The giants have started to settle the bills for AI, and the expenses incurred are not proportional to the value created. The good days of Anthropic are gone forever, and the big companies have collectively begun to restrict the use of internal AI Meta is building a central gateway called AI Gateway to monitor in real-time how many tokens each employee has burned and how much money they have spent, setting budgets and limits. Meta predicts that the internal use of AI will generate billions of dollars in costs by 2026 alone. The magical thing is that a few months ago, he had a different face. In November 2025, Meta was still notifying employees to make "AI driven influence" the core assessment standard for 2026, and to directly link performance bonuses to AI usage rates. As a result, it can be imagined that employees began to crazily increase traffic, creating a "tokenmaxxing" frenzy, and even creating an internal ranking list called "Claudeomics" to publicly compare who burned more. Before the ranking list was shut down, the total amount of tokens consumed by employees within 30 days skyrocketed to 73.7 trillion. CTO Bosworth later came out to put out the fire and said a very simple truth, burning tokens does not mean there is output. This sentence is a signal that the entire AI narrative is starting to turn. 1、 From 'encouraging burning money' to 'aligning productivity', the three giants are almost synchronized This is not Meta's business, it's the big companies collectively stepping on the brakes at the same time window. -Amazon: Shuts down employee run token ranking board 'KiroRank', clearly stating that token quantity is to clarify costs and efficiency, but do not use this as a measure of developer productivity -Microsoft: Directly revoked the developer's Claude Code license, which had only been open for a few months before -Meta: Building gateways, setting quotas, and pushing employees to gradually stop using third-party programming tools like Claude and switch to self-developed MetaCode The earliest news was that Uber's annual AI programming budget was burned out by April. ServiceNow has started monitoring daily usage of its employees. Some VC institutions have also started setting daily consumption limits for internal AI Token usage. The tone is surprisingly consistent: previously it was' quick use, versatile, KPI brushing ', now it is' cost aligned productivity'. To put it simply, the honeymoon period has ended and we're starting to settle accounts. 2、 Why now? Because Agentic AI punctured the bill Previously, when using AI, it was just a question and answer process with limited token consumption. Now that Agentic AI is running, a task can be broken down, called, and iterated multiple times on its own, consuming tens of times more tokens than regular usage, and in extreme cases, it can be estimated to be up to 1000 times more. Goldman Sachs predicts that by 2030, agentic AI will drive a 24 fold increase in token consumption. The more employees use it, the more they get involved, and the bills roll up, but the output has not increased year-on-year. The mismatch lies here. The input is exponential, while the output is linear. 3、 Who is paying for this carnival? On the other end of the bill is Anthropic This is the key. The billions of dollars in token fees burned by large factories are largely paid to upstream model vendors. In the current enterprise level programming scene, Claude is the absolute mainstay. In other words, the months when the company "burned tokens without restraint" happened to be the steepest months of Anthropic's revenue curve. Let's take a look at how exaggerated the Anthropic curve is: -December 2024: Annual revenue~1 billion US dollars -By the end of 2025:~9 billion US dollars -February 2026:~14 billion US dollars -By the end of February 2026:~19 billion US dollars -April 2026:~30 billion US dollars -May 2026: Exceeding $47 billion Series H financing was just completed at the end of May, raising $65 billion with a post investment valuation of approximately $965 billion. In just over a year, the annualized income has increased by almost 47 times. This growth rate is among the most outrageous in the entire history of business. But the question is precisely: what is the fuel for this curve? A large part of it is impulsive consumption on the enterprise side, such as the tokenmaxxing ranking list, the amount of performance linked flashing, and the wild volume increase of agentic AI before it is allocated quotas. The quality of this part of the demand is questionable, as it is not entirely driven by real productivity, but rather by KPIs and enthusiasm. Nowadays, big companies are setting limits, revoking licenses, and promoting self-developed alternatives, which is equivalent to actively squeezing out the water from the demand side. 4、 So what? My judgment My opinion is that Anthropic's annualized revenue this year is likely approaching a temporary peak. Firstly, major existing customers are actively reducing their usage. Microsoft's withdrawal of Claude Code, Meta's promotion of MetaCode replacement, Amazon and a bunch of companies setting quotas are all actions that directly cut into Anthropic's revenue plate. The fattest batch of corporate customers are shifting from "open use" to "cost saving". Secondly, the portion of income that has been squeezed out is the 'virtual obesity'. The run rate indicator itself is sensitive, as it is calculated based on the annualized income at a certain point in time. When this point happens to be at the peak of tokenmaxxing, the annualized number will be significantly amplified. Once the quota is implemented, the real water level will be revealed on a month on month basis. Thirdly, whether the new growth can fill the gap is a mystery. To maintain this curve without turning around, Anthropic needs to rely on new scenarios and new customers to fill in the water squeezed out by the enterprise. There is currently no consensus on whether it can be supplemented, how much to supplement, and how quickly to supplement. I'm not saying Anthropic is not good anymore, it's still one of the strongest players in this round, with a very hard base. But the issue of 'annualized income skyrocketing' is likely to encounter the first real resistance level this year. From 47 billion to above, every level will be harder to go than before. Ultimately, the AI business has finally reached the stage where ROI needs to be calculated. The decline of the frenzy of burning tokens is good for enterprises, and costs return to rationality; For upstream model merchants, it is a test that growth supported by impulse consumption must be replaced by growth supported by real productivity. As for this year's annualized peak revenue, I tend not to exceed the figure of $47 billion, as Codex is also catching up quickly on the other side.