EN / 中文

AI Computing Power Is Not Entering "Garbage Time", But Needs to Abandon Blind Expansion

by naojiti·November 17, 2026

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

At that time, adjustment signals had already emerged in the overseas AI supply chain. The maximum decline of Nasdaq tech stocks reached 30%, and I clearly felt that AI computing power was heading into a phase of adjustment, also known as garbage time.

Unexpectedly, this draft faced many twists and turns. Every time I intended to write, a major news event that could not be ignored would appear. For instance, the US midterm elections took place.

【Anti-data center】sentiments became voting issues across various parts of the United States. The wave of opposition led to Trump losing both chambers of Congress, subsequently spreading globally.

At that time, I was inspecting data centers and intelligent computing centers in Anhui Province, discovering that China's computing power construction was far from peaking, and overseas situations could not be simply applied to China.

In fact, today, November 16, DeepSeek, a national-level AI application, ceased supporting long conversations for users inactive for more than 7 days due to insufficient computing power, and also limited daily conversation quotas.

A most intuitive observation is that there are numerous aging data centers in various regions. To quickly meet the ever-growing demand for AI computing power, they can only undergo simple renovations based on existing server rooms.

However, because the electrical circuits cannot be significantly modified and power supply cannot keep up, some server racks can only run two AI accelerator cards per unit, falling far short of the design capacity for eight fully populated cards. Despite inadequate facility conditions, they force-run AI applications because the demand for computing power is rising too rapidly while infrastructure has not been built out sufficiently.

Thus, I prepared to document these field observations: one should not conclude that AI computing power no longer has construction value merely due to a round of volatility in overseas capital markets.

Unexpectedly, the Jensen Huang incident went viral again. Netizens worldwide collectively watched the drama unfold, culminating in his resignation on November 12. The founder of NVIDIA, which became the world's most valuable company by selling AI computing power, was thus ousted. But upon careful reflection, if Jensen Huang's affair had not been exposed, would NVIDIA's stock price have avoided a drop? As stated at the beginning, signs of an AI bubble bursting have already appeared in overseas capital markets.

In summary, every emergence of major news requires industry practitioners to think about and answer one question: Has the golden age of AI computing power ended?

My judgment has actually remained fundamentally unchanged: the past indiscriminate, arms-race-style expansion without regard for cost certainly cannot continue indefinitely; the bursting of the bubble is only a matter of time. Computing power construction needs to change, but it is not entering garbage time.

Talking About Bubbles, Who Is in the Bubble?

Will intelligent computing center construction enter a stagnant "garbage time" after the Jensen Huang incident? This depends on how well previous constructions were done and whether future usage is still needed.

It is precisely the US ban on NVIDIA chips that causes AI computing power construction in China and the US to diverge, each following its own path.

Let's start with overseas.The Bank for International Settlements specifically discussed the issue of circular financing in the overseas AI industry in its 2026 economic report. Simply put, chip manufacturers like NVIDIA and hyperscale cloud service providers like Microsoft Azure and Amazon AWS invest in AI companies or computing power service providers, who in turn commit to purchasing chips or computing services, creating mutually supportive relationships within the industrial chain.

This growth model of lifting oneself off the ground by stepping on one's own laces—using future growth expectations to support today's investments, and then using today's investments to prove future growth expectations—is obviously unsustainable.

But what is the actual situation domestically? It looks quite different.

Taking Wuhu as an example, our investigation revealed that local authorities require a domestic substitution rate of over 90% for computing center construction.This means that relevant projects must prioritize domestic computing products during the initial planning and construction phases.

However, constrained by factors such as production capacity and manufacturing yield rates, the supply of domestic computing products cannot fully keep pace with market demand. According to our exchanges with the industry, orders for certain domestic computing products have already been scheduled for two to three years in advance.

Corresponding to the tight supply is the fact that computing power demand is exceeding many people's expectations.During this visit to the industry, we found that AI is already playing a pivotal role in core business operations, even in industrial and scientific research contexts.

