Out of control, the first bubble of AI is programmers

1/4

Programmers Consume the Most Tokens

This year's AI market boom stems from the explosive growth of tokens, with both quantity and price rising. By early June, Anthropic's ARR (Annual Recurring Revenue) reached $47 billion, compared to only $14 billion in the same period last year.

However, the main downstream application of the token explosion is still programming. From the experience of reviewing the previous three technological revolutions in this series, the explosive growth of a single industry's support for such a revolutionary technology is only temporary.

Moreover, there are many bubbles in the current token consumption of AI programming. In fact, it is a loss of control in corporate budgets:

Many technology giants issued instructions to their engineering teams from 2025 to early 2026 to "use the most advanced models regardless of cost" in order to maintain the R & D speed in the fierce AI competition. Many companies did not configure strict usage limits or cost monitoring during deployment. As a result, there were many extreme cases: some teams exhausted their annual AI budgets in just a few months, and even received astonishing bills because they did not configure usage limits.

This kind of management out - of - control also occurred in the early days of cloud computing development. The technology teams had great purchasing power, but due to the lack of financial constraints and insensitivity to prices, many development environments and test clusters were forgotten to be shut down after operation and became "zombie assets". There were also many bills that reached the finance department, but no one could explain which specific functional module generated the costs.

This kind of unrestrained consumption is often temporary. Currently, the focus of AI usage by technology giants has shifted from "blindly pursuing model capabilities" to "cost auditing" and "setting financial red lines". Many companies are testing refined operation systems for token usage.

An even bigger bubble lies in the settlement mechanism of large models. The investors of the three major model companies in the United States are cloud computing providers, which "invest" in the form of token vouchers. The large models then use these vouchers to purchase cloud computing services from the investors. Although this also stems from real token demand, vouchers lack price signals and often over - stimulate consumption.

The problem is that programming is only an intermediate production process, and a large amount of terminal demand is still needed to digest it. Where are they?

In summary, the current token bubble is mainly due to the overly rapid increase in AI programming productivity and the temporary failure of the price signal mechanism, resulting in an oversupply in the "programming supply" stage. A team can now complete the programming volume of a month in the past in just one week. To digest this supply, either the terminal demand suddenly explodes, or the cost drops rapidly (through large - scale layoffs).

The development of AI cannot rely on just one or two industries. It can be speculated that although the penetration rate of AI programming applications will still rise, it is difficult to soar like in the first half of the year.

However, this does not mean being pessimistic about the AI trend. There will inevitably be multiple bubbles during the process of a technological revolution. Bubbles are not a by - product of a technological revolution but a financing method for it.

In the previous article of this series, "Many People Think Too Simply, and the Future May Experience Multiple AI Bubbles", I divided them into two categories:

The first category is that the technology itself is not mature enough, and no suitable demand and business model have been found. It belongs to the industrial - level bubble.

This stage is often a large - scale infrastructure construction period, and the industrial chain mainly relies on debt financing. The burst of this kind of bubble will lead to the bankruptcy of many enterprises and trigger a debt crisis, which is also one of the reasons for the frequent occurrence of economic crises in the 19th - 20th centuries.

The second type of bubble is that the capital's expectations exceed the growth rate of commercial applications. At this stage, downstream applications have exploded in large numbers, and investors over - linearly extrapolate and give overly high valuations. Once external events such as interest rate hikes and de - leveraging occur, the bubble will be punctured.

This type belongs to the capital - market bubble and will not affect the development of the industry.

Currently, we are still in the bubble period of the first stage. Its focus is not on the valuation of the capital market but on the industrial level. More specifically, whether AI programming has reached the peak of penetration rate.

According to the survey, more than 85% of professional developers use tools such as Cursor, Claude Code, and GitHub Copilot frequently in their daily work. If the penetration rate is really that high and no next more important demand is in sight, the bubble will burst soon.

However, this conclusion may be a bit hasty.

This article is edited based on the live - broadcast content. Welcome to make an appointment for the live - broadcast every Wednesday noon.

2/4

Code Is Becoming an Intermediate Language

Most tokens are indeed consumed by professional developers from Internet giants, but some are consumed by non - professional people who don't understand programming at all, and the latter are the most important demand.

