Is the cavitation period of the AI industry coming?

Last night, while browsing YouTube, I happened to come across a in - depth video just released by Bloomberg, titled "Big Tech's $650 Billion Gamble".

In the video, the analyst pointed at the bar chart and said:

By 2026, just Amazon, Google, and Microsoft are expected to invest a capital expenditure (Capex) of $650 billion.

Immediately afterwards, he put forward a rather embarrassing conclusion: The investment is increasing exponentially, while the revenue is increasing linearly. If this problem is not solved, the AI industry in 2026 is very likely to hit a huge air pocket.

Just like an airplane suddenly falling into a vacuum while flying, everyone should be able to imagine that feeling of weightlessness. So, after watching this video, I believe that this is not only the anxiety of Wall Street, but also a transitional moment for the entire AI industry.

Let's see where this $650 billion comes from and what it can actually achieve?

The $650 billion mentioned in the Bloomberg video is a rather delicate figure. I specifically went through Goldman Sachs' original research report and found that there is a very rare "inversion" behind this figure.

How to understand this inversion?

The infrastructure has soared into the stratosphere, while the applications are still slowly climbing. Look at Amazon, Microsoft, Google, and Meta. Their capital expenditure in 2026 will also be around this figure. Where will this money go?

All of it will be used to buy graphics cards, build data centers, and even compete for power resources. This level of investment is already an infrastructure investment at the "betting on the nation's future" level.

The problem is that while one side is bustling with activity, the other side is cold. Looking at the revenue side, we can find a very real temperature difference. Let's take Microsoft, which is the most promising company, as an example.

JPMorgan recently had an estimated data: Although Microsoft Office 365 has the largest number of 450 million business users globally, only about 15 million people are really willing to spend an extra $30 per month to buy Copilot.

Let's calculate this ratio, which is 3.3%.

What does this mean? Even for the world's strongest B - end giants, even if they directly place the entrance in front of Word and Excel, 96.7% of people still think that this thing is not worth the price.

This is very interesting. On one hand, there is a crazy infrastructure investment of hundreds of billions of dollars, and on the other hand, there is a calm purchasing rate of 3.3%. These two gears simply cannot mesh at present.

So, you will find that AI is currently in a very divided state. In the capital end, it is like a nuclear weapon, and everyone is afraid of falling behind and is desperately building it. In the application end, it is just a tickler, quite interesting, but still far from being "indispensable".

I estimate that everyone who has read the financial statements has the same big question mark in their minds.

Think about it. Zuckerberg (Meta), Nadella (Microsoft), and Pichai (Google) are all extremely smart people. Can't they understand how poor the 3.3% penetration rate is?

Don't they know that if they invest $650 billion and don't make a big splash in the next few years, Wall Street will hammer their stock prices?

Of course, they know. It's just that their thinking has long gone beyond the "financial statements" that ordinary people look at and has completely entered a "wartime state".

There are three extremely cruel and also extremely wonderful game logics hidden in this.

Let's start with the most practical one. For them now, not doing anything is like a "death sentence", and doing it wrong is at most a "life sentence". Remember the summer of 2024? At that time, Google's CEO, Sundar Pichai, in order to respond to the doubts of investors, put forward that very famous "AI Manifesto":

The risk of under - investing is dramatically greater than the risk of over - investing; (The risk of under - investing is much greater than the risk of over - investing)

This statement sounded quite heroic at that time, but by 2026, it has become a huge "prisoner's dilemma". Just last month, at the earnings conference call in January 2026, an analyst sharply asked if their spending was out of control, but Pichai still didn't dare to back down.

Why?

Because the current situation is much scarier than it was two years ago. If they didn't invest in 2024, they would at most fall behind. But if they stop in 2026, in case OpenAI or Microsoft develops an AGI at the level of GPT - 6, Google won't just lose money. It might even be kicked out of the game directly.

Although the Genimi model still has a slight edge in the global echelon, they don't dare to relax at all. So, you see, this $650 billion is essentially their payment for the strategy they made two years ago.

It's an extremely expensive "survival tax". They would rather let the money rot in the data centers than let their opponents grab the ticket to the next era first.

Now let's talk about another point. I think they are now reshaping the entire physical world for the sake of computing power. If you only focus on the money, you might think it's a digital game. But if you look at what they are doing in the United States, you'll really be shocked by their "infrastructure - building mania".

The most typical example is Microsoft.

Remember the agreement Microsoft signed at the end of 2024 to restart the "Three Mile Island Nuclear Power Plant"? By this year (2026), this plan has entered the substantial "heavy construction period".

You heard it right. In order to power hundreds of thousands of H100/H200 graphics cards, Microsoft actually went for nuclear energy. This shows that in Nadella's eyes, AI has long become an "energy business".

Their current crazy investment is actually building a physical barrier. When computing power becomes a basic resource like electricity and oil in the future, whoever controls the power plants (data centers) and oil fields (chip clusters) will have the final say and control the pricing power of the new world.

As for the third point, they are violently reducing costs and waiting for that "singularity" to appear.

Looking back at history, we can find that the current "bubble" is not that absurd. It's like the construction of 4G networks and optical fibers back then.

When the optical fiber bubble burst in 2000, the utilization rate was less than 5%. Everyone was scolding and saying, "Are you crazy? Why spend hundreds of billions on such high - speed internet? Do you really need it just to send an email?"

But history has already told us the answer: It was precisely because of the oversupply of optical fibers and the sharp decline in bandwidth costs that YouTube, Netflix, and Douyin were able to develop and become popular later.

The same goes for AI now. The reason why people think Copilot is expensive, not worth the $30 per month, and not useful is fundamentally because the computing power is too expensive and the model is too slow.

