Giant robot throws man into a trash can. Artificial intelligence replaces jobs. Vector illustration.
getty
Something strange is happening in the tech world right now: the technology that was supposed to make human labor obsolete is currently more expensive than the people it was supposed to replace. Companies are laying off workers to fund the very AI tools that cost more than the workers they just laid off. Circular logic would be darkly comical if tens of thousands of existences weren’t caught in the middle.
Uber’s CTO recently revealed that the company burned through its entire 2026 AI coding budget in four months. As of March, 84 percent of Uber engineers had adopted Claude Code, and around 70 percent of the code handed over now comes from AI. The consumption was enormous. The corresponding value was more unclear. Uber COO and President Andrew Macdonald publicly acknowledged that token usage does not appear to directly correlate with the useful features provided to users.
Uber is not an outlier. Microsoft, which has invested about $13 billion in OpenAI and writes up to 30 percent of its own code using generative AI, ordered engineers in a large department to stop using an AI coding assistant because the calculations were becoming untenable. An unnamed company billed $500 million in Claude in a single month after management forgot to set a usage cap, according to Axios. These are structural misjudgments about what intelligence costs if you buy it syllable by syllable.
Bryan Catanzaro, Nvidia’s vice president of applied deep learning, put it succinctly: His team’s computing costs now far exceed the company’s spending on the employees who use it. The company that makes the hardware powering the AI revolution admits the technology is more expensive than the people it was designed to empower.
And yet Catanzaro’s boss Jensen Huang tells the industry that a $500,000 engineer should consume at least $250,000 worth of AI tokens annually and that Nvidia is working toward a $2 billion annual token budget for its engineering workforce. He suggested that tokens should be a recruiting advantage. The message from the top of the supply chain is clear: Spend more, faster.
Companies have committed. Big Tech has announced $740 billion in investments this year, a 69 percent increase from 2025. Gartner expects spending on AI agent software alone to reach $207 billion in 2026, up 139 percent year over year.
This is where the arithmetic becomes perverse. In addition to these expenses, more than 115,000 tech workers were laid off at more than 150 companies in 2026. Meta eliminated 8,000 jobs. SentinelOne has cut 8 percent of its workforce to redirect resources to AI. Wix has laid off a fifth of its employees. Block halved its workforce. Atlassian has cut 1,600 jobs.
Automation worker concept with 3D rendering robot working in office
getty
The rationale presented is consistent: operational efficiency, redistribution towards AI. However, the MIT study found that AI automation only makes economic sense in about 23 percent of roles. For the remaining 77 percent, people remain cheaper. The chief economist at Goldman Sachs has made it clear that he does not see AI investments as being highly growth-enhancing. Sequoia Capital partner David Cahn has quantified the resulting gap: AI companies need around $600 billion in annual revenue to justify current infrastructure spending. From mid-2026, the gap will widen, not narrow.
So the current moment looks like this: Companies are reducing human labor to fund artificial intelligence, which currently costs more than the labor it replaces, and are pursuing productivity gains that most studies cannot yet verify, at a pace that is depleting annual budgets in weeks.
The cultural dimension is perhaps the most telling part. Amazon created an internal leaderboard called KiroRank to track AI usage in development teams. It was quietly phased out after employees started playing it – burning tokens on meaningless, stupid tasks just to move up the rankings. Meta has developed a similar tracker called Claudeonomics. Amazon encouraged its employees to “tokenmaxx” and viewed consumption itself as a performance indicator. When you reward people for how much they spend rather than what they produce, spending becomes the outcome.
Boards of directors demanded that their CEOs introduce AI. Then came the indiscriminate use – what the industry calls “tokenmaxxing”. In the third phase, leadership teams weigh in and ask a long-overdue question: Does every task actually require the most expensive model available? Approximately 95 percent of enterprise AI usage still occurs on the most expensive frontier models, even for work that does not require this level of complexity.
When a resource becomes so cheap that it can be wasted, people waste it without thinking about it. When it becomes expensive enough to matter, they suddenly become passionate about efficiency. Artificial intelligence seems to be headed for the same bill – except the waste is numbered in the billions and comes with a hefty monthly bill.