If we understand Cursor, Claude Code, and Codex as "software for writing code", we will indeed think that the market space is limited because there are only so many programmers in the world. However, if we understand them as "Agent systems with code as the general execution layer", many phenomena can be explained.

A large number of calls to Claude Code or Cursor are classified as "code generation" in the background analysis, but in fact, the tasks executed in the background may be just "write an automated script to summarize my invoices", or "crawl data from a certain web page and save it in Excel", or "summarize my materials into a personal knowledge base and update it continuously", or "send relevant news, research reports, and meeting minutes of my selected stocks every day", or "study my official account and automatically crawl and push 5 topic suggestions every day", etc. These are all demands that have nothing to do with traditional programming.

In the past, in the programming industry, a startup team or a large enterprise had an idea for a product, found programmers to develop software, and finally executed the task. But now, it is more likely that an individual has a personal task that needs to be repeated. He tells this idea to the Agent, and the background automatically generates code and then directly executes the task.

Code is changing from the "final product" that can be delivered in the business world to an "intermediate language" for executing tasks.

Why has everything become programmable all of a sudden? Because in the eyes of AI, most complex tasks in the real world are essentially a "state machine". Organizing an Excel table, calling a search interface, and operating a web page are essentially no different from writing a logical function. There is still "code, code, code, code" in the middle, but it is hidden.

Moreover, code is the most easily verifiable complex task. If it can run, it is correct; if there is an error, it is wrong. The feedback is extremely clear.

The real revolutionary progress of generative AI lies in the ability to automatically process unstructured data, such as human natural language, which in the past required writing special data pipeline code to complete.

Take stock analysis as an example. Before the emergence of the Agent, if a researcher wanted to analyze "the performance of gold and US Treasury bonds in different inflation environments in the past ten years", he would probably open a financial terminal, export various data, classify them by dimension, calculate yields, correlations, etc., and finally write a report.

Most of the work was dealing with data. The key was that the processing of unstructured data was a difficult point in the past, but generative AI has solved this problem, making the originally implicit program structure completely explicit.

Therefore, AI Agent is not about AI - enabling the programming industry, nor is it the growth of the software development industry. Instead, it "invades" various industries. An automation system that used to require a software engineering team to build can now be set up by an individual through an AI agent in just a few minutes.

The popularization of this ability means that "programming" has become a basic general ability like word processing or mathematical operations.

Just like when I was a child, typing was a skill and a profession, typing agencies were an industry, and schools taught Wubi input method as a labor - skills course. But later, intelligent pinyin made typing a simple general function, not even a skill. Now, AI voice has completely eliminated "typing".

In the past, work needed to be done through programmers to call software productivity. Now, non - programmers can directly call software productivity through the Agent, making the combination of all "knowledge workers × Agent" equal to the work ability of a complete team in the past.

In the past, only "standardized large - scale demands" were worth software - enabling; now, "personalized and segmented demands" also have the value of being software - enabled.

Don't underestimate this change. It will create an unprecedented blue - ocean market.

3/4

99% of Demands Have Economic Feasibility for the First Time

Writing code used to be a very expensive activity. It was not that the servers were expensive, but that the people were. There were costs for demand communication, programmer development, testing, maintenance, etc.

So, a rule was formed in the traditional software world: the more users there are, the more worth developing it is.

Why are large - scale software such as ERP, CRM, and Office? Why have WeChat, Taobao, and Douyin become national - level platforms?

It's not that they are well - made. The above applications target the most common demands in modern society, which can greatly reduce the product development and maintenance costs and still have extremely high gross margins to develop more segmented functions.

This also brings a problem. A large number of programming demands that were not common enough in the past were suppressed. You want a market - quotation software designed exactly according to your investment habits. Of course, this thing is better for you than Wind, and you are willing to pay a higher price than for Wind. But sorry, this demand cannot be realized because the cost of developing it specifically for you is too high.

When the programming cost drops significantly and "personalized and segmented demands" can be software - enabled, a large number of new demands will be released.

