According to Fortune, big firms are discovering that running AI often costs multiples of employee pay. The price tags come from exploding compute bills — think GPU hours, storage, and network egress — plus recurring token charges, and the extra layers of infra that support everything.
Microsoft cut back on licenses for the AI assistant Claude Code from Anthropic, moving engineers toward its own GitHub Copilot CLI. The strategic tie between the two companies hasn't disappeared: Microsoft put in $5B and Anthropic agreed to buy roughly $30B of Azure capacity (approx. numbers).
At Uber, the CTO burned through the company’s AI budget for 2026 in just four months — token costs were the culprit. That’s notable because Uber had been actively pushing generative AI internally, even ranking departments by how intensely they adopted the tech.
At Meta (recognized as extremist and banned in Russia) there was an internal leaderboard called “Claudeonomics” that tracked usage of Claude models.
At Amazon, employees were told to maximize their use of AI tokens.
Brian Catanzaro, Vice President of Applied Deep Learning at Nvidia, said his org already spends more on compute than on people.
Goldman Sachs projects a ~24x jump in global token consumption by 2030 — up to 120 quadrillion ops/month. Even if the cost per token drifts down, overall bills rise as agents scale and their task lists grow.
Analysts at Gartner expect that by decade’s end launching models with ~1 trillion parameters will be ~90% cheaper vs. 2025. But cheaper-per-param can be eclipsed by sheer load: agent-style models tend to chew through far more tokens than older NN setups.
Reasons for rising costs:
- IT & cloud infra getting pricier (capex/opex pressure);
- Token-based billing becoming the main driver of spend (per-use fees add up);
- Algorithmic complexity and the overhead of AI agents (more state, more requests).
Boards and investors increasingly demand clear ROI and pilots before greenlighting scale-ups. Instead of wholesale headcount removals, many orgs are hedging with hybrids where AI augments FTEs. Expect uneven rollouts: some teams will double down, others will throttle back while finance models the real run-rate and total cost of ownership.