Some people are burning 210 billion Tokens in a week, and some are spending US$150,000 on AI programming tools in a month. Today, the standard used by Silicon Valley to evaluate whether a programmer is good or bad is no longer code quality, project delivery or team contribution, but a crazy competition against an extremely expensive technical indicator, Token consumption.
An OpenAI engineer burned 210 billion tokens (equivalent to the total text of 33 Wikipedias) through the company's AI model in one week, ranking first in the company. At Anthropic, another AI company, one Claude Code user generated more than $150,000 in bills in one month.
This seemingly absurd comparison is not for employees to entertain themselves, but stems from the crazy admiration of the bosses. Since the boss is superstitious that AI can improve efficiency, burning money like crazy has naturally become a "performance" for employees to keep their jobs. In some technology companies, internal rankings will disclose the number of tokens consumed by each employee in real time; within giants such as Meta and Shopify, management even directly forcibly links the use of AI tools with performance evaluation.
The industry nicknames this phenomenon “Tokenmaxxing”.
But is this sky-high bill really worth it? Is what companies spend a lot of money to buy back super code that can subvert the industry, or is it electronic waste that employees let AI generate crazily in order to boost their rankings? When burning money itself becomes proof of diligence, this so-called productivity revolution may just be a self-deceiving bubble.
01 Token: Hard Currency in the AI Era
To understand the nature of this competition, we first need to understand what Token is.
Token is the smallest unit of measurement for information processing by the AI system, which is roughly equivalent to a word fragment. When the user inputs text into the AI model, the system will decompose the text into tokens for processing; when the model generates answers, the workload is also calculated in units of tokens. For text-based AI applications, generating 750 words requires approximately 1,000 Tokens.
Not long ago, even heavy users could consume thousands of Tokens a day, which was the limit. For example, a student uses AI to complete a paper and after several rounds of revisions, approximately 10,000 Tokens (approximately 7,500 words) are used. To consume millions of Tokens, a person needs to sit in front of a computer and type on the keyboard for several hours, while consuming billions of Tokens is almost impossible.
However, with the advent of so-called agent coding tools, the stakes have been raised. The essential difference between this type of tool and traditional AI assistants is their autonomy. These systems can work unsupervised for hours on end, reviewing and editing vast code bases and even writing complete software programs with just a word of prompt.Each agent can generate hundreds or thousands of subagents to handle different parts of the task, and each step of the operation generates thousands of Tokens.Some AI systems, such as the popular open source AI assistant OpenClaw, are even designed to run 24/7, consuming tokens crazily while human users sleep.
Ege Erdil, co-founder of the AI startup Mechanize, revealed: "If you have several agents that are running continuously, a full-time agent can consume 700 million tokens in a week. (I estimate my own weekly token consumption is between 1 billion and 10 billion.) It doesn't take any effort at all."
When Token becomes a universal yardstick for measuring AI workload, it naturally evolves into a scarce resource and even becomes a new bargaining chip for technology companies in talent competition. Nvidia CEO Jensen Huang made it clear in his keynote speech at this year's GPU Technology Conference that he plans to provide an additional token budget equivalent to half of the annual salary of engineers in addition to the annual salary of engineers."This has now become one of Silicon Valley's recruitment methods: How many tokens does my job come with?" Huang Renxun's words pushed tokens from a technical concept to the forefront of business competition.
Thibault Sottiaux, head of engineering at OpenAI’s AI coding service Codex, has observed a similar trend. He said that in job interviews, more and more people are asking how much dedicated inference computing resources they can get. Token is becoming the fourth major component in the competition for technological talents after salary, bonuses and equity.
02Glory and Anxiety on the Ranking List
In some technology companies, Token consumption is no longer just a technical indicator, but has been given a strong social attribute. In AI companies such as Meta and OpenAI, employees compete on internal rankings based on the number of tokens they consume. Ranking lists have become the new performance display board. Generous token budgets are becoming a job perk for programmers, like dental insurance or free lunches.
Shopify is another example. The company confirmed in a statement that token usage is just one of the metrics used to measure performance, and will also examine how AI improves and amplifies work results.The company's CEO, Tobi Lütke, has made the use of AI a basic expectation and tied it directly to performance reviews, with employees who use AI tools heavily being rewarded and those who use them less or not being penalized.
This atmosphere has given rise to a new kind of workplace anxiety. Venture capitalist Nikunj Kothari calls this “Token anxiety” in his Substack article. He observed that the topic of conversation in the technology circle is quietly changing. In the past, when people met, they often asked "What are you building?" but now it has become "How many intelligent agents have you run?"
To gain a spot on the leaderboard, some employees began resorting to extreme tactics. Some programmers have mastered the art of AI multitasking, opening multiple windows simultaneously and unleashing dozens of agents into their projects at once.
