For the past three years, the story around artificial intelligence has been simple: AI is cheaper than people, so companies are replacing people with AI. In 2026, that story is starting to fall apart. A growing number of executives, researchers and internal budget reports are saying the opposite. For a large share of real business tasks, a human employee is still the more cost effective option, and in some cases it is not particularly close.
That contradiction sits at the center of one of the more uncomfortable conversations happening inside tech companies right now. Layoffs tied to AI keep mounting. At the same time, the people actually running AI infrastructure are admitting that the compute behind it often costs more than the salaries it was supposed to replace.
What an Nvidia Executive Actually Said
The clearest version of this admission came from inside Nvidia itself. Bryan Catanzaro, the company’s vice president of applied deep learning, told reporters this spring that the cost of running AI for his own team frequently exceeds what it would cost to simply hire people to do the same work. “The cost of compute is far beyond the costs of the employees,” he said, a striking line coming from an executive at the company that sells the chips powering the entire industry.
That is not a small detail. Nvidia has every incentive to talk up the value of AI compute, not undercut it. When someone that close to the hardware says the math does not yet favor automation, it is worth taking seriously rather than dismissing as a one-off comment.
The Token Trap: Why Cheaper AI Keeps Getting More Expensive
Here is the part that confuses a lot of people outside the industry. The price of AI tokens, the basic unit that companies are billed for when they use a model, has fallen dramatically. Blended pricing across large enterprise deployments dropped roughly 67 percent year over year between early 2025 and early 2026. By most normal logic, that should make AI cheaper to run.
Instead, total enterprise AI spending climbed sharply over the same period. The explanation is volume. A simple chatbot answering one question used to cost a few cents. An agentic AI system, the kind now being rolled out across customer service, coding and operations teams, does not answer once. It plans, checks its own work, calls outside tools, retries failed steps and loops back through the reasoning process multiple times before finishing a single task. Industry estimates put agentic workflows at five to thirty times more token-hungry than a standard conversational query.
So even as the per-token price collapses, the number of tokens burned per task has exploded. Companies that rolled out broad employee access to coding agents have described burning through an entire year’s AI budget in a matter of months once usage caught on internally. Cheaper compute and a bigger bill turned out to be perfectly compatible.
What the Research Actually Shows
This is not just anecdotal. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory built a detailed economic model looking specifically at computer vision tasks, one of the more mature corners of AI. Their conclusion was that only about 23 percent of the wages currently paid for vision related work would actually be economically worth automating with AI today. For the other roughly three quarters of that work, a human remains the cheaper, more practical choice once you account for the full cost of building, deploying and maintaining a reliable AI system.
The researchers were careful to note that this is not a permanent ceiling. If AI costs keep falling at a steady annual clip, that 23 percent figure could climb toward 50 or even 68 percent of viable wages within a decade. But that is a projection about the future, not a description of where things stand right now. The gap between what AI marketing promises and what the underlying economics currently support is wide enough that a major published study had to put a number on it.
So Why Do the Layoffs Keep Happening?
This is where the picture gets genuinely strange. If AI compute is, by the industry’s own admission, often more expensive than human labor, why have companies cut well over a hundred thousand tech jobs in 2026 while explicitly citing AI as the reason? Oracle alone has shed roughly 21,000 positions. Meta, Block, Intuit and Atlassian have all announced significant cuts tied at least partly to AI investment.
Part of the answer has little to do with whether the automation actually works yet. Telling investors that headcount is being reduced “due to AI efficiency” reads as a forward looking, technology forward story. Telling them that a company simply overhired and is now correcting course does not carry the same weight on an earnings call, even when the latter is closer to the truth in plenty of cases. Some of these companies are also the same ones pouring billions into the AI infrastructure buildout described in our look at how AI data centers are consuming enormous amounts of land, water and power across the country. Cutting payroll while ramping up capital spending on compute is, in a sense, a bet that the economics described above will flip in AI’s favor before the bill comes due.
That bet is not guaranteed to pay off. A Goldman Sachs analyst noted earlier this year that AI contributed close to nothing measurable to United States GDP growth in 2025, and a separate survey found that even though roughly 70 percent of firms have adopted AI tools in some form, more than 80 percent reported no measurable change in employment or productivity as a result. That is a hard set of numbers to reconcile with a wave of layoffs explicitly blamed on the technology.
The Real Math Companies Are Starting to Track
To their credit, more sophisticated finance teams have noticed the disconnect and are changing what they measure. Instead of just tracking total token spend, a growing number of enterprise leaders are reportedly asking for what amounts to an apples to apples comparison: the cost per resolved customer ticket, the human equivalent hourly rate of running a given AI agent, and the actual revenue generated per AI driven workflow, weighed directly against the inference cost it consumed to produce that outcome.
That shift matters because it forces a more honest conversation. It is the same kind of scrutiny that showed up when Nvidia tried to justify its own enormous data center capital spending to investors during a recent earnings call, repeating the line that “compute equals revenue” while the hyperscalers buying its chips committed to nearly 700 billion dollars in AI infrastructure spending for a single year. Investors did not exactly cheer. Shares barely moved despite revenue beating expectations, a sign that Wall Street wants proof of return, not just promises about future token generation.
What This Means Heading Into the Rest of 2026
None of this means AI is a bad investment or that automation will not eventually win out in many of these roles. The MIT projections make clear that the cost curve is bending in AI’s favor over time, and chip makers keep promising the next generation of hardware will cut inference costs dramatically. What it does mean is that the simple narrative many companies have been selling, that AI is already cheaper than people and layoffs are simply following the math, does not hold up to scrutiny right now.
For workers, that distinction matters. A layoff driven by genuine cost savings is a very different situation than one driven by a narrative aimed at investors ahead of a possible public offering, a concern that has only grown louder as questions swirl around whether the broader AI investment boom is heading toward a bubble rather than a sustainable shift in how work gets done. For now, the people who actually run the hardware are the ones telling us the most expensive part of AI might still be the AI itself.
