After the Jensen Huang Incident, Has AI Computing Power Construction Entered Garbage Time?

11/18 2026 554

In mid-October, I began preparing an article about AI computing power entering garbage time.

At that time, the overseas AI industry chain had already shown adjustment signals, with Nasdaq tech stocks falling by as much as 30%. I clearly felt that AI computing power was about to enter a phased adjustment period—that is, garbage time.

Unexpectedly, this draft had a troubled fate. Every time I planned to start writing, a major news event that could not be ignored would appear. For example, the U.S. midterm elections.

"Anti-data center" became a voting issue across the United States, and the wave of opposition caused Trump to lose both chambers, then spread globally.

At that time, I was inspecting data centers and intelligent computing centers in Anhui and found that China's computing power construction was far from reaching its ceiling, and the overseas situation could not be simply applied to China.

In fact, just today, November 16, the national-level AI application DeepSeek, due to insufficient computing power, no longer supports long conversations inactive for more than 7 days and will also limit daily conversation quotas.

(NVIDIA bare-metal resources sold by a certain domestic data center)

One of the most intuitive observations is that many old data centers in various places, in order to meet growing AI computing power demand as quickly as possible, can only carry out simple renovations on the basis of the original machine rooms.

But because the circuits cannot be substantially modified and the power supply cannot keep up, some machine rooms can only run two AI computing cards per server, far from reaching the designed capacity of eight fully equipped cards. Even though the machine room conditions are inadequate, they still force AI to run, precisely because computing power demand is rising too fast while the infrastructure is far from being built in place.

So I prepared to write down these field observations: one cannot conclude that AI computing power no longer has construction value just because of a round of fluctuations in overseas capital markets.

Unexpectedly, the Jensen Huang incident became popular again, with netizens around the world collectively watching the drama, developing into his resignation on November 12. The founder of the company that became the world's most valuable company by selling AI computing power was thus driven out of NVIDIA. But on careful thought, if Jensen Huang's extramarital affair had not been exposed, would NVIDIA's stock not have fallen? As we said at the beginning, overseas capital markets have already shown signs of an AI bubble bursting.

In short, every time a major news event appears, practitioners are required to think about and answer one question: is the golden age of AI computing power over?

In fact, my judgment has never fundamentally changed: the past arms-race-style expansion at all costs certainly cannot continue forever, and the bubble will burst sooner or later. Computing power construction must change, but it is not entering garbage time.

Will intelligent computing center construction enter stagnant garbage time after the Jensen Huang incident? This depends on how well they were built before and whether they will still be needed afterward.

And it is precisely the U.S. NVIDIA ban that has caused AI computing power construction in China and the United States to bloom in two different directions.

Let us first talk about overseas. In its 2026 economic report, the Bank for International Settlements specifically discussed the circular financing problem of the overseas AI industry. Simply put, chip manufacturers such as NVIDIA and hyperscale cloud service providers such as Microsoft Azure and Amazon AWS invest in AI companies or computing power service providers, which then promise to purchase chips or computing services, forming a mutually supporting relationship of capital within the industry chain.

This growth model of lifting oneself into the sky by stepping on one's own foot uses future growth expectations to support today's investment, and then uses today's investment to prove future growth expectations. Of course, it cannot last.

But what about the actual situation in China? It is a different picture.

Take Wuhu as an example. We learned from our inspection that the local government has imposed a requirement of more than 90% localization rate for computing center construction. This means that relevant projects, at the very beginning of planning and construction, need to give priority to domestic computing power products.

However, constrained by factors such as production capacity and manufacturing yield, the supply of domestic computing power products cannot fully keep up with market demand. According to our exchanges with the industry, orders for some domestic computing power products have already been scheduled two or three years out.

Corresponding to tight supply, computing power demand is exceeding many people's expectations. But during this visit to the industry, I found that in industry and even scientific research, AI is already taking charge in core business.

For example, creating new molecules was previously thought to be work only scientists could do, but at the Precision Chemistry Laboratory of the University of Science and Technology of China, AI has begun to participate in the process from molecular design and candidate solution generation to automated experimental verification, requiring large models and agents throughout.

(An intelligent chemistry laboratory at a certain university, breaking through computing power limits)

From this perspective, the long-term growth logic of computing power demand still exists. Of course, this does not mean that China's intelligent computing centers should continue to engage in computing power arms races. Because having computing power demand does not equal having computing power consumption capacity. This is the biggest bottleneck in the current absorption of AI computing power.

In a certain central city, we noticed that the local government is not only promoting computing center construction but also cultivating computing power consumption through fiscal subsidies. Among them, the scale of computing power voucher subsidies has reached more than 100 million yuan.

This shows that local enterprises already have demand for using AI, but are not yet able to independently bear the full cost of computing power.

However, if the construction of computing centers requires government investment and enterprises' use of computing power requires policy subsidies, then how long can this development model, which relies on fiscal support at both the supply and demand ends, actually last?

