Last month, I shared thoughts on the looming AI bubble, suggesting it was nearing collapse as subsidies dwindled and the financial burden shifted onto consumers. The analysis was rooted in a troubling reality: AI vendors were masking this truth, impacting most serious users.
I had anticipated it would take a couple of years for the repercussions to manifest—maybe longer. Yet, within weeks, signs of distress emerged. Where I misjudged was the initial impact point; I expected smaller businesses to feel the strain first, with rising renewal fees. Instead, it was the largest corporations, many of whom produce and market AI technology, that began to flinch.
The message circulating was clear: “Join us or risk being left behind.”
Uber exhausted its entire AI budget in just four months
Praveen Neppalli Naga, Uber’s CTO, mentioned that by April, they had depleted their AI coding budget for 2026 in just four months. The reason? Anthropic’s Claude Code. They rolled it out to engineers in December 2025, and by February, adoption had surged from 30% to over 80%. According to Naga, a third of the planned budget vanished, returning him to square one. The monthly cost per engineer fell between $500 and $2,000.
In a “Rapid Response” podcast, Uber president Andrew MacDonald acknowledged that the link between token expenditure and visible customer benefit was hard to see. When asked if AI was enabling greater functionality, he candidly remarked: “That link doesn’t exist yet.” Once AI’s costs are on par with usage, justifying expenses will become more complicated.
In 2025, Uber recorded $3.4 billion in research and development losses, with a significant portion attributed to AI. Now, they’ve set spending limits. Employees receive tokens worth $1,500 monthly for each coding tool, yet the company is uncertain about the actual value purchased.
Microsoft has canceled Claude Code licenses for employees
Microsoft is discontinuing internal Claude Code licenses within its Experiences + Devices division, which covers products like Windows and Teams. The cutoff date is June 30, 2026, coinciding with the end of Microsoft’s fiscal year.
The pilot began in December 2025, and engineers quickly preferred Claude Code over Microsoft’s own GitHub Copilot CLI. Now, after just six months, the company is winding down operations and reverting all employees back to Copilot. What initially seemed like a straightforward license suddenly spiraled into uncontrollable costs.
Even though Microsoft had invested in Anthropic, maintaining these licenses proved unsustainable for its engineers’ utilization.
Meta has shifted from “tokenmaxxing” to “tokenminimizing”
“Tokenmaxxing” has emerged as a trendy Silicon Valley term, essentially describing the act of burning tokens to climb leaderboards and showcase “innovation,” with actual output being optional.
For two years, Meta had encouraged employees to integrate AI into their workflows, creating internal leaderboards that rated token usage. The message was clear: adapt or get left behind.
However, the internal memo’s tone has changed drastically. In June, Meta’s leadership unveiled plans to curb costs by restricting token usage and developing a dashboard to track usage intricately. Dubbed “token minimization,” they expect internal AI usage to soon cost billions.
Meanwhile, Meta has raised its 2026 capital expenditure forecast significantly, mostly to cover AI infrastructure. However, this has been accompanied by mass layoffs, signaling that even with increased spending, many workers are feeling the pinch.
Amazon has ended its AI leaderboards
Amazon once featured an internal leaderboard called Kilo Rank, where employees were scored on their AI usage. The aim was to encourage adoption and reward top users.
However, this gamified approach quickly led to “tokenmaxxing.” Employees assigned trivial tasks to AI just to climb the leaderboard, sometimes leading to inefficiency. As a result, Amazon shut down Kilo Rank in late May, with Senior Vice President Dave Treadwell reminding the team: “Don’t use AI just for the sake of using AI.”
In essence, Amazon wanted genuine adoption, but what transpired was more about performance metrics than productivity.
The token max hangover
Shockwaves from rising costs are echoing throughout the industry. Several employees at Priceline reported that certain updates began to cost four to five times more than anticipated. Anxiety over financial stability looms large.
By April 2026, many companies are already nearing their budget caps. Comments from industry insiders indicated that the previously unquestioned embrace of AI is changing; now, companies are raising the question of what tangible benefits they’re seeing.
Implications of this shift
I posited previously that AI firms were operating under loss-making subscriptions, consuming investor funds in a bid for market share, while masking actual costs. The products appear subsidized, but a price hike seems inevitable once users are dependent.
What caught me off guard was just how swiftly these subsidies began to dwindle. The deepest-pocketed companies—those that make and sell AI solutions—have come to similar crossroads. They are realizing they can’t sustain these offerings long-term. Uber’s budget is spent, Meta is limiting access, and even Microsoft’s backer is withdrawing licenses.
These are not struggling startups; they’re immensely wealthy corporations grappling with shared challenges.
As evidenced in a recent Economist article, companies are now racing to rein in skyrocketing AI costs. One tech executive described the tightening environment as “an absolute nightmare,” highlighting how the multitude of software programs leads to skyrocketing costs proportional to AI use. Corporate expenses linked to AI have surged dramatically—13 times within just a year. Currently, the average monthly bill for the top 1% of users is around $7,450 each, while typical customers average about $11 per month. It’s surprising to hear even Sam Altman express concerns over rising costs.
Presently, many companies find AI expenses outweighing revenue. Even those creating AI technology face difficulties maintaining viability.
Conclusions
While AI has practical applications, I rely on it regularly. Yet the economic rules are unyielding; higher costs emerge every time power users become adept at maximizing token use. This isn’t speculation—it’s precisely the underlying rationale behind why AI companies exist.
If you manage a business dependent on external models, it’s wise to scrutinize those providers. If those constructing these models struggle with their costs, what might that mean for you in the long run?
There are alternatives, and they are becoming increasingly affordable. For instance, mid-range models like Anthropic’s Sonnet are approximately 5% the cost of their flagship. As new companies enter the market with efficient models, outsourcing easy tasks no longer feels like a compromise but rather an intelligent choice. You shouldn’t feel tethered to vendors facing their own constraints.
Bubbles don’t just burst from headlines; they lose air when memos change. Budget cuts, canceled licenses—decisions to limit access to tools promised to revolutionize industries.
And those who marketed this transformation? They seem to be quietly stepping back.
