Nowadays, many people use AI every day, but each use 'consumes' something—Token. We previously discussed this in 'The Booming Token Economy: How It Affects Your Wallet, Career, and Future Spending | An Industry Report for Everyone.' Today, a reader asked about it, so let's revisit the topic.

Token is AI's 'smallest unit of speech'

Yes, the essence of Token is simple: it is the 'smallest unit' of AI-generated content.

In Chinese, each character, word, or even punctuation mark counts as one Token (roughly speaking, though calculation rules vary by model). In English, a word or even a half-word (e.g., 'un-', 'able') also counts as one Token. When you ask AI, 'Help me write a sweet morning message for my crush,' and it replies with a mushy paragraph, every character and punctuation mark corresponds to a Token. The total number of Tokens is the sum consumed in that reply.

Where do Tokens come from? How are they produced?

Tokens don't appear out of thin air—each one costs real money. The process resembles running an 'AI text factory'...

Step 1: Prepare the 'yeast'

To produce Tokens, you need large AI models, just as brewing requires yeast. These models have digested decades of human internet content—articles, news, novels, code, comments—learning to speak, reason, and express themselves. This forms the foundation of Token production.

Step 2: Build the 'factory'

This means amassing high-end graphics cards and servers. A single server houses multiple cards, and hundreds or thousands of servers together form an AI company's 'computing power data center.'

Step 3: Start 'production'

When a user inputs, 'Help me write a report,' the system breaks the command (instruction) into small units and feeds them to the AI model. The graphics cards in the data center then spin at high speed, performing billions of mathematical operations per second, 'spitting out' content character by character. Each character generates one Token, consuming electricity, graphics card performance, and server lifespan.

Step 4: Delivery and billing

After AI delivers the 'computed' content, the backend automatically tallies the Tokens generated and charges accordingly. This is the essence of 'pay-per-Token' billing: every character we use is built on graphics cards and electricity costs.

Who profits from selling Tokens?

Here's an interesting phenomenon: the hard workers and the big earners aren't always the same. This ecosystem has three layers:

Layer 1: Top-tier profiteers: Hardware manufacturers. These include companies producing graphics cards, chips, servers, and enterprises supplying data center electricity. As long as someone wants to produce Tokens, they must use this infrastructure.

Layer 2: Middle-tier stragglers: AI companies. They build models and data centers, charging per Token, but their lives aren't easy. More users mean higher costs (buying more graphics cards, paying more for electricity, etc.). Intense industry competition means raising prices could drive users to competitors. Despite high revenue and traffic, net profits are thin or even negative—essentially working for upstream hardware manufacturers.

Layer 3: Bottom-tier freeloaders: Token 'middlemen.' Many small AI tool providers don't build models or buy graphics cards. Instead, they wholesale Tokens from large firms via APIs and resell them at a markup. Without hardware or training costs, they profit purely from price differences, living more comfortably than the big firms—like earning money by reselling tap water without owning a water plant.

This is why Token may not be a lucrative business: the real profits don't lie in 'selling Tokens' itself.

Will AI get dumber if humans stop creating new things?

This brings us to a critical question tied to our earlier discussion: Tokens ultimately originate from human-created content. All knowledge, logic, and expression in large AI models come from learning text, videos, and scientific research generated by humans over decades. Without human thought, creativity, and execution, neither AI nor Tokens would exist.

So, here's the question: If humans become overly reliant on AI, stop producing new content—no new articles, research, or ideas—and AI systems start copying each other or feeding off themselves, will AI development become unsustainable?

To some extent, yes.

Today's AI stands on the shoulders of thousands of years of human civilization. Historical data acts as AI's 'granary.' Without replenishing this 'grain,' it will eventually run out. Without continuous human input of new knowledge, perspectives, and real-world experiences, AI can only rearrange old knowledge, never generating new insights or creativity.

Some signs are already emerging: Much online fiction, copywriting, and social media content is now AI-generated→posted online→then used to train other AIs, creating a closed loop.

In short, Tokens are the smallest units of AI 'speech,' generated in real-time by high-end graphics cards and large models, each consuming hardware and electricity. In the Token business, upstream hardware manufacturers profit effortlessly, AI companies struggle to keep up (or even lose money), and middlemen profit from price differences. The ultimate risk: If humans stop creating original content, AI will lose its 'nourishment,' making sustained development impossible.

After all, no matter how intelligent AI becomes, it remains an extension of human wisdom. No matter how popular Tokens get, they depend on human creativity. Machines can 'calculate' text but cannot 'create' genuine thought.