The end of "politically priced" AI

For two years, the big labs burned investor capital to offer below-cost computing power, creating usage habits that now must be paid for at full price. Global AI spending in 2026 is projected at around $2.5 trillion, of which $401 billion for infrastructure alone.

Microsoft has already begun cancelling numerous internal Claude Code licenses because the cost exceeded the salary of the developers using it. The era of subsidized AI is officially over.

From "tokenmaxxing" to the reckoning

For months the vendors' mantra was: maximize token usage. Every request was seen as a learning opportunity and a competitive advantage. Meta itself pushed its engineers in this direction, only to later hear its own Chief Technology Officer admit that AI shouldn't be used just for the sake of using it.

Now that phenomenon, dubbed "tokenmaxxing," has become a governance problem. Uber has publicly admitted that its head of operations increasingly struggles to justify AI spending in the face of productivity gains that are hard to measure.

Within a year, the share of organizations that consider AI a critical line item in FinOps management rose from 31% to 63%. AI and machine learning workloads now account for 22% of total cloud costs at SaaS and IT companies, in many cases exceeding traditional hardware spending.

The problem isn't the price, it's not knowing what it's worth

The most common mistake is thinking that "AI is too expensive." The real problem is that the vast majority of companies don't know how much each individual use case is worth. Without a clear KPI defined from day one of the pilot, every invoice becomes a shock.

The MIT "GenAI Divide" report (July 2025) revealed that 95% of enterprise pilots generate no measurable ROI. Gartner estimates that between 30% and 40% of these projects are abandoned not because the technology doesn't work, but because no measurement system was built before opening the spending tap.

The great migration toward the rational

This growing awareness is pushing many companies to redistribute their workloads:

Open-source or open-weight models for less critical tasks

Smaller, specialized models

Frontier models only where added value is proven

It's the same path already seen in the cloud: not everything needs to sit on the most expensive hyperscaler.

Who will win in the next 12 months

The dividing line will be clear-cut between two types of companies:

Those who built metrics – will be able to renegotiate contracts, demonstrate value, and replace overly expensive vendors.

Those who only built habits – will face a brutal choice: pay whatever it costs or shut everything down, risking having to relearn how their developers work without AI.

Companies aren't abandoning AI. They're simply entering the mature phase: the one where they must prove they actually know what to do with it.

Exactly as happened with the cloud between 2018 and 2022.

Those who have already seen that movie will handle this better. Those who haven't will pay for the ticket twice.