In economics, this is called "the supply curve shifts to the right": A demand worth $10 with a development cost of $5000 will not lead to a transaction; a demand worth $10 with a development cost of $5 will lead to a transaction, and the market scale will skyrocket.

This is actually the "long - tail revolution" of AI.

The long - tail in the Internet era is Taobao selling niche products. Although hot - selling products make money, niche products are the reason why you go to the "all - powerful Taobao" instead of other e - commerce platforms.

The long - tail in the Agent era is the demand for personalized software. Imagine that you have a group of 100 people with a certain niche hobby. You only need a very low development cost to provide some sustainable - charging personalized functions, and the charging is likely to be high because the more personalized the demand, the higher the price acceptance.

In addition to charging, the development of software that purely meets personal needs, which improves personal work efficiency or reduces costs, also has commercial value.

Traditional software makes money from economies of scale, while Agent makes money from personalization.

The changes don't stop there.

In the past, software was charged per user, and all functions were packaged together. Although you only used 5% of the core functions, you still had to pay the full fee. Heavy users could use all the functions to the fullest. In essence, most light users supported heavy users, which also led many occasional users to give up buying because of the "low cost - performance".

However, the core unit of Agent is the "task". This is a completely different business model. Excel has hundreds of millions of users, but Excel processes tens of billions of tasks every day, and tasks themselves can be charged by tokens. This has transformed the traditional software ecosystem.

These software are undergoing MCP transformation, encapsulating core functions into a set of standard instruction sets. These software are no longer just a UI interface for people to use, but a set of capability modules that the Agent can call at any time. Users issue personalized demands to the Agent through natural language, and the Agent disassembles them into processes and completes subsequent cross - system actions.

Agent is no longer a software but a new software distribution method. In the future, your most valuable ability will not be using software but defining tasks and judging results.

Since the number of tasks is much larger than the number of users and is closer to users' real personalized demands, it can also activate a large number of demands that were suppressed in the past.

In the past, people had to learn to adapt to software; in the future, software will have to "learn and adapt" to each individual.

Therefore, although the valuations and logics of all software stocks have been downgraded today, there will surely be some companies that actively embrace AI in the future with larger market values than now, and there will also be enterprises that bury their heads in the sand and pretend that the world has not changed and will be completely abandoned.

To simply summarize these two parts, AI currently not only increases the efficiency of programmers by 10 times but also makes it possible for 99% of the demands that were not software - enabled due to uneconomical reasons to be software - enabled for the first time. The results created can either be charged or improve work efficiency or reduce costs in various industries.

The result is that an AI Agent industry that is 10 times or 100 times larger than the current entire software industry will emerge.

However, because Agent changes our way of life, which is the most difficult thing to achieve, far more difficult than cultivating the Internet - surfing habit in the early days of the Internet, it has led to a split among users: A small number of people cannot live without tokens, while most people only use AI chat occasionally.

This split is not only one of the reasons for the extreme difference in the current capital market but also hides the reason why AI is about to experience its first bubble burst.

4/4

The First Bubble Burst Is Approaching

The real reason why AI programming has become the first killer - application scenario of AI lies not in the number of users but in the depth of use of these core users.

AI programming consumes a large number of tokens. A programming session of Claude Code often needs to read the index of the entire project library and conduct multiple rounds of code modification, refactoring, and compilation feedback. A Cursor user may consume hundreds of thousands to millions of tokens a day, while an ordinary ChatGPT user only consumes a few thousand to tens of thousands of tokens a day.

In contrast, those non - professional users who "solve life problems with programming thinking" are still in the early stage of trying AI Agent. Currently, the main usage scenarios are "single - tasks", such as writing a copy, making a plan, and organizing a table. As a result, their token consumption is usually much less than that of developers.

Although they represent the future of AI Agent, the current growth rate is actually not high. Most of those who used "Xiaolongxia" at the beginning of the year have given up.

In the pure "token consumption" statistics, the engineering delivery of pure professional developers still accounts for 60% - 70% of the share, and the penetration rate has increased to 85%. Coupled with the fact that the finance departments of large enterprises have begun to restrict budgets, this will bring two problems: