Some rush to the imperial examination under starlight and night, while others resign from office and return to their hometowns.Author: Zhang Yongyi
Source: GeekPark
For a long time, Microsoft employees maintained an unofficial, self-managed spreadsheet to exchange salary and bonus figures—a common underground “pay transparency” initiative in large corporations.
This year, a new column appeared on the table, later obtained by Business Insider: "Monthly AI Spending."
Approximately 350 U.S. employees filled in this column. The highest entry reads: 28 days, $28,000 (approximately RMB 190,000).
Microsoft's response was not to award the student. The company has begun tightening token usage management: department-level budgets are now in effect, individual token expenditures are visible on dashboards, and the default internal model has been switched to OpenAI's GPT-5.6 Sol.
Microsoft has begun an internal audit.
Token usage tiers
The most significant information in this table isn't in the numbers, but in the position—the column labeled "AI Spending" appearing in a salary disclosure table.
Employees list their AI usage next to their salaries, making it crystal clear: at Microsoft in 2026, how many tokens you consume each month is just as noteworthy as how much you earn.
These figures are not estimates made on a whim. An easily overlooked detail in the report is that Microsoft provided employees with an internal tool to check how much they spent on AI over the past 28 days. The current median voluntary self-reported amount across the entire company is $300 per month, approximately 2,000 yuan.
To be honest, this price isn’t really surprising—after all, a Silicon Valley software engineer’s total labor cost starts at $20,000 to $30,000 per month, so a $300 AI bill is a drop in the bucket.
But the real cost of the bill lies ahead. According to BI’s departmental breakdown, Microsoft’s Core AI engineering team had a median spend of $975 (approximately ¥6,577)—more than three times the company average; the Microsoft AI team spent around $490, Experiences and Devices around $250, and Azure around $241. Several departments saw individual spending exceed $10,000, with Core AI reaching as high as $16,000. The record holder—$28,000 over 28 days—came from the Customer and Partner Solutions department.
Although 350 people represent only 0.16% of the 223,000 global employees, and the data is voluntary rather than a full census, those willing to disclose their AI bills are inherently not a random sample. Yet the true distribution is likely even more extreme, not more moderate, than what this table shows.
Burning $28,000 over 28 days—what does that mean? That’s an average of $1,000 per day. Three months ago, GeekPark published an article titled “Microsoft Hits the Pause Button on Vibe Coding,” in which they calculated a hypothetical cost: an engineer earning $300,000 per year amounts to over $800 in daily costs. At the time, the article’s title was a declarative statement—“Burning tokens is already more expensive than hiring an employee.” Now, it’s become just one row in a table: the amount this employee has burned is more than the cost of hiring another engineer sitting beside them.
In fact, the Silicon Valley community has already coined a term for this behavior: tokenmaxxing, which refers to maximizing token usage.
Specific practices include deliberately sending massive prompts, filling up the context window, and running automated queries—not to solve problems, but to make the usage metrics on internal dashboards look better.
Clearly, a behavior is only named when it is common enough. The very existence of this term is evidence.
Over the past year, this token frenzy sent the same message to employees at nearly every tech company: embracing AI is a sign of advancement, and those who use AI will replace those who don’t. When bosses constantly emphasize AI usage—and usage is measured by a visible metric—it became the easiest way to manage up. Debating the quality of output takes hours, but the volume of usage is instantly clear.
When usage becomes a visible number, token burning inevitably slides into a new performance of overtime work.
This scenario has played out more than once this year: Uber’s CTO admitted to The Information in an interview that the company implemented an internal leaderboard to incentivize employees to use AI more, and as a result, the entire year’s AI programming budget was exhausted within four months.
Meta’s version was even more extreme: In April, an employee created a dashboard called Claudeonomics, ranking 85,000 colleagues by token usage; within 30 days, the entire company burned 60 trillion tokens, with the top individual using 281 billion tokens—Fortune estimated that one person alone was worth $1.4 million. The dashboard was taken down two days after The Information exposed it, and Zuckerberg himself didn’t even make the top 250. Android Headlines reported that a similar trend had also emerged internally at Amazon.
The difference is that Uber’s leaderboard was set top-down by the company, while Meta and Microsoft’s “leaderboards” emerged organically from employees themselves: no orders were given; the incentive structures blossomed on their own.
Microsoft executives clearly understood the full context of this event. In an internal memo in August, Jay Parikh, Executive Vice President and head of CoreAI, bluntly stated: “Tokenmaxxing is not the real goal we should be pursuing. I want everyone to focus on outcomes that truly create value for our customers and business.”
The unstated implication in the memo is clear: the company has realized that a significant portion of the tokens burned by employees did not buy productivity, but rather, symbolism.
BI’s report also starkly contrasted the salary data from the same table: so far, those who have adopted AI more heavily have not received higher raises, bonuses, or promotions. The rewards of performance haven’t materialized—the bill has already come due.
On the very day Microsoft began revoking Claude Code licenses, someone on social media posted another, much larger bill.
On May 15, 2026, OpenClaw developer Peter Steinberger posted a tweet with a casual tone, ostensibly promoting his menu bar widget CodexBar: the new version displays API costs much more beautifully. What truly stunned viewers was the accompanying image: $1,305,088.81 in OpenAI API usage over 30 days, 603 billion tokens, 7.6 million requests, powered primarily by the GPT-5.5 model—with nearly $20,000 spent in a single day on the day of the tweet.
Steinberger joined OpenAI in February, and this bill is clearly covered by OpenAI. The money is being spent on approximately 100 parallel instances of Codex coding agents, while only three people maintain the open-source OpenClaw project. He defines this “extravagance” as testing a question: If tokens were free, how would we write software in the future?
$1.3 million, roughly 46 times the amount held by Microsoft’s record-holder. But the two bills are fundamentally different: one represents 100 agents working on behalf of three people, while the other involves humans inflating usage metrics to meet performance targets. Both are displays, yet one showcases productivity, and the other, performance posture. In the same May, some raced through the imperial examinations under cover of night, while others resigned their posts and returned home.
From cutting tools to checking bills
Extend the timeline slightly: Microsoft changed its stance on AI costs three times in a single year—among major internet companies, this is among the fastest shifts in posture.
In December 2025, it made Claude Code available to thousands of employees, encouraging everyone to use AI to reshape their workflows.
In May 2026, it revoked Claude Code licenses for most employees, citing "toolchain unification," but retained a workaround: Claude models could still be accessed via the Copilot CLI.
By August, the tiered system was also removed. CNBC reported that the default model for GitHub Copilot for employees has been switched to OpenAI’s GPT-5.6 Sol. According to GitHub’s official statement, Sol is not a cost-saving model but the highest-reasoning-tier model in the GPT-5.6 family; external media interpretations suggest this move saves not on quality but on where the bills go—the report states it is more cost-effective than Claude’s family, and Android Headlines noted that Microsoft no longer encourages employees to use third-party tools like Claude for programming tasks.
First cut the product, then switch the model, and finally implement budgeting and dashboards. The level of control evolves from "what tools are you using" all the way down to "how much did you spend this month."
The timing itself is information: the revocation of licenses in May was deadline-driven to June 30—the end of Microsoft’s fiscal year; the implementation of budgeting and monitoring occurred in the first two months of the new fiscal year. Taken together, it reads as: the final action of the previous fiscal year was eliminating an uncontrolled cost item; the first action of the new fiscal year was establishing a formal budget line for AI.
AI spending at Microsoft completed, in one year, the journey from "experiment" to "line item." The cloud computing industry has already walked this path: first encouraging all teams to migrate to the cloud, then facing ballooning bills, and finally creating a dedicated role called cloud cost governance. Now it’s token’s turn.
Three months ago, when Microsoft paused Vibe Coding, Geek Park concluded that the real reason for the setback was not that AI was too expensive, but that organizations hadn’t adapted—and most companies won’t change anytime soon.
Microsoft has now provided its official response. It did not change the organization—it changed the budgeting system: updating internal guidelines, setting departmental budgets, and tracking individual spending on dashboards. In the words of Parikh’s memo, “Manage token spending with the same discipline applied to all other critical resources.”
Sounds flawless. Tokens are resources, and resources should naturally have a budget.
But this solution contains a circular logic: the internal tool that makes usage visible was originally released by Microsoft itself; employees inflated the numbers, and Microsoft’s remedy was to display everyone’s spending on another dashboard. The previous metric encouraged high consumption, and the next metric encourages frugality. Even though the metric changed direction, employees’ behavior toward metrics won’t change. Six months later, I wouldn’t be surprised to see a new column on this table: “token output per unit.”
The significance of this for domestic internet companies may involve some "pioneering" trial and error: many Chinese companies are still at Microsoft’s “December 2025” stage—allocating quotas, setting benchmarks, and incorporating AI usage rates into their OKRs.
Microsoft simply jumped ahead to the second half of the story: first encourage, then rank, then audit. The time lag is the trailer.
What matters isn't AI's price, but people's reaction to "visible numbers."
As long as there is still one column of numbers visible, someone will always be responsible for pushing it up.