Therefore, for the current construction of intelligent computing centers, we must neither miss the infrastructure construction window needed for large-scale AI application out of fear of bubbles, nor build a batch of digital white elephants lacking support from real demand, because that would not only cause economic losses but may also trigger political accountability.

Since computing power demand has not yet peaked, we should continue building computing centers, but we must change the past construction model.

For some time, data center construction has exposed many problems. In the United States, this is reflected in harm to the public interest, such as servers occupying local power supply and noise during operation, greatly affecting residents' living environment.

But China's problems are somewhat different.

Relying on the national integrated computing network and the "East Data, West Computing" project, China can coordinate computing power, electricity, and regional resources over a larger scope. The abundant wind and solar resources in the western region provide advantages in energy and environment for intelligent computing center construction. From our field inspections, the PUE of multiple domestic data centers has already been controlled below 1.5.

When the environment is not the problem, then what is the real problem in China's computing power construction? It is the existence of a large number of low-quality data centers.

After the rise of large models in 2023, in order to quickly absorb computing power demand, a large number of computing facilities converted from traditional data centers appeared on the market. But these machine rooms were originally designed according to the deployment needs of traditional servers, can only solve temporary problems, and simply cannot adapt to the requirements of high-density AI computing.

For example, the servers, power supply, and cooling facilities of traditional machine rooms are often concentrated on the same floor, and the floor load-bearing capacity is insufficient, while a high-density AI super node weighs more than 1 ton. Even if it is stuffed in, the original power supply system cannot support it, and heat dissipation is even more of a problem.

Therefore, some relatively advanced data center solutions in the industry, such as the latest 3D data center solution, use a layered sandwich design, with the water cooling layer, computing layer, and power supply layer each occupying a separate floor and carrying different functions. It can support high-density AI computing clusters such as super nodes, and subsequent equipment expansion and daily operation and maintenance are also more convenient.

(The cooling layer of some advanced computing centers)

From this, we can see that the logic of intelligent computing center construction is changing:

One is computing-power-electricity coordination. Before building an intelligent computing center, it is necessary to consider the matching of power resources and natural conditions, so that computing power and energy develop in coordination and do not bring negative impacts on the local environment. A practitioner told me that some intelligent computing centers are no longer planned by the computing power unit P, but by the power unit GW.

The other is system design. In the past, data centers used the number of racks and servers as indicators, but going forward they must move toward AI-oriented specialized design, fully considering power supply capacity, load-bearing design, overall operation and maintenance, and so on.

For example, at an intelligent computing center we visited, the outdoor chilled water storage tank was designed for one hour, while traditional data centers only planned for 15 minutes. For scenarios with extremely high requirements for business continuity, such as financial AI, such redundancy design is particularly important.

(A relatively advanced intelligent computing center with added redundancy design)

To sum up, intelligent computing centers are not impossible to build next, but they can no longer be built in the past way of blindly pursuing scale and homogeneous methods. They must be comprehensively designed from scratch for AI.

After construction, whether they are actually useful, who decides? The answer is still the industry.

Rarely is there a field like today's computing power that makes people feel it is so difficult to analyze.

Compared with algorithms, when ChatGPT first appeared, although we could not accurately predict its future, we could still roughly predict the direction of future development according to technological accumulation, model development paths, and barriers to replication.

But today's AI computing power industry has bound too many things to this chariot and has been interpreted in a very complex way. Today it is a capital conspiracy, tomorrow a political game, and the day after tomorrow it becomes gossip about a leading figure.

Passing through these complicated discussions, return to the most basic questions of the industry. That is, who is using this computing power? Who is willing to pay for it? And how many people can continue to bear this cost?

Everyone can find some answers to these questions from daily life.

For example, ordinary consumers use AI more and more frequently, but this has not automatically translated into revenue for model vendors. Some domestic large models still need to set call counts and quota limits for free users, or restrict access during peak periods. Because every call incurs corresponding inference costs for the vendor. This account of commercial revenue and computing power costs still does not balance.

(Subsidy policies in some regions require that computing power must be used for actual business needs)

The same is true on the enterprise side. Local governments can use subsidies to "help them onto the horse and see them off for a ride," but it is impossible to pay for enterprises forever. When policy subsidies are withdrawn, will enterprises still be willing to spend real money to purchase computing power services? If AI, after entering actual production, does not bring sufficient operating returns, then enterprises' computing power demand will also stop at taking advantage of subsidies.

These bottlenecks are more real and more worth questioning than Jensen Huang's gossip. From this perspective, the Jensen Huang incident merely lifted a fig leaf off the AI computing power industry. It also reminds us not to be led by every trending news story.

Intelligent computing centers have not yet entered garbage time, but we must say NO to low-quality intelligent computing centers built repeatedly, and say NO to computing power investment without support from commercial returns.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.