AI starts to squeeze out bubbles

After a year and a half of frantic growth, the global AI industry is collectively "squeezing out the foam."

Domestically, Zhipu and MiniMax, the two most high-flying AI stocks, have both entered a correction period recently after a sharp surge.

After listing in Hong Kong in early January this year, Zhipu's stock price once soared to HK$2980 at the end of June, with a market value exceeding HK$1 trillion. MiniMax rose to HK$1330 in mid-March, with a market value close to HK$420 billion.

After hitting record highs, the two stocks trended downward amid fluctuations.

On July 17, Zhipu's stock price fell by more than 28% in a single day, closing at HK$1107, and its market value dropped to HK$515.4 billion. MiniMax also fell nearly 16% to close at HK$216, with a market value of HK$75.4 billion. On July 20, the two stocks plummeted again, with intraday declines exceeding 12% and 8% respectively.

Compared with their peak periods, the market capitalization of both companies has shrunk by more than half.

Overseas, "AI concept stocks" have also fallen into a slump.

In the United States, Elon Musk's SpaceX was listed on NASDAQ in mid-June, rising from $135 to $225 in just three days, with a market value approaching $3 trillion.

But in the following month, SpaceX's stock price continued to decline. On July 16, its closing price fell below the issue price, and about $1.24 trillion of its market value evaporated from its all-time high.

Veteran giants are also having a hard time.

Since June, NVIDIA's market value has shrunk by more than 15%, evaporating $900 billion, barely holding the $5 trillion mark. The pullback rates of Google, Amazon, and Meta are 6%, 11%, and 12% respectively. Microsoft fell by 31%, and Oracle even dropped by 40%.

The sluggish performance of heavyweight stocks has dragged down the broader market. During the same period, the NASDAQ 100 Index pulled back nearly 6%.

Worse than the NASDAQ is the South Korean stock market.

In the first half of the year, driven by the AI storage concept, the South Korean stock market staged an epic bull market. The KOSPI (Korea Composite Stock Price Index) soared from 4200 points at the beginning of the year to nearly 9400 points at the end of June, outperforming all global capital markets in terms of gains.

But then, the South Korean stock market crashed in an avalanche, triggering intraday circuit breakers multiple times, and has now fallen above 6800 points.

The most damaging part is that Samsung and SK Hynix, which occupy half of the total market value of KOSPI, plummeted by 31% and 38% respectively, leaving a large number of investors who followed the trend to buy trapped at the "peak", causing massive losses for retail investors.

This round of sharp decline in global listed AI companies and "AI concept stocks" marks that the "de-foaming" of the AI industry has officially kicked off.

This AI market rally started in the first half of 2025. The AI sectors in China, the United States, South Korea and other regions rose all the way, pushing star companies to the center of the stage and creating countless wealth creation myths.

But discerning people can see that the so-called "AI bull market" that lasted for a year and a half, while containing reasonable elements, was also mixed with considerable speculative frenzy. Especially in the South Korean stock market, the surge of the two major memory manufacturers almost broke away from all valuation models.

The capital market is full of irrationality, but in the long run, it will always return to rationality. The booming AI industry cannot escape this underlying law either. "Squeezing out the foam" is only a matter of time and method.

Compared with the rise and fall of individual stocks and the broader market, what may be more worthy of attention is: how will the development path of the AI industry be reshaped in the "post-foam era"?

1

Since the second half of last year, many investment tycoons have frequently "blown the whistle" on the AI bubble.

As early as the end of October last year, Ray Dalio, founder of Bridgewater, publicly stated that a large number of bubble phenomena are concentrated in large-cap US AI tech stocks. He believes that this bubble will burst sooner or later, just not immediately.

But Dalio's warning obviously did not extinguish people's enthusiasm, and AI stocks continued to surge wildly. Half a year later, the "stock god" Warren Buffett lamented at Berkshire Hathaway's annual general meeting, "We have never seen people so enthusiastic about gambling as they are now."

During the same period, Wall Street big short Michael Burry put it more bluntly: "Everyone believes in the two letters 'AI'... The current market feels like the final frenzy stage of the 1999 dot-com bubble."

As Buffett and others said, the excessively high "price-to-dream ratio" is the most prominent feature of the AI bubble.

After more than a month of decline, the prices of star AI stocks are still very high. Compared with their issue prices, Zhipu still has an 8.5-fold increase, and MiniMax has also risen by 30% cumulatively.

Anthropic, which outperforms all others, is the key factor behind the abnormally high "price-to-dream ratio" of AI companies.

Anthropic is undoubtedly the fastest-growing large AI company at present. Multiple external institutions estimate that in May this year, its ARR (Annual Recurring Revenue) has reached as high as 47 billion US dollars, a 4.2-fold increase compared with 9 billion US dollars at the end of last year.

The internal expectations of Anthropic are even more optimistic: revenue will reach 70 billion US dollars in 2028, while achieving positive cash flow of 17 billion US dollars.

These stunning, red-hot numbers have completely ignited the enthusiasm of the capital market.

In February this year, Anthropic's valuation in its Series G financing was only 380 billion US dollars. Three months later, it raised another 65 billion US dollars, and its post-money valuation soared to 9.65 trillion US dollars. By mid-July, there were reports that the implied valuation of Anthropic's secondary share transfers had reached 1.15 trillion to 1.2 trillion US dollars.

In the context of the token economy, various capitals are willing to believe that AI companies naturally have extremely strong growth potential, and Anthropic is the best model.

But the problem is that the vast majority of AI companies cannot match Anthropic in terms of technical capabilities, revenue scale, or growth rate. Stories like "China's version of Anthropic" or "Anthropic in a certain field" are basically untenable.

Anthropic's biggest rival, OpenAI, had an ARR of 21.4 billion US dollars at the end of last year, which increased to 25 billion US dollars at the end of February this year, with an increase of only 17%. In addition, according to the exposed internal financial forecasts, OpenAI will not generate positive cash flow until 2030, which is a huge gap from Anthropic.

SpaceX is even further behind. According to its prospectus, its AI business revenue was 3.2 billion US dollars in 2025, and it is expected to be 3.8 billion to 4.5 billion US dollars in 2026, with an increase of only 19% to 41%.

The same is true for domestic companies. Zhipu's ARR growth is even faster than Anthropic's. It was only about 100 million US dollars in early January this year, and has now risen to 1 billion US dollars. But annual revenue of tens of billions of RMB cannot support a market value of hundreds of billions of Hong Kong dollars after all.

Anthropic is not a typical case, but the only one. Investors cannot generalize from it and take it for granted that others can grow as fast as Anthropic.

What's more, some AI companies are actively stepping on the brakes, and the first to bear the brunt is computing power.

In mid-June, it was revealed that Microsoft terminated the computing power negotiation with Oracle worth more than 30 billion US dollars. After SpaceX "acquired" xAI, it leased its computing power clusters to companies such as Anthropic and Google one after another, with a total contract scale of more than 800 billion US dollars. Blackstone, a PE giant, abandoned its original plan to invest 1 trillion US dollars in the world's largest data center project.

In the past, computing power was insufficient, but now there is more than enough computing power. This change reflects that the AI industry is quietly decelerating. The myth of Anthropic cannot be copied infinitely after all.

Most of these actions took place after June, basically in sync with the global collective pullback of AI stocks. It is not difficult to see that when the AI bubble becomes more and more obvious, market funds are often the most sensitive and fastest to react.

2

In addition to overvaluation, the deeper reason for the AI bubble is that the future vision of AI is magnificent and spectacular, but there are very few things it can do right now.

In life scenarios, AI has gone beyond the chatbot era and is evolving towards the agent era. "AI handling tasks" has become the development direction of the entire industry.

However, do ordinary people really need "one-sentence shopping/food ordering/taxi hailing"? I am afraid this still remains to be seen.

After all, even without the help of AI, these daily behaviors do not take much time; if you insist on using AI, in many cases the results you get are not much different, but it adds a secondary confirmation step, which drags down efficiency and experience.

In productivity scenarios, the most important application of AI at present is programming.

Compared with "manually writing code", AI programming has immediately improved work efficiency; especially after the popularization of agent frameworks such as OpenClaw, AI programming has become a skill that programmers must master and use frequently.

A senior insider in the software industry revealed that there are currently about 5 million programmers in China. On GitHub, there are more than 2.1 million active developers from China. According to a survey by *Caijing*, in companies such as Alibaba, Tencent, and Xiaomi, product R&D personnel consume as high as 200 million to 300 million tokens per day, and high-intensity development can even reach more than 500 million.

The millions of users and hundreds of millions of tokens consumed per person per day have made AI programming a thriving good business. The key reason why Anthropic is so prominent is that Claude Code is overwhelmingly dominant and has few rivals.

But the problem is that if AI is only used to improve the efficiency of programmers, its user scale and imagination space are limited after all, and it is far from the ultimate vision of AGI.

