Here is a number that should give any manager pause. At Nvidia, the company that makes the chips much of the AI boom runs on, one senior leader says the computers can cost more than the people. Bryan Catanzaro, Nvidia’s vice president of applied deep learning, put it plainly: “For my team, the cost of compute is far beyond the costs of the employees.”
That does not prove AI is more expensive than workers everywhere. Catanzaro was talking about his own team, inside a company doing unusually compute-heavy AI work. But it lands oddly because so much of the public conversation has assumed the opposite: that AI is the cheap substitute and people are the expensive part.
To be clear, I’m a writer, not an economist or a technologist. What follows is me reading the reporting and thinking out loud about my own work. The studies and examples I lean on below are findings from particular companies and particular moments in a fast-moving market, not settled laws about how any of this will shake out.
Why the swap felt obviously true
A human costs a salary, benefits, holidays, sick days. A model costs a subscription (sometimes, more on that later). You point the model at the task, cancel the salary, keep the difference. Some companies have encouraged that picture directly. Klarna said in 2024 that its AI assistant was doing the equivalent work of 700 full-time customer service agents, and headlines like that made the labour-saving case feel almost automatic.
There is also evidence that companies have started pointing to AI when cutting jobs. CBS News, citing Challenger, Gray & Christmas, reported that companies directly cited AI in announcing 55,000 job cuts in 2025. So the idea was not just internet panic. It was showing up in corporate language too.
The trouble is that the clean version assumes AI is priced like software you buy once and use forever. A lot of it isn’t. The powerful tools, especially when used through APIs or agentic coding systems, are often billed by usage. OpenAI’s API pricing, for example, bills tokens at different input and output rates. Anthropic’s Claude pricing is also listed by input and output tokens. That means you pay for what the model reads and writes. The more useful the tool, the more people use it. The more they use it, the more the meter runs.
What the numbers actually show
Uber is one of the clearer examples of what this can do to a budget. According to reporting cited by TechCrunch, Uber put a monthly cap of $1,500 per employee per agentic coding tool after blowing through its AI budget in four months. That is not the profile of a tidy fixed-cost software swap. It is a usage problem: unpredictable spend dressed up as automation.
This is the budget trap. A traditional software licence may be annoying, but at least it is relatively predictable. Token-based tools can behave differently. A developer who asks an AI agent to inspect a codebase, debug a problem, run tests, explain errors, rewrite code, and try again is not making one neat request. They are setting off a chain of model calls. In some products, even the reasoning or “thinking” process can add billable output tokens. Anthropic’s Claude Code documentation says thinking tokens are billed as output tokens, and that the default budget can be tens of thousands of tokens per request depending on the model.
None of this means AI is always more expensive. In some narrow, repeatable tasks, the economics may already be compelling. But the evidence has limits worth naming. A 2024 MIT FutureTech study looked specifically at computer vision tasks, where cost modelling was more developed. The researchers found that, at the costs of the day, only 23 percent of worker wages being paid for vision tasks would be attractive to automate. They were clear that this was about computer vision, current costs, and a picture that could shift if AI gets cheaper or is deployed at larger scale.
What this does and doesn’t mean for jobs
So does any of this settle the question everyone actually cares about, which is whether the job survives? Not really. The compute-versus-payroll story is a snapshot of a mismatch that may reverse as the cost of running these models falls. One executive’s team is not a general law. One company’s coding-tool budget is not the whole economy.
What worries me personally isn’t only AI being cheaper than me. It’s companies choosing cost over quality even when the sums don’t clearly favour it. There is a version of this where the AI bill itself becomes the pressure. Spend too much on the tool, struggle to prove the return, then look for savings somewhere else. That does not mean every AI investment leads to layoffs. But it is a plausible way for this to go wrong: pay too much for the machine, then cut people to make the spreadsheet work.
My own bet runs the other way. As the web fills with fast, generic, AI-generated content, quality becomes the thing that stands out. That points to using these tools to get better, not merely to make the job easier. Those are not the same goal. I’ve come to think a lot of career paths quietly go obsolete without continuous learning, and I’d rather treat AI as the thing that sharpens the work than the thing that hollows it out.
The post Most people think swapping a worker for AI saves money, but for some companies the compute bill has quietly overtaken the payroll it was meant to trim appeared first on Space Daily.