AI Wealth Truth (67): Why "AI Democratization" Is a Lie

I. "AI is democratized!" "Everyone can use ChatGPT!" "AI gives ordinary people abilities that only big companies used to have!" This is one of the most popular narratives today. But let us examine carefully what this "democratization" means.

II. What is "democratization"?

III. Technological democratization means: barriers drop, and more people can use and benefit. The spread of printers was democratization. Everyone could print. The spread of smartphones was democratization. Everyone could get online. Using AI is indeed democratizing. But that is not the whole story.

IV. Distinguish two kinds of cost:

V. Marginal cost: the cost to serve one more user. The marginal cost of using ChatGPT is low, from a few cents to a few dimes per query. This part is indeed "democratizing."

VI. Fixed cost: the cost to build the capability itself. Training GPT-4 may have cost more than $100 million. Building the training infrastructure may have cost billions. Accumulating training data takes years and massive resources. This part is not democratized at all.

VII. What does this mean?

VIII. You can use AI, but you cannot own AI. You can call APIs, but the model belongs to someone else. You can benefit, but others benefit more. You are a consumer, not a producer.

IX. The barrier has moved; it has not disappeared. The old barrier was "knowing how to code." Now many people can use AI. The new barrier is "training your own model." Only a few companies can do that. Technological progress always creates new inequality.

X. Let us look at the real structure of the AI industry:

XI. Top layer: model builders. OpenAI, Anthropic, Google DeepMind. They control the most powerful models. They capture the highest valuations and profits. There may be fewer than 10 companies in the world at this layer.

XII. Middle layer: API callers. Companies that build apps by using large model APIs. There are thousands. They depend on the top layer and pay fees upstream. They can make money, but their profits are skimmed by the upstream.

XIII. Bottom layer: end users. You and I, people who use AI services. We contribute data and attention. We get "free" or cheap services. We are at the bottom of the supply chain.

XIV. Is this structure fair?

XV. This is the classic platform economy structure. Apple controls iOS; app developers earn a bit on top. Taobao controls e-commerce; merchants earn a bit on top. AI is the same. Model developers control the core, and others attach themselves. Platforms always take the biggest slice.

XVI. Why cannot ordinary people or small companies train their own models?

XVII. The capital barrier. Training a frontier model takes hundreds of millions of dollars. That is venture-scale money. Ordinary people cannot access it at all. Capital is the first wall.

XVIII. The data barrier. Training needs massive high-quality data. Much of this data is monopolized by big companies, or requires paid access. Data is the second wall.

XIX. The talent barrier. Training large models needs top machine learning experts. These talents are locked up by big companies with high pay. Talent is the third wall.

XX. The ecosystem barrier. Users and developers have already gathered around a few major models. It is hard for a new model to get users. Network effects protect the incumbents. Ecosystem is the fourth wall.

XXI. What is the truth of "democratization"?

XXII. Consumption is democratizing, while production is concentrating. More people can use AI. But AI production is controlled by fewer and fewer hands. You get convenience, but power concentrates.

XXIII. Profit distribution is unequal. Using AI may raise your productivity. But most of the value flows to model builders and your employer. You get convenience; they get profits. "Democratization" hides the inequality of value distribution.

XXIV. How should you understand "AI democratization"?

XXV. Understand it as: the usage barrier is lower, but power is more concentrated. You can use AI, but you are not the owner of AI. This is not democracy. It is a new feudalism. You are a tenant farmer, using the lord's tools to farm, and most of the harvest goes to the lord.

XXVI. What would real "democratization" in the AI era require? The development of open-source models. Decentralized training infrastructure. Fair distribution of data. These have not happened yet. Until then, "AI democratization" is just a pretty marketing term. It makes you feel like you are sharing the fruits of progress. In reality, you work in the orchard, and the fruit belongs to the owner.