Take OpenAI as an example. From May to June this year, ChatGPT had more than 1 billion weekly active users, and the number of individual paid subscriptions exceeded 50 million.

On the other hand, its programming tool Codex has just exceeded 8 million daily active users. Roughly calculated, only one out of 125 ChatGPT users uses Codex, and the penetration rate of AI programming is less than 1%.

AI companies are well aware of the limitations of AI programming and are also exploring other paths, such as AI-assisted decision-making, AI collaborative office, AI multimedia creation, and AI scientific research. But these scenarios are not mature enough, and their commercial scale cannot keep up with AI programming.

This AI implementation deadlock of "not knowing what else to do except programming" has already begun to show signs.

The narrow application scenarios of AI, relying only on programming and chatbots, are not enough to drive the sustained and rapid growth of token consumption. The reduction in token consumption is directly reflected in the performance of major cloud vendors.

According to a report released by Shenwan Hongyuan in May, in the first quarter of this year, the total cloud business revenue of the four major US cloud vendors—Amazon, Microsoft, Google, and Oracle—reached 97.2 billion US dollars, a year-on-year increase of 37%, which is still growing rapidly.

However, according to Bloomberg's consensus expectations, the above revenue will grow to 110 billion and 120 billion US dollars respectively in the second and third quarters, with a quarter-on-quarter growth rate of about 13% and 9%, which is significantly lower than the year-on-year growth rate. This also means that global token consumption is still growing, but decelerating.

On the other hand, the price that enterprises pay for tokens is also decreasing.

According to estimates by market research firm Silicon Data, since May this year, the global token expenditure index has fallen by nearly 20%. This index weighted-statistics the global enterprise procurement prices of large model APIs, calculated based on the comprehensive payment cost per million tokens; since its launch in December last year, the increase once nearly doubled.

Data from China International Capital Corporation (CICC) also confirms this. It pointed out in an article that calculated by weighting the real market usage, the average expenditure per million tokens has fallen by 18% from its peak of $2.8 in mid-May to $2.31.

This is partly due to the successive price cuts by major AI companies, but it also reflects that most enterprises and individuals do not need to consume tokens at high intensity, let alone use expensive flagship models every day.

AI programming has proven to be a good business, but AI cannot only be used for programming. If AI companies cannot find more monetizable and scalable implementation scenarios beyond programming, then no matter how many features they make or how many demos they create, they are just piling up the foam until it bursts.

Before giving more answers, the second-best choice for AI companies is to actively manage expectations, burst potential risks, and promote a soft landing of stock prices. Returning from the sky to the ground has gradually become an urgent priority for AI companies.

3

The global AI "foam squeezing" is a recalibration of the current value and long-term path of the AI industry.

In the previous stage, even if AI companies could not become Anthropic or find a wide range of implementation scenarios, they could describe a grand blueprint to the market through large-scale investments, and then convert it into real money in the secondary market.

This value-building logic mixed with irrational elements was not only widely adopted by the AI industry, but also recognized by investors, and eventually became one of the motivations for tech giants to "compete with each other".

This year, the total capital expenditure of the five Silicon Valley giants—Microsoft, Amazon, Google, Meta, and Oracle—exceeds 700 billion US dollars. The huge bill brought them gains in the first half of the year.

In China, the annual capital expenditures of ByteDance, Alibaba, and Tencent are also measured in hundreds of billions of RMB, which is the first time in their respective development histories.

However, in the past month, the stock prices of large companies have corrected one after another, indicating that the simple and crude logic of "investing in exchange for market value" has begun to collapse. Compared with how much money the giants plan to spend, investors are more concerned about how well their AI businesses are doing.

On the other hand, the AI bubble has obscured the commercialization dilemma of the entire industry and distorted the strategies and tactics of participants.

During the bubble period, AI companies competed on user scale and financing capabilities. With users, there is no worry about getting investment; with investment, they can continue to attract more users through market promotion, subsidies, and other means. This is not a feasible business plan, but it is a story that investors are willing to hear.

But when the foam gradually dissipates, "spending money in exchange for users" also no longer works.

Compared with acquiring new users, AI companies need to find a commercial closed loop that suits them, verify its feasibility as soon as possible, calculate when they can cross key nodes to achieve scalable revenue, and eventually turn losses into profits and no longer need external capital injection.

This also means that it is increasingly important for AI companies to hold onto cash. Recently, many companies have intensively raised funds and sprinted for IPOs, which is actually preparing for the changes in the "post-foam era".

The AI industry is piercing through the foam. Its script is highly similar to the dot-com bubble and the mobile internet bubble back then.