics of mass layoffs don’t actually work.
You can’t cut your salary budget by 80% and quietly absorb a compute bill that grows faster than your headcount did. The math doesn’t hold up.
This doesn’t mean AI won’t change the shape of teams. It will. It already is. But the idea that companies will simply swap people for AI agents at a lower cost is looking less and less like the obvious outcome.
In reality, the most likely scenario is that teams stay roughly the same size, but individual developers become significantly more capable. The compute bill goes up. The salary bill stays flat or grows more slowly.
Token awareness is a skill now
If there’s a practical takeaway here, it’s this: knowing how to use AI efficiently is already part of the job.
Not just knowing what to ask an LLM, but understanding when it’s worth asking at all. When to reach for an agent versus writing the code yourself. When to break a long context down versus feeding everything in at once.
These decisions have a cost attached to them. Most developers haven’t had to think about that yet, because the bills are still going to centralized budgets and nobody is asking questions. That won’t last.
The developers who understand the economics of the tools they use will have a real advantage — both in producing output that’s actually worth the cost, and in making the case for their own value.
So, are humans safe?
For now, yes. Not because AI isn’t impressive, it clearly is. But because the costs of replacing humans with AI are higher than they look from the outside.
Depending on how pricing develops over the next few years, that calculus could shift. But right now, the compute bill at Nvidia being larger than the salary bill isn’t a warning about the future of human workers. It’s a warning about AI budget management.
Spend your tokens carefully.