The private equity side of the market comes to the same conclusion from the opposite direction. At the Crypto Valley Conference private equity panel, Giuseppe De Filippo (Head of Private Capital Markets at Julius Baer) explained that SaaS transactions are stalling because horizontal pricing is no longer working and valuations are not reflecting this reality.
AI can now generate a usable interface in hours, meaning the design layer that companies spent years polishing is worth less than it was a year ago. What AI cannot generate is twenty years of domain logic integrated into a niche ERP system for a mining operation or a water utility. The moat has shifted from what software looks to what it knows.
For a long time it was assumed that costs would fall. Prices per token have actually fallen, and Gartner predicts that the largest models could be nearly ninety percent cheaper to run by 2030. The catch is that consumption has increased faster than prices have fallen. A study by Faros AI found that “code churn,” meaning lines of code deleted compared to lines of code added, increased by more than 800% when AI adoption was high. More tokens in, more work thrown away.
The prices companies currently pay for AI use are not real prices. OpenAI, Anthropic, Google and Meta all price below deployment costs and burn venture capital to acquire market share. OpenAI spends nearly two dollars for every dollar it earns from inference. Sam Altman publicly admitted that the company was losing money on its $200 per month subscriptions. The dissolution of the subsidy model began this year.
The spending history and the return history were separate from each other. For years, subsidized conclusion prices, venture capital-funded losses, and the promise of ultimate productivity kept both trajectories moving in the same direction. In June 2026, the market noticed that they were diverging. Chipmakers lost about $1.3 trillion in market value in a single session, the largest one-day decline in the PHLX semiconductor index since the pandemic crash in March 2020. Nvidia, Micron and AMD led the losses. South Korea’s benchmark index fell 10% within a day and briefly halted trading. SpaceX slipped below its IPO price within days of listing. Accenture is down 52% in six months. The selloff wasn’t a judgment on technology. It was a judgment about the schedule.
TOPSHOT – A foreign exchange trader reacts as she monitors exchange rates in a foreign exchange trading room at Hana Bank’s headquarters in Seoul on Feb. 2, 2026. South Korea’s benchmark Kospi index plunged more than five percent on February 2, in line with a sell-off in Asian markets, amid fresh worries about an AI-powered technology rally that has sparked fears of a bubble in the sector. (Photo by Jung Yeon-je / AFP via Getty Images)
AFP via Getty Images
In April 2026, Anthropic transitioned enterprise customers from flat-rate plans to usage-based billing tied to actual computing power. GitHub followed weeks later with the same shift for Copilot, after years of quietly soaking up up to eight times the subscription value for heavy users. Analysts predict that companies’ AI bills will rise another 30 to 50 percent above current levels as prices normalize to reflect true infrastructure costs.
The profitability path requires either increasing prices or reducing computing and energy costs faster than consumption. Neither happens. OpenAI’s own forecasts call for losses of $14 billion this year, with cumulative losses of $44 billion before turning a profit in 2029. Ray Dalio has described the current moment as the early stages of a bubble. The parallel with the late 1990s is revealing. The Internet was real technology. And there was still a crash.
If the person selling the computing power calls the expense “the most valid criticism at the moment, there’s a lot of waste,” what does the person paying the bill call that? If the cost of tokens has already exceeded the cost of the employees they were supposed to replace, when does the comparison start in the other direction? And if the answer is that costs will eventually fall enough to close the gap, the follow-up question is: Who will bear the losses in the years leading up to that point?
History has already outlined the answer: the Internet existed, it still crashed, and what followed was no less Internet – it was the Internet finally paying for itself. The AI is heading towards the same sorting, and the gap is already visible. Lisa Emme, co-founder of Inversion AI, says: “The mistake is treating AI as a capability that you add on. AI-native companies rebuild around the model – and once you do that, you stop paying premium prices for the work that a specialized model does better and cheaper.
This is the duller future that the boom didn’t price in – not how much intelligence you can buy, but how much you can use. The industry has answered all questions about the possibilities of AI. It hasn’t answered the only question that matters now: whether it will pay for itself before the money runs out.
https://www.forbes.com/sites/jemmagreen/2026/07/02/ai-costs-more-than-the-people-it-replaced/