Those giants' current crazy investment is actually "violently reducing costs".

They are gambling that by the end of 2026, the inference cost can be reduced to one - tenth or even one - hundredth of the current level. By then, AI will become as affordable as "water, electricity, and gas at $3", and everyone can use it.

By that time, the application layer will truly experience a Cambrian explosion.

When this cold wind blows across the domestic AI circle, what's the temperature like? To put it bluntly, we might be more uncomfortable than the American giants. Why?

The core reason is "different constitutions".

If the AI circle in the United States in 2026 has a "rich man's disease" with too much money and too many graphics cards to digest, then in China in 2026, it's a typical "low - blood - sugar" situation with not enough money, difficulty in getting graphics cards, and still having to work desperately.

You can understand this by looking at two phenomena. Since DeepSeek, like a catfish, has brought the inference cost down to a "rock - bottom price", there is no longer any premium space in the domestic AI market.

Currently, the traffic is bustling, but the accounts are cold. This "involution - style" price war has directly pierced through the profit hole of the industry. In such an extreme environment, there are clearly two completely different types of players at the domestic AI game table.

The first type of players, I call them "gamblers", are the new forces still crazily betting on infrastructure and models, such as Zhipu, MiniMax, and Yuezhianmian, which are all familiar names to us.

Their business models were already difficult before 2025. By 2026, people suddenly found a rather embarrassing thing: the models have become homogenized.

So, they are likely to hold the best "hammer", that is, the model, and spend a lot of money looking for user scenarios all over the world.

Compared with them, the second type of players are much more relaxed. I call them "full - stack landlords".

ByteDance is a very typical example. It holds a winning hand. The data is its own (Douyin/TikTok), the traffic is its own (it doesn't need to buy it at all), and the scenarios are also its own.

Why can Doubao suddenly reach the top? Because it doesn't need to beg for traffic like a startup company. It can use its own traffic pool freely, and the customer acquisition cost is almost zero.

Looking at Baidu and Alibaba again, Robin Li has been shouting for two years about "re - creating all products with AI". Now, if you look at Baidu Wenku and Baidu Search, they are using the Wenxin large - scale model to reconstruct the "retention rate".

Alibaba is even more aggressive and directly integrates Qianwen into various scenarios.

What's even more amazing is their monetization logic. The profit comes from other sources. As long as the giants can use AI to increase the advertising click - through rate and conversion rate by 1%, or sell a few more points of Baidu Cloud's computing power, the money earned will be enough to cover the electricity bills of tens of thousands of graphics cards.

Moreover, don't think that these giants only focus on the base and compete on parameters. For them, the application layer is also "a territory to be fought for" and they don't relax at all.

So, in the current trough period of the industry, full - stack is the only life - saving circle. Giants with a complete closed - loop of "data - model - application - traffic" will, on the contrary, take this opportunity to deepen their moats and harvest the entire battlefield.

According to the current consensus on Wall Street and in Silicon Valley, this embarrassing vacuum period will last at least until the second half of 2026.

So, how can we get out of this dilemma?

The conclusion given by Wall Street is simple: The only variable is Agent. In essence, it's the change from Copilot (co - pilot) to Autopilot (autonomous driving/employee).

Why is it specifically this? We need to figure out this business account clearly.

Why can't current AI make a lot of money? Because the current AI (that is, the Copilot model) is essentially still selling software (SaaS). It's like a very knowledgeable intern. You ask a question, and it answers. But in the end, it's still us humans who make decisions, take responsibility, and deliver results.

This leads to a huge "value paradox":

Since I still have to check, modify, and take care of things myself, in the end, it's just a tool. Even if a tool is sold for only $20 a month, I still think it's expensive. This is also why the penetration rate of Microsoft Copilot is only 3.3%. As software, its upper limit is really too low.

The logic of Agent is completely different. As the inference ability becomes more and more mature and the price becomes lower and lower, it can directly sell labor.

You know, the global software market (SaaS) is only a few hundred billion dollars in scale, but the global labor market (Labor) is a multi - trillion - dollar market.

Let's take a very practical example:

If you use Copilot to write a collection email for you, after it's written, you still have to check, copy, paste, and send it yourself. But if it's an Agent, you just need to give it a vague instruction, and it can handle everything.

For example:

Ask it to collect the overdue payments for this quarter. The Agent will break down the task by itself: first check the list of debtors in the CRM, then analyze the personality and repayment records of each customer. If the customer doesn't reply, it will automatically send another collection email three days later or directly make an AI voice call to follow up. Finally, it will send you the collected money and a detailed report.

The important thing is that its identity has changed. It has become a digital employee.

So, now let's do the math. If you hire a real human sales assistant, the monthly salary plus social security will cost at least 8,000 yuan, and you have to give them weekends off and they may take sick leave.

Now, an Agent that can work 24/7 without any emotions and handle thousands of customers in minutes, even if it's sold for 800 yuan a month, which is 10 times more expensive than the original software, you'll still think it's incredibly cheap.

This is the only way to get out of the industry's "air pocket":

When AI's ability is strong enough to truly "replace labor costs", the original linearly - growing revenue curve will suddenly turn into an exponential explosion.

So, in 2026, in the AI circle, besides comparing the basic strength of models, it's also necessary to compare who can complete the thrilling leap from "selling tools" to "selling services" first. When AI is no longer charged by "Token" but by the work result, this air - pocket period will finally come to an end.

This article is from the WeChat official account "Wang Zhiyuan" (ID: Z201440), written by Wang Zhiyuan and published by 36Kr with authorization.