⌘K
Change language Switch ThemeSign In
Narrow Mode
黄仁勋发 Token 当工资！硅谷兴起刷量大赛，一人烧掉 33 个维基百科
量 量子位 @梦晨
One Sentence Summary
Silicon Valley is witnessing a 'Tokenmaxxing' craze centered on token consumption. Tokens are evolving from a technical term into a new form of 'digital currency' for measuring employee performance and serving as compensation.
Summary
This article reports on the emerging 'Tokenmaxxing' phenomenon in Silicon Valley tech companies. With the proliferation of Coding Agents (such as Claude Code and Codex), individual token consumption has grown exponentially, with some cases exceeding $150,000 in monthly bills. Companies like OpenAI, Meta, and Shopify have incorporated token usage into leaderboards or performance reviews. NVIDIA CEO Jensen Huang has even proposed including token budgets as part of engineers' compensation, making it the fourth form of pay after salary, bonuses, and stock options. This trend reflects the scarcity of AI compute as a core productive force, while also sparking industry concerns regarding output quality and financial sustainability.
Main Points
* 1. Token consumption is becoming a new metric for productivity in Silicon Valley.AI companies track employee token usage via internal leaderboards and even incorporate it into performance reviews, reflecting how AI-assisted development has been deeply integrated into core workflows. * 2. Coding Agents are the primary drivers of the token consumption explosion.Compared to manual human input, tools like Claude Code can automatically review codebases and spawn sub-tasks 24/7, leading to an exponential leap in individual token demand. * 3. Tokens are evolving into a new form of compensation and benefit.Industry leaders like Jensen Huang have proposed including token budgets as part of compensation packages, and engineers are increasingly prioritizing 'dedicated inference compute quotas' during interviews. * 4. 'Tokenmaxxing' has sparked concerns about output quality and costs.Blindly chasing consumption can lead to resource waste and the illusion of being 'busy,' prompting corporate finance departments to re-evaluate the ratio of compute costs to human output.
Metadata
AI Score
87
Website qbitai.com
Published At Today
Length 1766 words (about 8 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
> 梦晨 发自 凹非寺 > > 量子位 | 公众号 QbitAI
OpenAI最烧Token的人有多狠？
一位匿名员工，上周处理了2100亿Tokens，是全公司之最，足够把整个维基百科填满33遍。
他不是在做什么惊天大项目，就只是公司内部排行榜第一名而已。
与此同时，在隔壁Anthropic一位Claude Code用户单月账单超过15万美元，折合人民币过百万元。
这便是硅谷新风尚Tokenmaxxing，直译过来就是“Token刷量大赛”。
具体来说：
* AI公司内部开始出现排行榜，追踪每个员工的token消耗量； * 招聘时，”你能给我多少token预算“正在成为工程师最关心的问题之一； * Meta和Shopify甚至把AI使用量写进了绩效考核标准。 Token这个AI处理的最小文本单位，正在从技术术语变成硅谷的新型货币。
风投机构Theory Ventures创始人Tomasz Tunguz亲身经历了Token账单的指数级膨胀。
六个月前，他每月在Claude上花200美元。然后加了三个agent订阅Codex、Gemini和Claude Code，月费涨到600美元。
接着他开始用AI把待办清单自动变成完成清单，每天处理31项任务，日均推理账单飙到92美元。再加上每月400美元的智能体浏览器。 半年之内，他的AI推理支出从年化7200美元涨到4.3万美元，再到超过10万美元。
但在一年前，一个人想用掉这么多token几乎不可能。
假设一个学生写篇论文，来回修改几轮，大概消耗1万个token，约等于7500个英文单词。
要烧掉几十亿个Token，得在电脑前不停下指令好几十小时。 Coding Agent改变了一切。
Claude Code、Codex这类工具可以在无人监督的状态下连续工作数小时，审查和编辑大型代码库，从一条指令生成完整程序。每个agent还能派生出子agent处理不同子任务，每一步都在生成成千上万个Token。
龙虾OpenClaw更是24/7不停工作。
Token消耗的爆炸直接推高了AI公司的收入。
Anthropic今年在两个月内将收入预期翻了一倍多，Claude Code年化收入达到25亿美元。
OpenAI的Codex周活跃用户超过200万，年初以来增长两倍，Token使用量增长五倍。
Google去年透露，其AI模型每月处理超过1.3万万亿（quadrillion）个Token。 不过，这场增长背后有个关键推手：补贴。
OpenAI和Anthropic都在200美元/月的订阅计划里提供了价值约1000美元的Token额度。 和当年打车、外卖用发优惠券抢市场的逻辑一模一样。
英伟达GTC 2026上，黄仁勋把这股暗流推到了台面上，抛出了一个让所有人竖起耳朵的提议：
> 工程师年薪几十万美元，我会在基础薪资之上再给他们相当于一半年薪的token，让他们能力放大10倍。当然我愿意这么做。
黄仁勋成了第一个公开谈论“公司Token预算”的重量级CEO。
在他的框架里，Token正在变成继工资、奖金、期权之后的第四种薪酬。
根据薪酬追踪网站Levels.fyi的数据，硅谷75分位软件工程师的年薪是37.5万美元（约262万人民币）。如果再加10万美元的token预算，总包就是47.5万美元，其中21%是token。
OpenAI Codex工程负责人Thibault Sottiaux最近在X上写道，AI算力正变得越来越稀缺、越来越值钱：
> 候选人面试时越来越多问我：我能有多少专属推理算力。
在OpenAI内部，员工已经可以在排行榜上看到同事消耗了多少token。
公司token预算正在成为一种员工福利，就像医疗保险或免费午餐。
另一面，Shopify和Meta已经把AI使用纳入了绩效考核，奖励重度使用的员工，批评不用的。
风险投资人Nikunj Kothari这样描述弥漫硅谷的新情绪Token焦虑。
> 晚饭时的开场白过去是“你在做什么？”现在变成了“你跑了几个agent？”
但质疑声音也在出现。一位匿名OpenAI员工评价同事们的token竞赛：这看起来不可持续。
