June 17, 2026, (Inside AI) — Tech companies are slamming the brakes on runaway AI token spending after months of aggressive internal mandates backfired. Quarterly bills shocked executives at firms like Uber, while internal leaderboards at Amazon tracked “tokenmaxxing” — a race to use AI for every task regardless of value. Now, the same leaders who pushed for blanket adoption are ordering teams to use AI only for fruitful purposes, but the financial damage is done and the path to sustainable budgeting remains murky.
The Hangover After the AI Binge
Earlier this year, companies cracked the whip to get staff using AI more. Teams were told to integrate it into workflows without clear need. The predictable result: token usage soared, costs exploded, and business outcomes disappointed. As one industry observer noted, it was like a construction firm buying drills nonstop and forcing workers to use them for every task — you end up with Swiss cheese walls and a drained budget.
The pivot is abrupt. Some businesses still buy drills, but big players realize the cost-benefit ratio makes no sense. They now tell staff to use AI only where it generates real value. Yet this course correction is harder than it looks.
The Unpredictable Calculus of Token Costs
Budgeting for AI tokens is not like forecasting server costs. When a worker prompts a model, two unknowns collide: how many tokens the response will contain, and how many attempts are needed to get a useful answer. Output tokens cost roughly 5x more than input tokens, and agentic tools compound the problem by generating prompts on their own.
“You have only the most minimal control over the number of tokens that any model responds with,” explains Stephanie Kirmer, a data scientist and columnist. “For the most part, the number of output tokens is a part of that nondeterministic unknown.” Multiply that uncertainty across thousands of employees and dozens of models, and finance departments face a budgeting nightmare.
Even past usage data offers little guide. Model architectures, problem types, and hidden randomness make costs swing wildly. Companies must set limits, but those caps will inevitably cut off access mid-project — forcing jarring switches between AI-assisted and manual work.
When the AI Spigot Runs Dry
The practical effects are already surfacing. Will teams revert to manual coding in Q3 after months of AI reliance? Will marketing documents be handwritten once thresholds are hit? The disruption of toggling between workflows could erode whatever productivity gains AI provided.
This belt-tightening ripples outward. Hyperscalers like Anthropic and OpenAI — both planning IPOs this year — have pushed startups to embed AI features, betting on usage-based revenue. If enterprise clients slash consumption, that revenue pipeline dries up. With billions owed to investors and no clear path to profit, a slowdown is the last thing they need.
Consumer demand also looks fragile. Apple’s recent WWDC reveal of a privacy-focused Siri powered by Google Gemini — free and on-device — could lure users away from paid chatbots like ChatGPT and Claude. If the quality holds, subscription models face fresh pressure.
The Same Story, Different Angles
Headlines about “companies shocked at AI bills” and “record IPOs” are not separate stories. They are two sides of a single, unsustainable equation. Enterprises cannot fund infinite token budgets, and consumers — squeezed by rising prices and gloomy economic sentiment — won’t fill the gap. Add public backlash against data centers and AI hype, and the hyperscalers’ revenue expectations look increasingly detached from reality.
Kirmer, who has long warned about opaque AI costs, puts it bluntly: “If they do not have unlimited budgets, we have to come back and ask where the billions and billions that OpenAI, Anthropic, and others are expecting to generate in revenues are going to come from.” The answer remains elusive.
