The good news is that everyone doesn’t need to worry too much about losing their jobs because of AI. Because after the boss calculates the accounts, he will find that your price/performance ratio may be higher. In March this year, Huang said, "If an engineer with US$500,000 did not consume at least US$250,000 worth of Tokens every year, I would be deeply disturbed." This brought the magic of this world to a new level.
Companies began to encourage employees to consume as many tokens as possible, and even incorporated token consumption into employee KPIs.
Two months ago, an old man from a major domestic factory broke the news on Xiaohongshu and published the department’s Token consumption list in March. He also said that whether they can become a regular employee during the probationary period, how much performance they can get at the end of the year, and promotions must refer to the Token consumption data.
It’s just as crazy abroad.
Silicon Valley technology companies are promoting a "Tokenmaxxing" (Token maximization) culture internally. Taking Meta as an example, employees built a Claudeonomics dashboard to count the Token consumption of about 85,000 employees in the company. When the data was pulled, more than 60 trillion Tokens were burned throughout the company in 30 days.
Disney, which theoretically has little to do with technology, has also launched an AI Adoption Dashboard on its intranet to track employees' use of AI.
Slowly, this trend has become more and more crooked. Token consumption is even regarded as the threshold of social interaction. If you don't use it enough, you won't be able to enter their circle.
Everyone is in this competition.
It seems that everyone has acquiesced from the beginning that AI is a perfect existence that can reduce costs and increase efficiency, so don’t worry about it, just close your eyes and do it all.
When I saw the bill, I realized that that was not the case...Reducing costs and increasing efficiency has become reducing costs and increasing laughter.
Some time ago, according to Bloomberg news, Uber implemented a new rule: when employees use various intelligent programming tools (Anthropic’s Claude Code or Cursor), the monthly consumption limit for a single person and single tool is US$1,500.
The key is not the amount, but Uber’s initiative to limit it.
You know, in December last year, in order to keep everyone up to date, Uber opened Claude Code to about 5,000 engineers across the company, and also created an internal ranking list to track usage.
The original intention was to get everyone to embrace the trend of the times, but before they could embrace it, Uber’s chief technology officer revealed that the company had burned through the entire Claude Code budget in four months.
Therefore, Uber had to take emergency measures and manually open a wave of gates. Only special business scenarios that have been approved at all levels can exceed the $1,500 limit.
At the same time, Microsoft can't sit still.
They are busy taking back the Claude Code licenses from employees in the E+D department (experience and device department). Before June 30, everyone must switch to the embrace of Microsoft's own son GitHub Copilot CLI.
Although the official statement is for integration, Claude's model can still be used when migrating to GitHub Copilot, but The Verge reported that sources said there are still financial considerations.
Because after June 30, Microsoft will start a new fiscal year.
In addition to Microsoft and Uber, foreign media Axios also broke even more fierce news. One company burned through $500 million in just one month because it didn't cap usage on employees' Claude licenses.
Although there is no specific company pointed at, this amount of Token consumption has led the outside world to directly focus suspicion on the seven sisters of Silicon Valley.
As luck would have it, the day after the Axios report was released, Amazon shut down an internal AI ranking list called "Kirorank". Executives said, "Don't use AI for the sake of using AI."
So it’s hard not to arouse doubts about whether their family burned 500 million US dollars in a month. After all, Amazon was quite aggressive before, requiring more than 80% of developers to use AI every week, which caused the employees below to start all kinds of meaningless operations.
Classic Goodhart's Law, when an indicator becomes a target, it ceases to be a good indicator.
Fortunately, this farce of Token worship did not last long.
As soon as the bill came out, everyone came to their senses and thought of a more essential question:Is it worth the money?
It is undeniable that the company allowed everyone to burn Tokens in the early stage, which also meant experimentation.
After all, no one knows how much value AI can bring. If you can really see the effect, it doesn’t matter if you spend some money on it.
But the reality is often that Tokens flow away like water from a faucet, but no actual business value can be seen, or it is difficult to find a standard to measure this value.
