A recent line from Nvidia CEO Jensen Huang has been making the rounds. “An engineer earning $500,000 a year should be spending at least half of that — $250,000 — on tokens. If you tell me they only spent $5,000, I’m going to explode.” It sounds like a joke. He means it. What exactly are these “tokens” he’s talking about, that produce a number like that?
A token is the smallest unit of text that AI processes. A human reads a sentence in a single breath, but AI chops the text into tiny pieces and reads it that way. In English, one token averages about four characters; a paragraph runs roughly 100 tokens. AI companies bill by token count. The more you use, the more you pay.
And the unit of “more” is beyond imagination. One OpenAI engineer used 210 billion tokens in a single week. That’s the equivalent of feeding all 33 volumes of an encyclopedia into the system from scratch. Another developer paid $150,000 in a single month on just one of Anthropic’s coding tools. Even a one-line requirement, fed in, prompts the AI to build, debug, and rebuild thousands of lines of program code on its own.
By this point, Silicon Valley has invented a new form of bragging. “Token-maxing” — a culture of competing over who spent more. An internal token-usage leaderboard, built voluntarily by Meta employees, ranked all 85,000 staff and awarded titles like “Token Legend” to the top users. After the press picked it up, the board was shut down within two days. Canada’s Shopify went further, sending an internal memo: “Before requesting a new hire, prove that AI cannot do the job.” Use more tokens before you hire more people.
There are scenes that aren’t quite funny. A developer at a U.S. big-tech firm, in an anonymous interview, confessed: “Because of leaderboard rankings, I keep the AI agent running for hours even when I have no real reason to use it.” Fake token usage, for the sake of evaluation scores, has arrived.
The reason this rampage is possible is price. At the time of the ChatGPT shock in 2022, inference cost $20 per million tokens. Two years later it had fallen to $0.07. A 280-fold cut. By 2030, it is projected to fall another 90% from there.
In the middle of this price war, the side swinging the sharpest blade is China. Early last year, the Chinese startup DeepSeek released a model whose performance approached OpenAI’s flagship at one-twenty-seventh the cost, startling the industry. Today, the top six AI models that developers use most are all Chinese. According to Chinese government announcements, the country’s daily token usage has grown by roughly 1,400 times in two years, reaching 140 trillion tokens. Going one step further, the Chinese government formally designated the Chinese name for “token” as “词元 (cíyuán).” 词 means “word”; 元 means “basic unit” and is simultaneously the character for the Chinese currency, the yuan. An open declaration of intent: the settlement unit of the token economy will carry China’s own name.
So who is making the most money in this enormous market? Surprisingly, not the companies that sell the tokens. Internal documents obtained by The Wall Street Journal show OpenAI doesn’t turn a profit until 2030, and Anthropic not until 2028. Revenue is exploding, but subtract the cost of model training and the picture is deeply in the red.
The real money flows to Nvidia, which makes the graphics processing unit (GPU) chips, and to the cloud companies that stack those GPUs in data centers and rent them out. Anthropic’s recent ten-year contract with Amazon Web Services is reported at $100 billion. The compute capacity bought with that amount of money is the equivalent of five nuclear power plants.
This isn’t only happening abroad. Each token requires enormous amounts of electricity and cooling water. Every time you ask a chatbot a light question, a giant data center spins up behind it and pulls a power-plant’s worth of electricity to answer you. In Korea too, requests from global big-tech companies to build new data centers keep piling up. Without anyone quite noticing, the token has crossed beyond being a unit of measurement, becoming both the currency of a new era and the electricity bill of that same era.