Two years of open-ended trials with generative AI tools have delivered plenty of excitement inside corporate offices. Now the invoices are landing. And they are bigger than expected.
Uber burned through its full 2026 artificial-intelligence budget in just four months. The ride-hailing giant had opened access to Anthropic’s Claude Code for 5,000 engineers late last year. By early 2026, AI generated 70% of committed code. Usage hit 84% to 95%. The company’s chief operating officer, Andrew Macdonald, later admitted the connection between that spending and tangible benefits for consumers “is not there yet.”
Executives across industries echo similar frustrations. One consultant told Axios of a client that spent $500 million in a single month on uncapped Claude licenses. The bill arrived without warning. Stories like these have multiplied in recent weeks. They signal a clear turn. The era of unrestricted AI trials is giving way to tighter controls and harder questions about value.
From freewheeling pilots to measured budgets
Amazon Web Services senior vice president Peter DeSantis draws a parallel to the early cloud days. “Just like every technology, when we first launched the cloud, some of our most successful customers were delighted by the agility… but many of them woke up one day, and they’re like: ‘Wow, we’re spending a bunch of money.’” He sees the same pattern now with AI. Companies must learn to budget for it and use it efficiently. (Fortune)
Schneider Electric’s chief AI officer, Philippe Rambach, puts it plainly. His teams choose models carefully. “You don’t always need to use the latest frontier model. Quite often you can use relatively cheap models.” Cost awareness now shapes every business case. “The question of the cost of AI is becoming more and more important. We need to have that under control. We need to measure it. We need to include that in our business case, business plans, and decisions.”
Visibility remains poor. A KPMG survey found only one in four companies possess a full picture of their AI expenses. Half have partial sight. One in five see nothing until the bill hits. “It’s a new resource that needs to be managed that didn’t exist quite that way, and we’re seeing exponential growth,” said Steve Chase, KPMG’s global head of AI. (Fast Company)
Sam Ransbotham, professor of analytics at Boston College, highlights the structural mismatch. Users and bill-payers often sit in different departments. “They turn on usage, and suddenly the people paying the bill are not the people using the product, and whenever you have that sort of mismatch, there’s going to be problems.” Token pricing adds opacity. A token represents a fragment of text or code. Some cache. Others process fresh. The math rarely reveals itself until month-end.
Corporate America has begun to ration. The Wall Street Journal reported executives pushing workers toward cheaper internal tools, sharpening skills for better returns, and imposing limits. Meta, Microsoft, and Salesforce have urged more productive use or outright caps on certain models. Amazon killed an internal leaderboard that rewarded raw token volume; employees gamed it with pointless queries. Similar experiments at other firms produced the same waste.
Microsoft canceled most Claude Code licenses in its Experiences and Devices division, which builds Windows and Microsoft 365. Per-engineer costs ran $500 to $2,000 monthly. The projected annual tab approached nine figures with little evidence of faster output. Google retired its Mariner browser agent after it proved slow, brittle against cookie banners and CAPTCHAs, and expensive per task. Starbucks abandoned NomadGo, an AI inventory system using cameras and LIDAR. It mislabeled items. Baristas corrected more than they saved. ROI turned negative after nine months.
Klarna scaled back after AI chatbots handled 75% of customer contacts in 2024, equivalent to 700 agents. Simple tickets closed fast. Complex ones escalated more often. Customer satisfaction slipped. The firm quietly brought humans back in early 2026. Walmart imposed per-employee caps on its Code Puppy tool in June after demand exploded. Duolingo dropped AI from performance reviews in April when engineers questioned whether they used it for substance or optics. Meta itself shut down an internal Claude token leaderboard after one user burned 281 billion tokens—roughly $1.4 million—in 30 days. The system had rewarded volume over results.
These pullbacks come as hyperscalers pour unprecedented sums into infrastructure. Goldman Sachs analysts now forecast $527 billion in capital spending by major AI players for 2026, up from earlier estimates. Some projections reach $700 billion. Power has become the binding constraint. Data centers could consume 1,000 terawatt-hours globally this year, rivaling Japan’s total electricity use. In the United States, demand may climb 130% by 2030. Hyperscalers compete for grid capacity, long-term power contracts, and even nuclear restarts. (Goldman Sachs)
Vendors have shifted pricing too. The “all you can eat” phase fades. Usage-based token charges let providers recover massive compute and energy bills. Startups that built wrappers around frontier models now face their own squeeze. Many subsidize free users while paying full freight to OpenAI or Anthropic. One analysis warns 99% of such AI startups could fail by year-end as margins collapse. (Medium)
Yet adoption keeps rising. Ramp data shows 50.6% of U.S. companies now pay for at least one AI subscription, up from 46.8% at the start of the year. Global AI app spending heads toward $4.2 billion in the first half of 2026. ChatGPT’s market share slipped below 50% while Gemini and Claude gained ground. Still, many executives admit they cannot yet tie those subscriptions to clear profit gains.
Builders of the technology sense the change. Even inside AI labs and cloud providers, conversations have turned from experimentation to measurement. Cheaper, specialized models gain favor for routine tasks. Frontier models stay reserved for work only they can handle. Rambach’s advice at Schneider Electric captures the new discipline. Match the tool to the job. Track every dollar. Build the business case before scaling.
So the free-for-all ends not with a ban but with budgets, dashboards, and accountability. Companies that master this transition will keep investing. Those that treat tokens like free air risk nasty surprises. The bills have arrived. The reckoning has begun.
WebProNews is an iEntry Publication