In the world of artificial intelligence, there's a technical concept that's starting to become common outside of the lab: the token. It's not a digital currency or a security code, but rather the minimum unit by which the effort, use, and growth of AI systems are measured. Understanding what a token is and why it matters is key to understanding how this technology actually works and how the performance of the companies behind it compares.
For those unfamiliar with the term, in the context of language models—such as those used by OpenAI, Google, or Anthropic—a token is a small unit of text. Unlike a word, which can be short or long, a token could be:
A whole word ("hello")
A part of a word ("car," "mobile")
A punctuation mark (".", ",")
Even a space
AI models don't work with phrases or words the way a human would; instead, they process, generate, and calculate in tokens. Every time you ask a question, you generate tokens. Every time the model answers, it produces tokens. All AI work is measured in tokens.
Tokens, we could say, are a measure of effort and scale. As models become more complex and the tasks they perform become more advanced, the number of tokens required to process an input and generate a suitable output also increases. Therefore, the number of tokens used is a good way to measure how much "work" an AI system has done.
And this data isn't just useful for engineers or scientists. Increasingly, financial analysts are considering it a key metric for valuing companies in the sector.
For example, according to recent figures, Microsoft processed more than 100 billion tokens in the first quarter of the year, five times more than the same period the previous year. This figure was revealed by company executives and highlights not only the growth in the use of its AI systems, but also its growing investment in infrastructure and development.
Meanwhile, at Google's latest annual conference, CEO Sundar Pichai revealed that the number of tokens processed monthly by Google products had increased 50-fold in a single year. A figure that is not only impressive, but also shows the depth to which AI is being integrated into everyday products such as the search engine, Gmail, Google Docs, and Android itself.
For users, tokens are "hidden": we don't see how many are used when we query a model like ChatGPT or another. We only see the response. But for tech companies and their investors, tokens are like kilowatts for a utility or flight minutes for an airline: a direct metric of usage, capacity, and business.
Furthermore, AI models are marketed based on tokens. Platforms charge developers, companies, and business users based on the number of tokens they consume. The more complex, the more tokens. The more tokens, the higher the cost. Therefore, optimizing token usage has become an essential part of AI-based product design.
But the impact of this "tokenization" doesn't stop with the software. The companies that build the hardware needed for these AIs to work—such as Nvidia, AMD, TSMC, Intel, and others—are also deeply involved. The more tokens a chip can process in less time, the greater its value to data centers and, therefore, to the market.
This dynamic has led companies like Nvidia, maker of the world's most powerful GPUs, to see historic increases in their market value, directly driven by the demand for infrastructure capable of supporting the explosion of tokens generated every second worldwide.
OpenAI and Token Limits
To put this in perspective, here are the current limits on OpenAI models, one of the industry leaders:
GPT-3.5: up to 4,096 tokens per input + output combined
GPT-4 (standard version): up to 8,192 tokens
GPT-4 Turbo: up to 128,000 tokens (equivalent to about 100,000 words)
These limits determine how much information can be processed and responded to in a single session. And, of course, the price also varies depending on the model and the number of tokens used. A long, detailed query in GPT-4 Turbo, for example, can cost significantly more than a short answer in GPT-3.5.
Toward a universal metric?
Since tokens represent both the volume and intensity of an AI's work, they are becoming a de facto standard for analyzing and comparing companies in the sector. While there is still no universal way to verify these figures—each company publishes the data it chooses to display—the trend is clear: tokens are the new "currency of effort" in the AI ecosystem.
And as more sectors integrate AI into their operations, it will become increasingly common to hear about tokens not only in technical settings, but also in economic reports, stock market analyses, and discussions about digital sustainability.
In short, tokens are much more than a technical curiosity: they are the key unit for understanding the present and future of Artificial Intelligence. Where lines of code were once counted, tokens are now counted. Because in the new era of intelligent computing, everything—absolutely everything—goes through there.
EN 
ES