Uber’s Chief Technology Officer walked into 2026 with a full AI budget. It was gone before the year found its stride.
Not because of bad planning. Because of tokens.
That story, quiet, almost embarrassing, leaked from an internal report and quickly became a kind of workplace folklore. Here was one of the most sophisticated technology organizations in the world, and it had underestimated what one day of AI actually costs.
The Invisible Price Tag
For two years, the story was simple: AI is cheaper than people. Headcount goes down, automation goes up, margins improve. That was the pitch. That was the cut. Now, in the spring of 2026, something quieter is happening inside finance meetings and engineering all-hands, the math is getting complicated.
Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, the company that literally powers the AI revolution, told Axios something that stopped people mid-scroll: “For my team, the cost of compute is far beyond the costs of the employees.”
“For my team, the cost of compute is far beyond the costs of the employees.” — Bryan Catanzaro, VP Applied Deep Learning, Nvidia
.
The Numbers, Side by Side
A knowledge worker in the U.S. costs an employer significantly more than their salary alone. According to the U.S. Bureau of Labuor Statistics, total employer compensation for private-industry workers averaged $46.15 per hour in December 2025, including wages and benefits. For a standard 8-hour day, that’s roughly $369 per person, per day. For a $80,000-a-year employee, the true all-in cost, salary, healthcare, retirement, payroll taxes, and office overhead, lands closer to $104,000 to $112,000 annually. That’s about $400–$430 per working day.
Now look at AI. A Microsoft 365 Copilot Enterprise seat runs $30 per user per month, but that’s always an add-on. Add the mandatory M365 base subscription, and a 100-person enterprise deployment runs closer to $66 per user per month all-in. ChatGPT Business is $25 per user per month. Claude Pro is $20 per month. At the subscription level, AI looks almost absurdly cheap, roughly $1–$2.20 per workday.
But subscription pricing is only part of the story. The real cost lives in token consumption, the per-query charges that stack invisibly as teams actually use these systems. OpenAI’s GPT-5.2, the current flagship, is priced at $14 per million output tokens. Heavy enterprise usage, agents reading codebases, processing documents and running multi-step workflows can push actual costs ten to thirty times higher than the sticker price.
What Companies Are Actually Experiencing
A 2024 MIT study found that AI automation was economically viable in only 23% of roles studied, jobs where vision-based tasks dominate. In the remaining 77%, it was still cheaper for humans to do the work. Despite that, Big Tech firms announced $740 billion in AI capital expenditures in 2026 alone, a 69% jump from the prior year.
Simple math on one high-profile case: that $740 billion, divided across AI teams, amounts to roughly $28,000 per person per month in compute costs, likely more than the salaries of the people running the systems.
Some CEOs are reframing this. Amos Bar-Joseph, CEO of Swan AI, bragged about his Anthropic bill on LinkedIn: “We’re building the first autonomous business scaling with intelligence, not headcount.” NVIDIA’s Jensen Huang reportedly wants engineers whose productivity is measured in the AI tokens they spend, not the code they ship.
What Most People Are Missing
The real cost of AI deployment isn’t the model subscription. It’s everything wrapped around it. Hacker News engineers who’ve lived this describe the full bill as: tokens plus the engineer wrapping them, plus orchestration infrastructure, plus the supervisor, plus the eval pipeline, plus the rebuild every time a model update subtly shifts behaviour. None of that overhead disappears when AI shows up. Most of it stacks on top of the existing payroll.
A $20/month Claude subscription can meaningfully extend what a single knowledge worker produces. A sprawling agentic deployment running autonomously against a live codebase is a different cost category entirely.
The question was never AI or people. The question most organizations skipped was: which tasks, at what scale, under what oversight.
The Honest Part
Productivity gains from AI remain genuinely hard to measure. The Yale Budget Lab found no widespread data supporting the idea of AI displacing jobs at scale, yet layoffs continue, often with AI cited as the reason. Federal Reserve data shows only about 18% of companies had adopted AI tools as of late 2025. Gartner estimates 80% of enterprise AI projects fail to scale beyond initial pilots.
And then there’s reliability. One engineer’s widely shared post described an AI agent that destroyed his database and network through what he called ‘overuse.‘ Hallucinations, behavioural drift across model versions, and the constant need for human oversight are real costs that don’t appear in any pricing page.
Key Takeaways
- A fully-loaded U.S. knowledge worker costs approximately $400–$430 per working day.
- AI subscription tools cost $1–$2.20 per user per day, but enterprise token usage can multiply this dramatically.
- Nvidia’s own AI compute costs now exceed its employee payroll for applied deep learning teams.
- A 2024 MIT study found AI automation economically viable in only 23% of vision-heavy roles.
- The real expense isn’t just the subscription fee. It’s the orchestration layers, human oversight, eval pipelines, and repeated rebuilds stacked on top of it.
- The companies winning aren’t asking AI vs. human. They’re asking which tasks, at which scale, for which workflows.
The Reckoning
There’s a version of this story that’s embarrassing for the companies that laid off thousands of people, pointed at AI, and then quietly discovered the token bill was bigger than the payroll. But that version misses something.
The more accurate read is that we are mid-experiment, and the numbers are finally honest. AI isn’t uniformly cheaper. It isn’t uniformly smarter. It isn’t a replacement. It’s a capability, one with a real cost structure that rewards precision over enthusiasm.
The organisations that will get this right aren’t the ones spending the most on tokens. They’re the ones asking the question most companies are still too excited to ask:
What exactly are we paying for, and is it actually working?
