Software long tailAuthor and source: Thought Stamp
One-fourth of developers consumed the most tokens.
This year's AI market surge stemmed from the explosive growth of tokens, with both volume and price rising sharply; by early June, Anthropic's annualized recurring revenue (ARR) reached $47 billion, up from $14 billion a year earlier.
However, the main downstream application of the token boom remains programming. Based on our analysis of the three previous technological revolutions in this series, explosive growth in a single industry provides only temporary support for such a revolutionary technology.
Moreover, the current token consumption for AI programming is itself filled with a lot of inflation, stemming from uncontrolled corporate budgets:
In 2025 and early 2026, many tech giants, in an effort to keep pace in the intense AI race, instructed their engineering teams to use the most advanced models without regard to cost. Many companies deployed these models without setting strict usage limits or cost monitoring. As a result, extreme cases emerged: some teams exhausted their entire annual AI budget within just a few months, and others received shockingly large bills due to the lack of usage caps.
This lack of governance also occurred in the early days of cloud computing, when engineering teams had significant purchasing authority but, due to the absence of financial constraints, were insensitive to costs. Many development and testing environments were left running after use, becoming "zombie assets," and numerous invoices reached the finance department with no one able to identify which specific feature or module generated the charges.
This excessive consumption is often temporary; major tech companies have shifted their AI focus from "blindly pursuing model capabilities" to "cost auditing" and "setting financial boundaries," with many enterprises testing refined operational systems for token usage.
The larger bubble lies in the settlement mechanism of large models: the investors in the three major U.S. model companies are all cloud service providers, making "investments" in the form of token vouchers; these large models then use the vouchers to purchase cloud computing services from their investors. Although this stems from genuine token demand, the vouchers lack price signals and often lead to excessive consumption incentives.
The problem is that programming is only an intermediate production process, and it still requires substantial end-user demand to absorb it—where is that demand?
In summary, the current token bubble is primarily due to an overly rapid increase in AI-driven programming productivity, coupled with a temporary failure of price signaling mechanisms, resulting in an oversupply at the programming stage. A team can now accomplish in one week what previously took a month; to absorb this surplus, either terminal demand must suddenly surge, or costs must drop rapidly (through large-scale layoffs).
The development of AI cannot rely solely on one or two industries. It can be inferred that although the adoption rate of AI in programming will continue to rise, it is unlikely to surge as dramatically as it did in the first half of the year.
However, this does not mean doubting the AI trend; technological revolutions inevitably involve multiple bubbles along the way. Bubbles are not byproducts of technological revolutions—they are the financing mechanisms of technological revolutions.
In the previous article of this series, "Many People Are Oversimplifying It—The Future May Experience Multiple AI Bubbles," I categorized them into two types:
The first category is insufficient maturity of the technology itself, lacking suitable demand and business models, representing an industry-level bubble.
This stage is often characterized by large-scale infrastructure development, with the industry primarily relying on debt financing. The bursting of this bubble can lead to widespread corporate bankruptcies and trigger a debt crisis, which is one of the reasons economic crises were common in the 19th and 20th centuries.
The second type of bubble occurs when capital expectations outpace the growth of commercial applications. At this stage, downstream applications have exploded in volume, and investors excessively extrapolate linear trends, assigning inflated valuations. When external events such as interest rate hikes or deleveraging occur, the bubble bursts.
This type of bubble belongs to the capital market and will not affect the development of the industry.
It is still in the first stage of the bubble phase, where the focus is not on capital market valuations, but rather on the industrial level—specifically, whether AI programming has reached a peak in adoption.
According to the survey, over 85% of professional developers frequently use tools such as Cursor, Claude Code, and GitHub Copilot in their daily work. If the adoption rate is truly this high and no heavier-duty demand is in sight, the bubble is about to burst.
However, this conclusion may have been reached a bit hastily.
This article is edited from the live stream content. Feel free to reserve your spot for the live stream every Wednesday at noon.
2/4 The code is being converted into intermediate language.
Most tokens are indeed consumed by professional developers from major internet companies, but some are also consumed by non-professionals who have no programming knowledge—and it is this latter group that represents the most important demand.
If you view Cursor, Claude Code, and Codex as "software for writing code," it may seem like the market potential is limited, since there are only so many programmers worldwide; but if you understand them as "agent systems with code as the universal execution layer," many phenomena become much clearer.
Frequent calls to Claude Code or Cursor are classified as "code generation" in backend analytics, but the actual tasks being performed may simply be "write an automated script to summarize my invoices," "scrape data from a webpage and save it to Excel," "compile my information into a personal knowledge base and keep it updated," "daily send relevant research reports and meeting minutes for my selected stocks," or "analyze my official account and automatically suggest five topic ideas each day," among others—all of which have nothing to do with traditional programming.
