AI Is Now More Expensive Than Paying Real Humans, and Even Nvidia Is Surprised
So there we were, scrolling through the morning headlines, coffee in one hand, mouse in the other, and then, screech. We nearly spat out our coffee.
The headline read: “‘The cost of compute is far beyond the costs of the employees’: Nvidia executive says right now AI is more expensive than paying human workers.”
Not from a sceptic. Not from a union leader. That was Bryan Catanzaro, Nvidia’s vice president of applied deep learning — a guy whose entire career is about making AI work, telling Axios that for his own team, the math has flipped. If the people who sell the compute are saying it’s too expensive, maybe we should all sit down and listen.
And it gets worse, or funnier, depending on where you sit.
Wait, AI Was Supposed to Save Us Money. What Happened?
Like a lot of you, we’ve watched the narrative shift. First it was “AI will replace all the jobs” (cue dread). Then it softened to “AI will be your co-pilot” (cue relief). Now, somewhere in late April 2026, the conversation has landed on something far more awkward: AI is actually bleeding corporate budgets — and it’s bleeding them faster than a human payroll ever did.
Meta just announced plans to cut 10% of its workforce, about 8,000 people, while scrapping 6,000 open roles, all in the name of efficiency and AI investment. Microsoft offered thousands of employees its largest-ever voluntary buyout. The optics look like “machines are taking over.” But the ledger is telling a messier story.
Uber’s CTO, Praveen Neppalli Naga, revealed that his team burned through the company’s entire 2026 AI budget in just a few months. He’s now “back to the drawing board,” rethinking everything.
“I’m back to the drawing board because the budget I thought I would need is blown away already.”, Uber CTO
This isn’t a tech glitch. It’s a pattern.
By the Numbers: When 1 + 1 = Empty Wallet
Let’s lay out the receipts. You’ll want to sit down for these.
The MIT Study That Predicted This Mess
Back in 2024, a team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) asked a beautifully simple question: Is AI actually cheaper than human workers?
They studied jobs where computer vision is central, inspecting parts on an assembly line, reading medical scans, that sort of thing. Their finding? Only 23% of these tasks could be automated at a cost that beat human wages. In the other 77% of cases, installation, maintenance, and operational costs made humans the cheaper option.
Yes, really. Three out of four times, it’s still cheaper to pay a person.
Inside AI Labs: Compute Eats Everything
Now peek inside the companies that build the AI models. Epoch AI broke down spending at Anthropic, Minimax, and Z.ai. The numbers are staggering: compute costs account for 57% to 70% of total spending at these firms, dwarfing every single other expense, including those legendary Silicon Valley salaries.
Anthropic alone is estimated to have spent $9.7 billion in 2025, with roughly $6.8 billion of that going just to compute (training and inference combined).
Even setting aside the AI labs, the broader picture is intense. Morgan Stanley reports Big Tech has announced $740 billion in AI capital expenditures so far in 2026, a 69% leap from 2025. And according to Gartner, worldwide IT spending will hit $6.31 trillion in 2026, a 13.5% year-on-year jump driven almost completely by AI infrastructure.
That’s not a market. That’s a flood.
The Hidden Cost No One Put in the PowerPoint
Here’s where the story gets human, messy, and, honestly, a little absurd.
Somewhere along the way, software engineers started treating their token consumption like a high score in an arcade game. The slang for it is “tokenmaxxing,” and we regret to inform you it’s a real thing.
Developers run multiple AI agents simultaneously, unsupervised, churning through tasks while the meter ticks ruthlessly in the background. One Stockholm-based developer, Max Linder, told The New York Times: “I probably spend more than my salary on Claude.” Some power users are racking up monthly token bills north of $150,000.
When Nvidia CEO Jensen Huang floats the idea of “half your salary in AI tokens” as a recruiting perk, you start to grasp the scale of the problem. Huang suggested that a $500,000 engineer should consume at least $250,000 worth of tokens a year, and that he’d be “deeply alarmed” if they didn’t..
Think about that for a second. The world’s most influential AI visionary is telling companies they should budget an extra quarter-million per engineer, to do… what, exactly? Be more productive? Automate tasks? Win at tokenmaxxing leaderboards?
The rational incentive is unclear. The spending, however, is all too real.
So… Is AI Ever Going to Be Cheaper?
Yes. Almost certainly. But not today, and maybe not next quarter.
The cost per token is falling fast, a kind of Moore’s Law for model inference. One research group noted that the cost of running GPT-3.5-level models has dropped from about $20 per million tokens in late 2022 to roughly $0.07 now, a nearly 300× decline in three years. Some forecast that AI training and inference efficiency will keep compounding.
Sam Altman, OpenAI’s CEO, has predicted AI usage costs could continue dropping 90% annually.
But here’s the twist: as tokens get cheaper, usage explodes. It’s the classic Jevons paradox, make a technology more efficient, and people consume exponentially more of it. The total bill doesn’t shrink; it grows.
Jensen Huang wants to turn tokens into a kind of corporate currency, a fourth element of compensation alongside base pay, equity, and benefits. His bet is that the ROI will eventually flip. For now, though, it’s a leap of faith backed by very little productivity data. As the Yale Budget Lab notes, there is no widespread evidence to support the idea that AI is currently displacing human jobs at scale.
A Survival Kit: How to Think About AI Spend Right Now
If you’re a founder, a team lead, or just someone trying to make sense of this for your own career, here’s a practical, non-panicked way to frame things.
1. Ask the Only Question That Matters
“Is this AI tool augmenting a valuable human, or is it trying to replace one I didn’t need to replace in the first place?”
Augment → likely worth it after a pilot. Replace → run the maths ruthlessly, using real token-cost projections, not vendor slide decks.
2. Remember the 23% Rule
Right now, the data says roughly 1 in 4 tasks pencils out cheaper with AI. Not zero, but not most. Identify the jobs in your world that sit squarely in that 23% sweet spot. Vision-based inspections? Maybe. Summarising a thousand PDFs? Probably. The rest? Keep the human in the loop and sleep better.
3. Beware the Tokenmaxxing Trap
Token leaderboards create bragging rights, not ROI. If you’re evaluating AI tools for your company, look for usage guardrails and seek tools that let you set hard limits per user. Uber’s four-month budget blowup is a cautionary tale, not a case study to emulate.
4. Think Like a CFO (Even If You’re Not One)
Fixed salaries are predictable. Token costs are shockingly, horribly variable. Every AI query, every agent loop, every “just let it run overnight” job adds to a bill that doesn’t ask permission. If you can’t cap it, you can’t control it.
We’re in a strange, transitional moment. The AI revolution isn’t a hoax, but neither is the price tag.
Nvidia’s own VP just admitted what many suspected: right now, it’s cheaper to hire a human being than to offload their work to a machine in most cases. MIT’s researchers crunched the numbers and landed in the same place. And some of the world’s biggest companies are discovering this only after the budget has gone up in smoke.
As one wry observer noted: “Maybe human labour will be more cost-efficient after all.”
Maybe we’ve been so busy worrying about AI taking our jobs that we forgot the oldest business truth in the book: you can’t replace a bicycle with a Ferrari and call it a savings plan.