排行榜不衡量产出质量。那些刷到数十亿token的人，到底在产出有用的东西，还是只是在空转、看起来很忙？
当一家公司为每个员工支付的Token费用接近甚至超过这个人的工资时，财务部门对“人头”的算法就会发生变化：
如果算力在干活，到底需要多少人来协调它？
参考链接：
[1]https://tomtunguz.com/inference-as-compensation/
[2]https://www.nytimes.com/2026/03/20/technology/tokenmaxxing-ai-agents.html
[3]https://www.wsj.com/tech/ai/claude-code-cursor-codex-vibe-coding-52750531
量 量子位 @梦晨
One Sentence Summary
Silicon Valley is witnessing a 'Tokenmaxxing' craze centered on token consumption. Tokens are evolving from a technical term into a new form of 'digital currency' for measuring employee performance and serving as compensation.
Summary
This article reports on the emerging 'Tokenmaxxing' phenomenon in Silicon Valley tech companies. With the proliferation of Coding Agents (such as Claude Code and Codex), individual token consumption has grown exponentially, with some cases exceeding $150,000 in monthly bills. Companies like OpenAI, Meta, and Shopify have incorporated token usage into leaderboards or performance reviews. NVIDIA CEO Jensen Huang has even proposed including token budgets as part of engineers' compensation, making it the fourth form of pay after salary, bonuses, and stock options. This trend reflects the scarcity of AI compute as a core productive force, while also sparking industry concerns regarding output quality and financial sustainability.
Main Points
* 1. Token consumption is becoming a new metric for productivity in Silicon Valley.
AI companies track employee token usage via internal leaderboards and even incorporate it into performance reviews, reflecting how AI-assisted development has been deeply integrated into core workflows.
* 2. Coding Agents are the primary drivers of the token consumption explosion.
Compared to manual human input, tools like Claude Code can automatically review codebases and spawn sub-tasks 24/7, leading to an exponential leap in individual token demand.
* 3. Tokens are evolving into a new form of compensation and benefit.
Industry leaders like Jensen Huang have proposed including token budgets as part of compensation packages, and engineers are increasingly prioritizing 'dedicated inference compute quotas' during interviews.
* 4. 'Tokenmaxxing' has sparked concerns about output quality and costs.
Blindly chasing consumption can lead to resource waste and the illusion of being 'busy,' prompting corporate finance departments to re-evaluate the ratio of compute costs to human output.
Key Quotes
* Tokens, the smallest unit of text processed by AI, are evolving from a technical term into Silicon Valley's new currency. * Engineers make hundreds of thousands of dollars a year; I would give them tokens equivalent to half their salary on top of their base pay to amplify their capabilities by 10x. * Candidates are increasingly asking me during interviews: 'How much dedicated inference compute do I get?' * Leaderboards don't measure the quality of output. Are those racking up billions of tokens actually producing something useful, or are they just spinning their wheels and looking busy?
AI Score
87
Website qbitai.com
Published At Today
Length 1766 words (about 8 min)
Tags
Tokenmaxxing
AI Compute
Engineer Culture
Performance Review
Coding Agent
Related Articles
* #400. Unveiling xAI's Internal Operations: A Single Code Commit Worth $2.5 Million, Musk's 'Physics' Speed and Crazy Leverage * MiniMax M2.1 Emerges as a Leader in Multilingual AI Coding, Achieving SOTA * AGI is Already Here, and This Time We're a Little Scared... * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native * AI Starts to "Take Action", Alibaba's Qwen Leads the World * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * Coding Agents Have a Sweet Spot That Most People Skip * After OpenClaw, I Only Care About the Next 3-6 Months | A Conversation with Wang Wenfeng, Founder of Sheet0 * OpenAI Open-Sources Symphony: A Deep Dive into the Autonomous Coding Agent Scheduling Service * After Topping Open-Source Rankings with its Programming LLM, the Zhipu GLM Team Faced a 3-Hour Questioning Session HomeArticlesPodcastsVideosTweets