What's more, they take advantage of the loopholes in the subscription package to obtain a large amount of Token quota at a cost far lower than the market price. A startup founder revealed that he discovered a vulnerability in an AI tool developed by the design startup Figma, which allowed a $20-per-month account to use the equivalent of $70,000 worth of Claude Tokens. He exploited the vulnerability to build six software projects simultaneously until the feature began enforcing AI quota limits in recent days.
This phenomenon has also triggered reflection within AI companies. An anonymous OpenAI employee said this seems unsustainable.
However, no one is willing to stop easily, because no one wants to be the person who still mainly relies on manual coding in the AI era. As technology newsletter writer Gergely Orosz notes, within large tech companies, not using AI at an accelerated pace is becoming a career risk, regardless of the quality of the output.
This culture has deviated from the track that engineering should be in the eyes of observers. Some netizens commented that real engineering pursues efficiency and uses the least resources to obtain the best results.The current behavior of comparing server bills is simply mistaking money burning for productivity. If companies pay for hasty experiments without restraint, they will eventually pay the price.
03 Token consumption: Is it measuring effort or achievement?
The employees try to use it, and the boss is happy to pay for it. It seems like a win-win situation. But there is a classic management problem hidden here,Does high token consumption necessarily mean high output?
Automation platform maker Zapier has begun tracking employee token usage with a new dashboard. Its chief AI transformation officer Brandon Sammut said that if they find that an employee uses five times more tokens than their peers, they will be curious, is this person extremely inefficient, or is he a real superstar? The answer depends on what exactly these Tokens are exchanged for.
Vercel provides a positive example. A senior engineer asked a group of AI agents to build a new set of core infrastructure services based on a research paper within a week. If left to human engineers, the task would take weeks or even months. The bill for the job was around $10,000. Vercel CEO Guillermo Rauch thinks the investment is well worth it, saying it's a bit like giving people a fire hose that sprays fuel. He estimated that spending $10,000 for one day of work could save the company millions of dollars.
Kumo AI Company sees the value of Token investment from another angle. Co-founder Hema Raghavan revealed that her excellent engineers using AI agents are like having an army of junior assistants. While some engineers were skiing on the weekends, their agents continued to perform tasks. What's more, Raghavan found that agents sometimes helped write better code, which in turn reduced the company's overall cloud costs.
However, not all high consumption yields such positive returns. Some employees may perform a large number of inefficient, repetitive or even meaningless calculations just to appear on the ranking list, or out of the simple logic that what they use is correct.Some netizens commented that this is like only looking at how many phone calls a salesperson makes, but not caring about how many deals he ultimately completes. If companies do not differentiate between effort and results, incentives will lead to superficial busyness rather than real contributions.
Some engineers shared a smarter strategy of refining natural language prompts into key variables instead of passing complete paragraphs intact to the model. In this way, Token usage can be reduced by about 99%, while the effect can still be retained by nearly 90%. This commenter bluntly said: "Tokenmaxxing is burning money to pretend to be productive, and semantic efficiency is the real unlocking method."
After interviewing a number of heavy users, Kevin Roose of the New York Times raised a more serious question. Rankings do not measure the quality of output, which raises an obvious question:Has any of these token maximizers produced anything good? Or are they just spinning in circles, churning out useless code and wasting precious processing power just to look busy?
04 Cost bill: Who will pay for “digital presenteeism”?
Currently, most of the Token consumption costs are borne by enterprises. But as usage surges, sooner or later this account will have to be reckoned with.
AI companies themselves are already benefiting from this wave. Anthropic more than doubled its revenue forecast in the first two months of the year, largely due to rapid growth in its agent coding tools. OpenAI’s Codex tool has seen weekly active users triple since the beginning of the year, and overall usage measured in tokens has grown fivefold. Google said last year that its AI models process more than 1.3 quadrillion tokens per month.
But for businesses using these tools, the soaring cost of tokens is a reality that cannot be ignored. Some netizens pointed out that the actual resource consumption of current AI subscription services on the market, such as Claude’s $200 monthly package, far exceeds the pricing, and is behind the high subsidies from AI companies.Once AI companies increase prices to achieve profitability, or companies shift to a pay-per-use API model, Token costs may increase several times or even more. By then, those large token owners who are famous in the rankings are likely to quickly become a cost black hole in the eyes of their bosses.
This situation happened at the beginning of the popularity of cloud computing. Many enterprises have paid a heavy price due to out-of-control cloud costs, such as idle resources, over-provisioning, and lack of governance, which ultimately led to bills far exceeding expectations.
Today, token cost is becoming a new cloud cost issue. Exceeds AI founder Mark Hull said that he recently used Claude Code to develop three workflow tools, totaling about 300,000 lines of code, and the token cost was about US$2,000. He decided to make the platform available to his entire company, but within 48 hours costs soared, forcing him to set usage limits.