Including Uber COO Andrew Macdonald also said in an interview show that it is difficult to find any connection between “higher token consumption” and “new feature implementation”.
in other words,The consumption of Token cannot be directly equated with the actual output value.
AI reads and understands your needs, then thinks and generates the content you want, all of which consumes Tokens. This means that as long as there is interaction, consumption will occur, but the output may not always be valid content.
After understanding this, it’s a bit strange to look back and treat “Token consumption” as a list.
This is like writing an article in an editorial department. If the word count is an important assessment criterion, then Shichaoda can keep writing anecdotes like this and add more meaningless nonsense literature to make up the word count.
In order to cope with the assessment, employees can completely stop doing real work and ask the AI to run a few useless lengthy codes in different ways every day, or let the AI do some work that may be faster.
When the data was finally pulled, everyone’s Token consumption exploded. It was extremely advanced, but perhaps no substantive business was advanced.
MiHoYo previously worked on a multi-agent collaboration project. For 13 hours, these Agents did nothing serious. They just called each other and chatted enthusiastically, burning 2 million yuan in one night.
And not only at the company level, but also in the circles of developers and ordinary users, it has become a popular trend to show how many tokens you have consumed. It seems that the larger the number, the stronger your ability and the more geeky you are.
But to be honest, Shi Chaoguang saw how many tokens they burned, but he really didn’t see much results.
Previously, Peter Steinberger, the developer of OpenClaw, revealed that the team burned through a bill of 1.3 million US dollars a month, and was also questioned by netizens as not delivering anything.
Although Peter responded that all the consumption was spent on OpenClaw, Shichao thought about it and found that OpenClaw didn't seem to have updated any explosive features...
The current Token consumption is awkward. It can only prove that the big model is working hard, but it cannot prove how much good work you have done with it.
Just like back then, some people questioned that GDP was not objective enough when reflecting the real economic situation. Later, economists slowly explored another set of measurement standards that could be used as a supplement.
Therefore, blindly letting employees use AI without clarifying the connection between Token consumption and output, or without finding an indicator that can accurately quantify the actual output value of AI, is simply giving money to large model manufacturers.
To take a step back, even if it is not an extreme example like MiHoYo, this account cannot be settled.
Because AI cannot completely replace humans at this stage, and Potian also plays an auxiliary role, the true cost composition of an enterprise’s introduction of AI should be “employee wages + AI computing power cost”.
The actual workflow often becomes that after the workers put forward the requirements, the AI generates a bunch of preliminary usable things, and the workers continue to retry and correct errors. During this process, Tokens are kept burning, which may be much more expensive than hiring two interns directly.
At the end of the day, I really can’t tell whether it would be more economical to lay off employees or use AI.
Goldman Sachs predicts that by 2030, global token consumption will increase 24 times compared with 2026, reaching 120 quadrillion per month.
In the past, everyone always thought that AI could replace some highly repetitive low-end jobs, but now from a cost perspective, low-end jobs are safer.
In general, there are now some voices in the industry that are gradually returning to rationality and no longer blindly pursue token consumption.
Major domestic companies like Tencent are also rumored to have begun to limit the amount of tokens their employees can use. After early experiments, everyone gradually realized that the use of Token requires more consideration of actual output.
At the same time, the charging logic of SaaS companies is also changing.
For example, the marketing platform Hubspot began to modify its pricing model in April, from charging by token to charging based on actual results.
Not long ago, Shichao went to Suzhou to participate in an event. Wang Dong, vice president of Kingsoft Office at the scene, made a point that I think is worth thinking about:The implementation of enterprise-level AI requires "double-high scenarios" of high value and difficulty.
To put it bluntly, good steel must be used on the blade.
In the end, this farce of Token worship came and went quickly, but Shichao still felt a little complicated in his heart.
Because Tokens are too expensive, people are joking on the Internet, "If you let the cows and horses work overtime, you don't have to pay overtime, but if you let the AI work overtime, you can't lose a penny of this money."
When one day capitalists discover that hiring humans is more cost-effective than AI, will it be a tragedy for us?