There is a particular kind of corporate awakening that arrives with an invoice. For two years, large companies handed engineers powerful AI tools and cheered as they used them. The cheering has now collided with the accounting. Microsoft, according to TechRadar, has begun pulling back internal licences for Anthropic's Claude Code, which was costing between 500 and 2,000 dollars per engineer each month, and is moving one division to GitHub Copilot's command-line tool ahead of its fiscal year-end. Uber, which deployed Claude Code to roughly 5,000 engineers and stoked adoption with internal leaderboards, found the experiment worked almost too well: assisted systems generated nearly 70 percent of committed code, but token usage scaled so fast the company exhausted its entire AI coding budget in four months.
This is the texture of what executives in India are calling AI's "year of truth." The pattern is consistent across the giants. Meta has made cutting inference costs a strategic priority. Amazon, Walmart and Cisco have introduced usage caps or routed staff to cheaper models. The shift is less a retreat than a change of question. "Token consumption became a vanity metric during the first wave of enterprise AI," Deepak Dhanak of the platform Rocket told The Times of India. "The overspend is not an AI problem, it is a measurement problem. Businesses now need to match the right model to the right task and measure outcomes, not activity."
The discipline is being forced partly by pricing. As Forbes reported, several AI services recently switched from flat subscriptions to usage-based billing, and the result was sticker shock: ballooning costs "with little productivity gain to show for it." That has pushed many enterprises to discover that open-source and in-house models are often good enough, and far cheaper, for routine tasks. The newsletter's framing was sharp: "What does it profit a company to build AGI if nobody can afford it?" The pressure runs both ways, landing on the labs that need their customers to pay more before those customers can see the value.
Curiously, the scrutiny is arriving alongside evidence that the most feared consequence of AI may be overstated. A study published this week by the startups Ramp and Revelio Labs, tracking 22,000 US firms from 2021 to early 2026, found that the heaviest AI adopters are not slashing jobs at all. They grew headcount about 10 percent faster than peers after adopting the technology, and entry-level hiring rose 12 percent. "If you are reading headlines where CEOs blame layoffs on AI, be skeptical," the authors wrote. Nvidia's Jensen Huang has called such claims "lazy"; even Sam Altman has played down talk of a jobs apocalypse.
The two findings fit together more neatly than they first appear. The companies spending most on AI tend to be larger and faster-growing, so they hire, but they are also the ones now scrutinising every token. What both trends puncture is the same fantasy: that AI is a switch you flip to make costs vanish and output soar. The reality is a productivity paradox, where workers finish tasks faster without the gains reliably showing up in revenue or profit. The winners of this phase will not be the firms that spent the most or fired the most. They will be the ones that figured out, invoice in hand, which AI was actually worth paying for.