Vercel CEO Guillermo Rauch also admitted that although the employees currently consuming the most tokens are also the best performers, he did not deny that abuses will occur in the future. He bluntly said that employees may use these tokens for side projects, such as their own startups, part-time jobs to make extra money, or anything else. There will definitely be a lot of abuse.
Some commentators call this phenomenon a replica of digital presenteeism in the AI era. In the past, some people hung their coats on chairs to pretend to be at work, and some used physical mouse jigglers to keep instant messaging tools online while working from home. Nowadays, token consumption is used to prove their value.Only this time, the cost is no longer free acting, but real money. Once companies start to strictly calculate the input-output ratio, these performances will become particularly dazzling.
FinOps expert Kevin Prokopetz pointed out that unmanaged adoption of AI tools will lead to a large number of tokens being burned without any visibility into the actual return on investment. Another commenter, Nate Patel, put it more bluntly: “If token consumption is not tied to deliverables or time saved, it’s just burning money.”
05Back to the basics: What should be measured?
The problem returns to the origin of management. Since Peter Drucker first systematically discussed the productivity of knowledge workers, how to effectively measure output has been puzzling various organizations.There is always a tendency to measure the metric that is easiest to calculate rather than the metric that is most valuable. In the past, it was the number of lines of code, the number of emails sent, and the number of working hours. Now it is the Token consumption.
Some commentators cited Goodhart's law to analyze this phenomenon. When an indicator itself becomes the target, it is no longer a good indicator.This is exactly the case with the Token ranking list, which drives employees to compare consumption rather than pursue real results.
Brian Jabarian, a researcher at the University of Chicago Booth School of Business, said that companies must start measuring token usage, not to compare who uses more, but to see the input and output.
He believes that everyone thinks that productivity will increase as long as using AI Token, and then it will be over, but the reality is much more complicated. If a company saves upfront costs through AI recruitment, but needs to spend more manpower or tokens to make up for mistakes later, then the overall loss will be a loss. When a company provides AI tools to 500,000 employees, these token issues become paramount.
Some companies have begun to explore more refined management methods. Zapier's Sammut said they use analysis to draw conclusions about whether a usage pattern is a gold medal worth promoting among colleagues or a negative pattern that needs coaching to break away from. Exceeds AI’s Hull suggested that companies should develop governance rules around token usage, such as setting limits on which models can be used for specific tasks, and even leverage AI itself to automate such selections.
A netizen shared his evaluation criteria. To evaluate the value of an AI project, it is not how many tokens are consumed, but how much lasting value each token creates.If any manager insists on using Token consumption as the only assessment criterion, he will choose to leave without hesitation.
Perhaps what we need is not more Tokens, but smarter ways of using them. Many companies use state-of-the-art top models for all scenarios, while in many workflows lower cost models are sufficient. Lack of context caching and poor context management are also important reasons why enterprises waste tokens.
Others have begun to explore making AI more frugal and autonomous, such as using smaller-scale models to complete specific tasks. IBM's Granite 4 series of models, with its 3B and 350M parameter versions, run at a fraction of the cost of larger models and can even run on low-power devices such as the Raspberry Pi.
Another commentator put forward more fundamental thinking from the perspective of technical architecture. The real breakthrough, he believes, lies in replacing parameter inflation with structural efficiency.Using brute force that consumes tens of billions of tokens to solve deterministic logic problems is like driving a jet engine to a bar. This is a huge and unreasonable waste of energy.The future of agent AI should be frugal, autonomous, and deterministic, offloading rigid logic to efficient solvers, rather than uncontrollably generating hundreds of sub-agents in a software sandbox.
These views point in the same direction: Token itself is not an end, but a means. Measuring Token consumption is ultimately to measure what it is exchanged for.
Conclusion
When this evil trend of "pretending to work hard" by burning computing power spreads throughout Silicon Valley, managers will eventually face an extremely cold business reality: There are no real superheroes in this crazy involution game. The only absolute winners right now are the computing power suppliers hiding behind the scenes.
It is undeniable that today's "Token addicts" who are consuming their quotas like crazy may really be able to explore the potential of the tool in the future and evolve into the legendary hundred-fold efficiency engineers. But it could just as well be just an expensive piece of workplace performance art. Once companies re-examine the true definition of "effective output", this bubble blown by computing power will collapse at any time.
No matter how this absurd drama ends, one thing is certain:The future world is destined to require larger data centers. And when the big waves wash away the sand, the ones who can really win in the competition will not be those "ranking machines" who are obsessed with ranking numbers, but those organizations and people who know how to convert every Token into actual business value.
But before imagining the ultimate triumph of this productivity revolution, it’s best to pray that your company’s finance chiefs are still emotionally stable when they see next month’s bills.