In the past, the programming industry worked like this: a startup team or large company had an idea for a product, hired programmers to develop software, and then executed the task. Today, it’s more likely that someone has a personal task they need to repeat, tells the idea to an Agent, and the backend automatically generates and executes the code.
Code is transitioning from deliverable "end products" in the business world to an "intermediate language" for executing tasks.
Why has everything suddenly become programmable? Because, in the eyes of AI, most complex tasks in the real world are essentially "state machines"—organizing an Excel spreadsheet, calling a search API, or interacting with a webpage are no different in essence from writing a logical function; it’s still "code code code code" underneath, just hidden from view.
Moreover, code is the most straightforward complex task to verify for correctness: if it runs, it’s correct; if it errors, it’s wrong—with extremely clear feedback.
The true revolutionary advancement of generative AI lies in its ability to automate the processing of unstructured data, such as human natural language, which previously required writing specialized data pipeline code.
Here’s an example of stock analysis: Before the advent of agents, if a researcher wanted to analyze “the performance of gold and U.S. Treasuries under different inflation environments over the past decade,” they would likely open a financial terminal, export various datasets, categorize them by dimensions, calculate returns and correlations, and finally compile a report.
Most of this work involves handling data, and previously, processing key unstructured data was a challenge. Generative AI has solved this issue, making the previously implicit program structure fully explicit.
So, AI agents aren't about AI-ifying the programming industry or driving growth in software development—they're about infiltrating various industries. Tasks that once required a full software engineering team to automate can now be built by individuals in minutes using AI agents.
The widespread adoption of this ability means that "programming" has become a fundamental, universal skill, much like word processing or mathematical calculation.
This is like when I was a child—typing was a skill and a profession, typing clubs were an industry, and schools taught Wubi input method as a vocational course; but later, intelligent pinyin turned typing into a simple, universal function, no longer even qualifying as a skill; now, AI voice technology has made "typing" disappear entirely.
Previously, leveraging software productivity required programmers to make calls; now, non-programmers can directly use agents to access software productivity, enabling every "knowledge worker × agent" pair to match the output of a full team in the past.
In the past, only "standardized mass demand" was worth softwareizing; now, "personalized niche demand" also has value in softwareization.
Don't underestimate this small change—it will create an unprecedented blue ocean market.
3/4 of demand, economically viable for the first time
Writing code was once a very expensive endeavor, not because servers were costly, but because of the high costs associated with human resources—communication of requirements, development, testing, maintenance, and more.
Thus, in the traditional software world, a rule emerged: the more users, the more worthwhile it is to develop.
Why are large software applications mostly ERP, CRM, and Office? Why have WeChat, Taobao, and TikTok become national-level platforms?
It’s not that they did well; these apps capture the most universal needs of modern society, which greatly reduces product development and maintenance costs while enabling high gross margins to fund the development of more specialized features.
This also creates a problem: many past, non-universal programming needs have been suppressed. You want a market analysis software entirely designed according to your investment habits—something that, for you, is clearly better than Wind, and you’d be willing to pay more than Wind’s price. But unfortunately, this requirement cannot be fulfilled—developing it solely for you would be far too costly.
When programming costs drop significantly and "personalized niche demands" can be softwareized, a large number of new demands will be unleashed.
This is known in economics as a "rightward shift in the supply curve": a demand worth 10 yuan with a development cost of 5,000 yuan will not result in a transaction; a demand worth 10 yuan with a development cost of 5 yuan will result in a transaction, causing the market size to plummet.
This is essentially AI’s long-tail revolution.
In the internet era, the long tail is about Taobao selling niche products—while bestsellers generate profits, it’s the niche items that make you turn to “Taobao, the everything store” instead of other e-commerce platforms.
The long tail of the agent era consists of personalized software needs. Imagine a group of 100 people who share a niche hobby—you can deliver specific, customizable features at extremely low development cost, and charge sustainably for them. This pricing can often be high, because the more personalized the need, the higher the willingness to pay.
Beyond fees, software development that purely satisfies personal needs—enhancing individual productivity or reducing costs—also holds commercial value.
Traditional software earns money through economies of scale, while agents earn money through personalization.
The changes don't stop there.
Previous software charged per user, bundling all features together—so even if you only used 5% of the core functions, you still paid for everything. Heavy users could exhaust all features, effectively relying on the majority of light users to subsidize them, which led many occasional users to abandon the purchase due to poor value for money.
The core unit of the Agent is the "task," a fundamentally different business model: while Excel has hundreds of millions of users, it processes tens of billions of tasks daily, and each task can be charged by token, thereby transforming the traditional software ecosystem.
