.jpg)
The price of a token fell roughly tenfold in a year. Enterprise spending on generative AI still more than tripled, from $11.5 billion to $37 billion.
That sounds like a contradiction, but in fact it’s one of the oldest principles in economics in action. Make something useful cheap enough and people don't buy a little of it, they buy a lot. William Jevons spotted the pattern with coal in 1865; we're rediscovering it with GPUs, and coining a whole new word for the economic impact: tokenomics.
For many enterprises, AI bills are now big enough to reach the quarterly earnings call. Royal Bank of Canada's token usage jumped 500% in six months. Cisco's CEO says a third of staff use its internal chatbot daily and "the token usage is getting pretty, pretty crazy." Roughly 300 companies fielded questions about AI tokens from analysts in April and May. A year earlier, just 93 of them mentioned the word at all. Global AI spending will hit $2.59 trillion this year, up 47%. When a number climbs that fast, the reflex is to reach for the fire alarm.
Pulling the plug is not a strategy
It wasn’t that long ago that ‘tokenmaxxing’ was all the rage, and leadership encouraged adoption and usage with leaderboards. Times have changed.
Amazon retired an internal leaderboard called Kirorank after engineers started aiming agents at pointless tasks to climb it. Meta took its version down too.
Other businesses have gone further to curb usage: Microsoft reportedly canceled a chunk of its internal Claude Code licenses, and Meta, Uber and Salesforce have all floated usage caps. The instinct is understandable. When you can't explain a number, the quickest way to shrink it is to switch off whatever's producing it.
Turning off the AI to control AI spend is a confession that you can't tell good spend from bad. Most companies can't. Only 7% of leaders say they've cracked AI ROI, and the ones closing in share one trait: real visibility into what they're spending. Leaders who have it are five times likelier to see a return, 15% against 3%.
Most requests don't need your priciest model
Give an identical coding task to a handful of models and some burn over 1.5 million more tokens than others to reach the same answer. Same job, wildly different invoice, decided by which model you happened to pick. Marc Benioff, reportedly staring at a $300 million bill, wants what every operator wants: a smart router that sends only the hardest queries to the priciest model. Executives everywhere are now buying or building systems to do one thing: pick the lowest-priced model that can handle a given prompt.
The companies staying in the black already work this way. 8x8 reckons it's saved around $5 million a year by dropping tools that Claude replaced, and its Claude bill sits well below that. It doesn't ban usage; it puts a dashboard in front of all 1,800 employees so everyone can see the meter. When Opus 4.8 landed (about 1.7 times pricier than an earlier Claude model), its operations chief didn't move to lock people out. He asked whether they could downgrade the model a little and still get the same outcome. When he later noticed one of his own automations was eating tokens, he had Claude rewrite it and cut usage by 80%. Instead of banning behavior, he optimized it.
Govern the meter before you bill for it
This gets more pressing every quarter. Nearly half of large enterprises now say most of their compute goes to inference, up from under a third a year ago. The cost has shifted from a one-off training run to an always-on meter that ticks with every prompt. 98% of FinOps teams now manage AI spend, up from 31% two years ago. And Gartner expects more than 40% of agentic AI projects to be scrapped by 2027, with runaway cost and murky ROI top of the kill list. Most of those projects will probably work, but function isn’t enough to justify unaccountable expense.
The fix is a control plane that watches the meter long before anyone talks about billing. Every token gets attributed to a tenant and a project, so the spend has an owner. Quotas — token budgets, rate limits, access ceilings — stop a runaway agent at a threshold, well before it becomes a five-figure line on next month's invoice. And each request goes to whichever model suits the job and the sensitivity of the data, so you're not paying flagship rates to answer trivia. We built PaletteAI to do that.
None of this shrinks the bill on its own. Global AI spend is heading one way: up. Tokens will keep getting cheaper and paradoxically, the invoices will keep getting bigger. The companies that come out ahead will be those that look at an enormous AI bill and can confidently account for every line of it, and the value it generated.