The era of large language model companies subsidizing AI subscriptions at a loss is accelerating toward its end. Analyst Hedgie emphasizes that the real cost per token has far exceeded what enterprises expected, forcing the industry into a dilemma: either companies slash agent usage due to budget overruns, or model vendors cut prices and absorb the losses themselves. Either path ultimately leads to writedowns.
https://twitter.com/HedgieMarkets/status/2057531661785628841
Recent budget overruns at Microsoft and Uber have exposed the cost pain of AI agents. At Uber, adoption of AI tools drove per-engineer monthly API costs to between $500 and $2,000, burning through the company's entire AI budget for 2026 in just the first four months. Microsoft's Experiences + Devices division also announced it would terminate its Claude Code license by June 30. These cases show that when enterprises shift from traditional sentence autocomplete to multi-step autonomous reasoning agents, per-token consumption increases exponentially, completely breaking the old flat-rate subscription business model.
The industry now faces two possible paths forward. On one path, if companies cut back agent usage to meet budgets, LLM developers will see slowing revenue growth, making it harder to justify high pre-IPO valuations. On the other path, if model vendors lower prices to maintain usage, their unit economics will worsen at a particularly unfavorable time. GitHub Copilot's move to a usage-based AI Credits model starting June 1, 2026, marks a turning point where compute cost pressure is being passed to end users. Eventually, the bill for all that compute will have to be paid through writedowns by one side or the other.