After adopting AI, the company seems to be poorer than ever. When AI first emerged, executives all thought it was a golden opportunity to cut staff and reduce costs. They imagined: one AI could replace three people—working nonstop, available on demand, requiring no salary raises or social insurance, and always online. It sounded perfect. But in reality, AI doesn’t slack off, and it doesn’t work overtime—it simply charges more every time it does more. As a result, many companies are now openly saying: “We can’t afford the tokens anymore.” Many people’s first reaction is: “Is it really that bad? Isn’t AI getting cheaper? Didn’t DeepSeek’s arrival prove that large model costs have plummeted?” But many overlook one key point: while models have become cheaper, companies are using them far more intensively. From occasional individual use, to entire teams adopting AI, to dozens of agents running autonomously in the background 24/7—the cost per call may have dropped, but the monthly bill keeps soaring. For example, Uber granted Claude Code access to 5,000 engineers—and within months, they nearly burned through their entire annual AI budget. Microsoft has recently started hitting the brakes, restricting internal access to Claude Code and no longer allowing engineers unlimited usage. In short: the era of “use it however you want” is over. Amazon went even further—they simply removed their internal AI usage leaderboard. Why? Because they realized that when “how much AI you use” becomes a performance metric, employees start gaming the system—flood the system with token-heavy requests just to climb the ranks. It looks like everyone is enthusiastically embracing AI—but in reality, many of these calls generate zero value; they’re just “using AI for the sake of using it.” In one multi-agent experiment by miHoYo, dozens of agents kept calling each other, waiting for responses, and confirming back and forth: you ask me a question, I reply, you confirm again. No one ever closed the loop—and the call chain kept growing longer and longer. By the end of one night, they had burned approximately 2 million RMB in tokens—while generating negligible real value. At this point, many might wonder: what exactly are tokens? How can they drain a company’s budget like this? Think of tokens as electricity in the AI world. When you type a question into a chatbox and get an answer in seconds, it feels free. But behind the scenes in enterprise systems, every input, every output, every model invocation, every tool call by an agent—even AI-to-AI conversations—consume tokens. Crucially, AI’s pricing model is fundamentally different from traditional software. Previously, software costs were fixed: one license costs X dollars; annual budget is predictable from day one. AI is different—it’s usage-based pricing, and usage scales exponentially with business complexity. One employee asking a few questions occasionally? Minimal cost. An entire team using it regularly? Costs start climbing. Add in agents that let AI call other AI systems? The bill can jump from thousands to hundreds of thousands—or even millions—in no time. And over the past two years, society as a whole has been pushing everyone to use AI more: increase AI penetration, boost usage frequency, maximize automation. Some companies even made token consumption part of employee KPIs. In economics, there’s a well-known principle called Goodhart’s Law: when a metric becomes a target, it ceases to be a good metric. Abroad, people have even coined a new term: “Tokenmaxxing”—roughly meaning “pushing token usage to its absolute limit.” Some people make AI rewrite the same code snippet dozens of times. Others have it generate a dozen versions of a single report. Some even break simple tasks into dozens of interconnected agents—just to make the system look smarter. In reality, AI has become nothing more than a fancy facade. Minor overuse can be tolerated—but what truly pushes costs out of control are multi-agent systems. Theoretically, they sound beautiful: one agent plans, another executes, another checks, another summarizes—a digital team collaboration model. But in practice, they resemble a meeting with no chairperson. You ask me; I ask you. You wait for me; I wait for you. One round of confirmation isn’t enough—let’s do another. Everyone is active—but nothing ever gets finished. In most multi-agent systems, 30% to 60% of tokens are wasted on these meaningless loops. In plain terms: vast sums of money aren’t turning into results—they’re vanishing as AIs hold endless meetings with each other. The irony? These agents aren’t slacking off—they’re being overly diligent. They strictly follow protocols: Agent A calls Agent B; Agent B confirms back with Agent A; this cycle repeats until the entire system gets stuck in an infinite loop. It’s like a company holding a meeting with dozens of people—from evening until dawn—everyone speaking passionately, fully engaged—but no one makes a decision. And the meeting charges by the second. Worse still, these “meetings” keep being replicated, split apart, and nested deeper. As scale increases, costs spiral out of control exponentially. Because AI costs aren’t one-time—they compound along every link in the call chain—and they’re nearly impossible to predict. At this point, companies aren’t debating whether AI is useful—they’re calculating something far more urgent: will this thing blow up our budget? That’s why domestic models like DeepSeek and Doubao are suddenly being reconsidered—not out of sentiment—but because of one simple reality: for the same task, they might cost several times less. In short: don’t always use the most expensive model. Give simple tasks to cheap models; reserve the big ones for complex work. Companies are finally realizing: AI isn’t a tool where “the more you use, the better.” It’s a system where “the more you use, the more you burn.” Wall Street has noticed too. Previously, investors judged AI companies by who used the most AI, who grew fastest, who burned the most tokens. Now they care about only one thing: ROI. You burned so many tokens—did you actually make any money back? A harsh reality: improved efficiency doesn’t equal profit. If code is written twice as fast but no additional products are sold—it’s not profit. It’s just spending faster. Even more surreal? This isn’t just happening at one or two companies. One company burned $500 million in a single month on Claude alone—and there were cases where no usage cap was set, causing token consumption to skyrocket uncontrollably. Meta took it even further—they once ran an internal leaderboard called “Claudeonomics,” tracking who used AI the most. The top user consumed 31.2 trillion tokens in one month. That’s equivalent to hiring two senior engineers for an entire year—just for one person’s AI usage in a single month. Meanwhile, CEOs shout “Go all-in on AI!” while finance teams break out in cold sweats. Ultimately, it’s not about avoiding AI—it’s about stopping mindless token burning.People are starting to ask a more practical question: Do these tokens actually get exchanged for real money? #AI #AIAgent @grok