Polymarket called it first. Forbes just ran the numbers on AI costs versus human salaries. And the numbers are worse than you think.
Polymarket posted “JUST IN: Forbes analysis reveals companies are spending more on AI than on the workers it replaced.” The post went viral for a reason. The Forbes analysis, from Forbes contributor Dr. Jemma Green published July 2, lays out the full picture across the biggest names in tech. Uber. Microsoft. Nvidia. Amazon. The story is the same everywhere.
Companies are spending more on AI inference and tokens than on the salaries of the workers they laid off to fund it.
The AI costs numbers that matter
Let me walk through the data points that stood out to me.
Uber burned its entire 2026 AI coding budget in four months. By March, 84% of engineers had adopted Claude Code. Roughly 70% of committed code now originates with AI. But here’s the part that matters: token usage did not correlate with features shipped. Uber’s COO Andrew Macdonald said publicly that the spending wasn’t translating to user-facing value.
Microsoft told an entire division to stop using AI coding assistants. The bills became untenable. This is a company spending roughly $80 billion on AI infrastructure telling its own engineers to pull back because the token costs outpaced the value.
One company ran up a $500 million Claude bill in a single month. Management forgot to set a usage cap. Five hundred million dollars. In one month. Because nobody configured a throttle.
Nvidia’s own VP of applied deep learning, Bryan Catanzaro, said compute costs for his team now exceed what the company spends on the employees using it. The company that manufactures the hardware powering the AI revolution is telling you the technology costs more than the people it replaced.
And yet Jensen Huang, Nvidia’s CEO, says a $500,000 engineer should consume at least $250,000 worth of AI tokens annually. Nvidia is working toward a $2 billion annual token budget for its engineering force. Spend more, faster, from the top of the supply chain.
The tokenmaxxing culture
Amazon built an internal leaderboard called KiroRank to track AI usage among engineering teams. Employees gamed the system by burning tokens on meaningless tasks just to climb the rankings. Amazon quietly took it down.
Meta built a similar tracker called Claudeonomics.
Amazon had encouraged its teams to “tokenmaxx,” treating AI token consumption itself as a performance metric. When you reward people for how much they spend rather than what they produce, spending becomes the output.
The Forbes analysis reports that roughly 95% of enterprise AI usage still runs on the most expensive frontier models, even for work that doesn’t need that level of sophistication. The default is the priciest option.
The macro picture
Alongside this spending frenzy, more than 115,000 tech workers have been laid off in 2026 across over 150 companies. Meta cut 8,000. SentinelOne cut 8% of its workforce to redirect resources toward AI. Wix cut a fifth of its people. Block halved its headcount.
Big Tech has announced $740 billion in capital expenditure this year, up 69% from 2025. Gartner projects AI agent software spending alone will reach $207 billion in 2026, up 139% from the prior year.
But the economic viability is not there yet. An MIT study found AI automation is economically viable in only about 23% of roles. For the remaining 77%, humans remain cheaper. Goldman Sachs’ chief economist has stated he does not view AI as strongly growth-positive. Sequoia Capital’s David Cahn estimates AI companies need roughly $600 billion in annual revenue to justify current infrastructure spending. The gap is widening, not closing.
The pricing reckoning
The prices companies are paying for AI usage are not real prices. OpenAI, Anthropic, Google, and Meta are all pricing inference below the cost of serving it. As I covered in the OpenAI Financial Leak, OpenAI loses roughly two dollars for every dollar it earns on inference. Sam Altman has acknowledged the company loses money on its $200 per month subscriptions.
Anthropic moved enterprise customers from flat-rate plans to usage-based billing in April 2026. GitHub followed for Copilot. Analysts project enterprise AI bills will rise another 30 to 50% when pricing normalizes to reflect actual infrastructure costs.
OpenAI’s own projections show $14 billion in losses this year, with $44 billion in cumulative losses before any path to profit in 2029.
The market noticed
In June 2026, chipmakers lost roughly $1.3 trillion in market value in a single session. The PHLX semiconductor index suffered its steepest one-day drop since the pandemic crash of March 2020. Nvidia, Micron, and AMD led the losses. South Korea’s KOSPI fell 10% in a single day and briefly halted trading. SpaceX slid below its IPO price within days of listing. Accenture is down 52% in six months.
The selloff was not a verdict on the technology. It was a verdict on the timeline. Ray Dalio has described the current moment as the early stages of a bubble. The parallel to the late 1990s is instructive. The internet was real technology. It still produced a crash.
Bottom line
Companies are spending more on AI than on the workers they replaced. The companies making the hardware admit it. The companies selling the software acknowledge it. The market is starting to price it in.
The question that matters is not whether AI is transformative. It is whether the current AI costs justify themselves before the money runs out. The Forbes analysis suggests the industry is nowhere near that point yet.
And Polymarket called it before the article even published.
