Companies Traded People for Tokens. The Returns Haven't Shown Up
80% of $1B+ companies have cut staff to fund AI compute, billed in tokens per call. Gartner's 350 executive survey found no correlation between those cuts and improved returns.
80% of $1B+ companies have cut staff to fund AI compute, billed in tokens per call. Gartner's 350 executive survey found no correlation between those cuts and improved returns.
A US$500,000-a-year engineer should burn US$250,000 in AI tokens annually, Jensen Huang told the All-In Podcast at GTC 2026. Nvidia is working toward a US$2 billion yearly token bill for its own engineering force. Read those numbers as a budget reallocation, not a productivity target: money is moving out of human capital and into compute, priced in tokens, the unit AI vendors charge per call or per million model invocations.
"Token" here does not mean cryptocurrency. It is the chunk of model usage a chatbot, code assistant, or autonomous agent consumes, billed in fractions of a cent. The substitution Huang is sketching, a quarter-million dollars of tokens per senior engineer, is the line item the rest of corporate America is racing to fund. The data on whether that substitution is paying off is starting to come in.
In May, Gartner surveyed 350 executives at companies with more than US$1 billion in revenue that are deploying AI agents or automation. Roughly 80 percent had cut headcount as part of the AI rollout. The finding that broke through the AI-productivity hype: among those companies, there was no statistical correlation between the scale of headcount cuts and improved returns. "Workforce reductions may create budget room," Gartner analyst Helen Poitevin said, "but they do not create return." Among the 80 percent that cut headcount, the companies that reported improved returns on AI investment were, according to Gartner's data, disproportionately those that used AI to amplify the people they kept.
Challenger, Gray & Christmas reported that AI was the most-cited reason for US job cuts for a record fourth consecutive month, with tech accounting for 31 percent of first-half layoffs in 2026. June layoffs cooled to 45,849, down 53 percent from May, but AI still led cited reasons. "Companies are shifting budgets toward AI investments at the expense of jobs," Challenger's Andy Challenger said.
The four largest US hyperscalers, the cloud platforms that build the data centers AI runs on, have guided roughly US$700 billion in combined 2026 capital expenditure — an aggregate of individual company guidance figures, nearly double the prior year. Gartner separately projects AI agent software spending will reach US$207 billion in 2026, up 139 percent year-over-year. Together those numbers describe a single corporate flow: payroll dollars leaving the operating budget and reappearing as AI capex and software spend on the other side of the income statement.
An internal Meta memo, reported via Reuters and summarized in industry coverage, described the May 2026 cut of 8,000 roles as offsetting substantial AI investment, even as the company posted 33 percent quarterly revenue growth. Oracle filings show headcount down 21,000 as savings feed its data centre buildout. Klarna, which cut roughly 700 jobs for AI, has reportedly reversed course and started rehiring. Uber, according to one trade-press account citing a LinkedIn post from Uber's Head of AI Products, reportedly burned through its entire 2026 AI budget by April, prompting its chief operating officer to publicly question the return.
The pattern that emerges is not "AI doesn't work" or "AI will replace workers." It is more specific than either. Companies that treated the AI line item as a headcount substitute, fewer bodies and more tokens, have so far failed to translate the swap into measurable ROI. Among the surveyed cohort, Gartner found no correlation between the depth of headcount cuts and improved returns. The companies that did report improved returns, Gartner found, were those using AI to amplify the people they already had. The substitution framing has been the wrong mental model for a meaningful share of the S&P 500's AI programs.
Huang's GTC keynote and the token-economics session that followed made the math explicit. The next test is the next earnings cycle. If hyperscalers start reporting AI revenue at the same line-item detail they report capex, the substitution case either gets its first quantitative support or its first quantitative refutation. The companies that get there with intact headcount will have had a head start on the ones that didn't.