Enterprise AI bills are spiking so fast that some companies are burning their entire annual budget in three months, and a few are seeing spending double or triple.
The culprit is agentic AI, which consumes far more tokens than a simple chatbot, and the budgets companies set in late 2025 are already gone.
Corporate America is starting to ration AI because the cost is climbing faster than anyone anticipated. New reporting from Axios and The Wall Street Journal shows AI bills exploding across big companies, with some enterprises hitting their full annual budget in just three months, and others watching their spend double or triple. The pain comes down to one unit of measurement: tokens. These are the basic chunks of text AI bills for. Every word you send a model and every word it sends back is counted in tokens, and tokens cost money.
What changed recently is the rise of agentic AI, meaning AI that takes multiple steps on its own to finish a task instead of just answering one question. Because an agent repeats queries in sequence, it burns far more tokens than a chatbot ever did. As the WSJ's Bradley Olson reported, "Use of artificial intelligence by big companies is exploding, and the soaring cost has some of them pumping the brakes. Some enterprises have hit their annual budget in just three months or reported seeing their AI spending bills double or triple. Now, corporate leaders are scrambling to bring down expenses by finding ways to ration AI use."
The examples are eye-opening. Microsoft reportedly canceled most of its internal Claude Code licenses, partly over cost, six months after rolling them out. Uber's CTO went viral for blowing the company's entire 2026 Claude Code budget in four months, and Uber's COO told Business Insider the costs are harder to justify because the higher usage was not translating into proportionally more useful features.
Axios reported one company spent half a billion dollars in a single month after failing to set usage limits on its Claude licenses. Paul Roetzer, SmarterX and Marketing AI Institute founder and CEO, discussed what the spike means on Episode 217 of The Artificial Intelligence Show.
3 months - Time it took for some enterprisesto burn through their entire annual AI budget
$500 million - One company's AI spend in a single month because no usage limits was set
24x - Projected growth in token consumption between 2026 and 2030, per Goldman Sachs
3.2 quadrillion - Tokens processed by Google alone in May 2026, up 7x year-over-year
4 months - How fast Uber's CTO blew the entire 2026 Claude Code budget
The demand has no ceiling in sight. These bills keep climbing because appetite for AI keeps growing. "The demand for intelligence and agents is seemingly insatiable," says Roetzer. "We're just at the ground floor here, and I think it's only going to increase. This is the top of the first inning." The numbers back him up. On May 5, Goldman Sachs projected that token consumption will multiply 24x to 120 quadrillion tokens per month between 2026 and 2030, noting that "agentic AI requires a lot of tokens because many queries are repeated in sequence. It's like taking a simple chatbot request and blowing it up 10-fold, 20-fold, 50-fold." Google offered a window into the scale: Sundar Pichai shared at Google I/O that the company processed 9.7 trillion tokens a month two years ago, 480 trillion in May 2025, and 3.2 quadrillion in May 2026, a 7x jump in a single year. That is just one company.
The budgets are already obsolete. Most companies set their 2026 AI budgets in the fall of 2025, before agentic tools like Claude Code detonated their assumptions. "If you set an AI budget at fall of 2025 during budgeting season, that was before the explosion in capabilities with Claude Code that has sort of set off this massive, exponential increase in the capabilities and use of AI tools.," says Roetzer. "So those budgets are basically obsolete, and no one I've talked to has any clue how to handle this." He has watched leaders try in real time. "I have actually met with executives at major enterprises who are in charge of AI access and token budgets in their companies, and they are scrambling to solve for how to do this, how to manage it, how to monitor it, how to meter it, how to decide who gets what budgets."
Monitoring usage does not solve it. The instinct is to track every employee's token usage and clamp down on the heavy spenders. But seeing the data does not help, because acting on it means micromanaging each person's AI use every single day, which no manager can sustainably do. Some companies are going the other direction entirely: Meta reportedly incentivized employees to use more tokens, which Roetzer calls absurd as a key performance indicator. The trouble is that usage does not map to value.
As Christopher Penn put it in his piece on saving AI token budgets, "if your organization is proposing some measure of AI adoption, token usage is the absolute worst metric."
His six non-technical fixes are practical:
Do more planning and less execution
Keep things on deck
Use templates
Plan big and act small
Use smaller models
Get better at prompting
The model fits software, not knowledge work. Roetzer's core argument is that per-token pricing doesn't work now. "Metering this stuff by tokens might make sense in the software world. It does not make sense in a broader knowledge worker for marketing and sales and success and HR. It has to be a flat fee."
"I just want unlimited usage. And I don't care what you charge me for it. If a lab just told me $15,000 flat, and it covers all usage for that employee for the year, I'd be like, sold. The value is obvious to me. So this game of playing this credits, it's just unsustainable."
— Paul Roetzer, founder and CEO of SmarterX, Episode 217 of The Artificial Intelligence Show
SmarterX feels this firsthand. The team hits Claude Code limits daily on standard licenses, not even the metered API, and roughly $200 of credits can vanish quickly. Usage is wildly uneven across about 20 employees, ranging from a few cents to one person consuming half the total, and nobody can fully explain why. That unevenness is the problem with token metering in miniature: It punishes the people getting the most value and hands leaders a number that looks precise but says almost nothing about return.
The deeper signal is that the pricing model is colliding with the adoption curve. "It has to be unlimited usage of some sort because there's no way this scales," Roetzer says. The companies generating these jaw-dropping bills are not overusing AI. They are using it the way it is meant to be used, and the metering scheme underneath was never designed for knowledge work, where the value of an output has nothing to do with how many tokens it took to produce.
The introduction of flat-fee, unlimited pricing. The current credit-based system creates anxiety on both sides, with companies afraid to let employees use AI freely and labs leaving obvious value on the table. The moment a major lab offers true unlimited usage for a flat per-seat fee, expect enterprise adoption to jump, because the budgeting nightmare disappears.
Falling model costs might quietly defuse the crisis. The other force at play is the relentless drop in the cost of intelligence. "Twelve months from now, the cost of that model's going to be 10x to 100x less, and we probably don't have this issue," Roetzer says. Smaller, cheaper, good-enough models could absorb much of today's agentic workload. The companies that learn to match the right model to the right task, rather than throwing the most expensive model at everything, will get the economics under control first.
Only 25% of organizations have reached the Scaling phase of AI adoption, according to the 2026 State of AI for Business Report. The largest share, 47%, is still Piloting, running a limited number of narrow projects, and 28% are still just Understanding AI.
How does this relate to the token and budget chaos? This is what happens when companies ramp AI spending faster than they build the operational foundation to scale. Most are paying for experiments, not a scaled capability, so the bills rise while the return stays fuzzy or minimal.