Preface
I was chatting with a friend recently about a phenomenon: the API pricing of large models hasn't been dropping as fast as many expected — in some cases, it's actually going up. At the GPT-4 tier, the input/output cost per million tokens is still substantial. Claude and Gemini's flagship models hover at similar price points.
Two years ago, many said "AI costs will decline as fast as cloud computing did." But the reality is, as model capabilities skyrocket, context windows stretch longer, and multimodal becomes the default, token unit prices haven't experienced the "Moore's Law" everyone imagined. This made me rethink a question: in this AI era, what exactly is a token?
Token: The "Raw Material" of the AI Era
A token is the raw material of this era.
Think about traditional industry logic: factories purchase crude oil, ore, and cotton as raw materials, then through processing, assembly, and packaging, turn them into valuable goods — gasoline, steel, clothing. Raw material cost is the starting point of the production chain, and profit comes from the value added during processing.
Work in the AI era is converging on this same model. We feed prompts into large model APIs (consuming tokens), and models return results (output tokens). A conversation, an article generation, a code completion — all essentially consume tokens as "raw materials," then rely on our expertise, judgment, and creativity to process raw outputs into truly valuable products.
A designer uses Midjourney to generate 10 sketches, then selects, modifies, and combines them into a final design — she's buying tokens and selling design capability. A programmer uses Copilot to generate a code framework, then reviews, refactors, and tests it before delivery — he's also buying tokens and selling engineering judgment.
Why Tokens Won't Get Cheaper
Many people expect token prices to plummet, but I don't think that will happen anytime soon. Three reasons:
- Compute cost is a hard constraint — The GPU clusters needed for training and inference are real physical resources. NVIDIA's production capacity, electricity costs, and data center construction timelines can't be quickly compressed through "economies of scale"
- Model capability is inflating — The GPT-4-level model you use today will be replaced by a stronger one next year. Stronger means bigger, and bigger means more expensive. The pace of price decline gets offset by the pace of capability improvement
- Demand is exploding — From individual developers to Fortune 500 companies, everyone is integrating large models. Demand growth may outpace supply expansion, supporting prices at elevated levels
So rather than waiting for tokens to get cheaper, accept a reality: tokens are expensive, and they'll stay expensive. Just as oil will never be free, tokens as the core production resource of the AI era reflect genuine resource scarcity.
Workers = Factories
If tokens are raw materials, then every AI-using worker is a factory.
A factory's core competitiveness doesn't lie in the raw materials themselves — everyone can buy the same oil, the same ore. Competitiveness comes from processing capability: the sophistication of your craft, the uniqueness of your product design, the reliability of your quality control.
Similarly, two programmers using the same large model to write code can produce vastly different quality. The difference isn't who uses a better model, but who has stronger "processing capability" — understanding requirements, controlling architecture, maintaining code quality standards, handling edge cases.
This means: in the AI era, a person's value lies not in "whether they can use AI," but in "what they can produce with AI." Knowing how to use ChatGPT isn't a competitive advantage; creating things others can't with ChatGPT is.
Where Does Added Value Come From?
Since tokens are raw materials, added value comes from "processing." Specifically, I see three levels:
Level 1: Professional knowledge. With the same prompt, a medical professional and a layperson can get completely different conclusions. Professional knowledge determines whether you can judge whether AI output is correct, valuable, and actionable. This is the most fundamental added value.
Level 2: Engineering capability. Turning AI's raw output into a reliable product requires engineering skills — system design, quality assurance, performance optimization, security. This is what most "AI-native applications" truly need.
Level 3: Creativity and judgment. This is the hardest to replace. Asking the right questions, defining valuable directions, making decisions under uncertainty — these abilities won't depreciate with AI progress; they'll become scarcer as AI proliferates.
Everyone can buy the raw materials, but those who turn raw materials into art will always be the few.
What Should We Do
Once you grasp the "token = raw material" logic, the direction for action becomes clear:
- Invest in processing capability, not hoarding raw materials — Instead of worrying "will AI replace me," ask "can I use AI to create something better." Deepen professional expertise, cultivate engineering literacy, exercise creative thinking
- Understand token cost structure — Every API call has a cost. Learn to optimize prompts, reduce wasteful calls, choose appropriate model tiers. Not to save a few pennies, but because this reflects "cost awareness"
- Find your "processing advantage" — Everyone's knowledge background, experience, and thinking patterns are different. Find your unique "processing capability" and combine it with AI's capabilities — that's your core competitive advantage
One last thought: the essence of token economics is really the "re-pricing of human value." As AI handles more and more fundamental work, what's truly scarce is no longer "execution ability," but "judgment," "creativity," and "the ability to do things right."
Raw materials will keep getting more expensive, but good factories never lack orders.