These applications are undergoing MCP transformation, encapsulating their core functionalities into a standardized set of commands. As a result, they are no longer merely user-facing UI interfaces, but rather modular capabilities that agents can invoke on demand. Users can submit personalized requests through natural language, and the agent will decompose them into workflows to execute cross-system actions.
An agent is no longer just software—it’s a new way of distributing software. In the future, your most valuable skill won’t be knowing how to use software, but rather how to define tasks and evaluate outcomes.
Since the number of tasks far exceeds the number of users and better aligns with users' personalized, real-world needs, it can also unlock a large volume of previously suppressed demand.
In the past, people had to learn to adapt to software; in the future, software will “learn and adapt” to each individual.
So, although software stocks were all hit today—valuations and logic both crushed—some companies that actively embrace AI will inevitably have larger market caps in the future, while others that bury their heads in the sand and pretend the world hasn’t changed will be completely left behind.
In summary, AI is not merely increasing programmers' efficiency by 10 times—it is, for the first time, making it feasible to softwareize the 99% of demands that were previously uneconomical to automate, resulting in either monetizable solutions or significant gains in efficiency and cost reduction across industries.
As a result, an AI Agent industry ten to a hundred times larger than today’s entire software industry will emerge.
However, because agents change our way of life—and this is the hardest part, far more difficult than cultivating online habits in the early days of the internet—it has led to a split among users: a small minority cannot live without tokens, while most people only occasionally use AI chat.
This split is one of the reasons behind the stark contrast in today’s capital markets and also conceals the reason why AI is on the verge of its first bubble burst.
4/4 The first bubble burst is approaching
The real reason AI programming became the first killer application for AI is not due to the number of users, but rather the depth of use by these core users.
AI programming consumes a huge number of tokens; a single Claude Code programming session often requires indexing the entire codebase and involves multiple rounds of code modifications, refactoring, and compilation feedback. A Cursor user may consume hundreds of thousands to millions of tokens per day, whereas a typical ChatGPT user consumes only a few thousand to tens of thousands of tokens per day.
In contrast, non-professional users who apply programming thinking to solve everyday problems are still in the early stages of experimenting with AI agents and primarily use them for single tasks, such as writing copy, creating plans, or organizing spreadsheets, resulting in significantly lower token consumption than developers.
Although they represent the future of AI agents, their current growth rate is actually quite low—most users who started with "crayfish" at the beginning of the year have since abandoned it.
In pure "token consumption" statistics, engineering deliverables from professional developers still account for 60%–70% of the share, while penetration has increased to 85%. With large enterprises' finance departments now beginning to constrain budgets, this will create two issues:
The first issue is already happening: if AI adoption is key to future business strategy, the first step is to cut non-AI IT budgets, the second is to limit investment in non-AI business areas, and the third is to eliminate roles that AI can replace.
The first step has already occurred, leading to the "killing logic" within the tech industry against traditional software and the internet;
The second step is underway, leading to a valuation decline in non-AI traditional industries;
If a third step truly occurs and macro demand is hit, it becomes a "revenue killer" for most industries.
The third step is coming soon. Why are U.S. Treasury yields remaining high? It’s not due to oil prices—they’re already below pre-war levels—but because AI companies are heavily using bond financing, pushing up credit bond rates. Higher financing costs will inevitably weigh on investment and demand across most non-AI industries.
This is also the “crowding-out effect” inevitably experienced in the early stages of a technological revolution, where massive investments have displaced normal investments in other sectors, even though productivity gains and new applications from the revolution have not yet scaled up.
The second issue: even after squeezing non-AI budgets to the extreme, companies will still need to revisit and optimize their AI expectations—similar to the early days of cloud computing—because this is the scenario with the highest price sensitivity. The impact on token prices will precede the impact on traffic, which is extremely damaging to the narrative around ARR growth.
This will lead to the first AI bubble bursting—both an industry bubble and a capital bubble—resulting in widespread bankruptcies and consolidations among startup AI companies, delayed or canceled capital expenditures by tech giants, and a drastic reduction in the valuations of large models. I expect this to occur within the next one to two years.
The hardest-hit were the companies in the mining chain, whose valuations are heavily dependent on the capital expenditures of tech giants and which operate in intensely competitive markets with rapid technological evolution, leaving their business models utterly vulnerable.
Of course, with every technological revolution, there is initially an overestimation of short-term changes, followed by an underestimation of long-term impacts. For investors who truly believe AI will change the world, the bursting of the bubble is a good thing:
First, the bursting of the capital bubble allows us to acquire companies with genuine long-term investment value at reasonable prices;
Second, the bursting of industry bubbles allows us to identify the niche sectors within the supply chain that truly possess competitive advantages and moats.