For instance, in creating new molecules, previously considered work exclusive to scientists, AI has begun participating in the process from molecular design and candidate generation to automated experimental verification at the University of Science and Technology of China's Precision Chemistry Laboratory. The entire process requires invoking large models and agents.

From this perspective, the long-term growth logic for computing power demand remains intact. Of course, this does not mean that China's intelligent computing centers should continue engaging in a computing power arms race. Because having demand for computing power does not equate to having the ability to consume it. This is currently the biggest bottleneck in AI computing power consumption.

In a certain central Chinese city, we noted that the local government is not only promoting the construction of computing centers but also cultivating computing power consumption through fiscal subsidies. Among them, the scale of computing power voucher subsidies has reached hundreds of millions of CNY.

This indicates that local enterprises already have a need to use AI, but do not yet have the capability to independently bear the full cost of computing power.

However, if the construction of computing centers requires government investment, and enterprise usage of computing power requires policy subsidies, how long can this development model, where both supply and demand rely on fiscal push, sustain itself?

Therefore, regarding current intelligent computing center construction, we must neither miss the window for infrastructure construction required for large-scale AI application due to fears of bubbles, nor build a batch of digital unfinished buildings lacking real demand support, which would not only cause economic losses but potentially trigger political accountability.

How Should Future Intelligent Computing Centers Be Built?

Since computing power demand has not peaked, computing centers must continue to be built, but the past construction model must be transformed.

Over the past period, data center construction has exposed many problems. In the US, this manifests as harm to public interests, such as servers occupying local power supplies and operational noise significantly affecting the living environment of local residents.

But China's issues are somewhat different.

Leveraging the national integrated computing network and the "East Data West Computing" project, China can coordinate computing power, electricity, and regional resources on a larger scale.The abundant wind and solar energy resources in western regions provide advantages in terms of energy and environment for intelligent computing center construction. From our field inspections, PUE (Power Usage Effectiveness) values for multiple domestic data centers have been controlled below 1.5.

If the environment is not the problem, then what is the real problem with 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, to quickly absorb computing power demand, a large number of computing facilities converted from traditional data centers appeared in the market. However, these server rooms were originally designed according to the deployment requirements of traditional servers, solving only temporary problems and being fundamentally unable to adapt to the demands of high-density AI computing.

For example, in traditional server rooms, servers, power supply, and cooling facilities are often concentrated on the same floor level, resulting in insufficient floor load-bearing capacity. A single high-density AI super node weighs over 1 ton. Even if squeezed in, the original power supply system cannot handle it, and heat dissipation becomes an even greater problem.

Therefore, some advanced data center solutions currently in the industry, such as the latest 3D data center solution, adopt a layered sandwich design where the water cooling layer, computing layer, and power supply layer occupy separate floors respectively. Each handles different functions, capable of supporting high-density AI computing clusters like super nodes, making subsequent equipment expansion and daily operation and maintenance more convenient.

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

One aspect is computing-electricity synergy.Before constructing an intelligent computing center, the matching of power resources and natural conditions must be considered, allowing computing power and energy to develop synergistically without negatively impacting the local environment. One practitioner told me that some intelligent computing centers are now planned not by computing power units (Petaflops), but by power units (GW).

Another aspect is system design.Past data centers used rack and server quantities as indicators. Going forward, they must move towards professional designs tailored for AI, fully considering power supply capabilities, load-bearing design, overall operation and maintenance, etc.

For example, in an intelligent computing center we visited, the outdoor chilled water thermal storage tank was designed for a one-hour duration, whereas traditional data centers only plan for 15 minutes. For scenarios with extremely high requirements for business continuity, such as financial AI, this kind of redundant design is particularly important.

To summarize, intelligent computing centers can still be built, but they can no longer blindly pursue scale or be built in a homogeneous manner as in the past. They must be comprehensively redesigned from scratch for AI.

Who decides whether the constructed centers are useful? The answer is still the industry.

Penetrating Through Complexity, Returning to the True Scene of the Industry

Few fields make analysis seem so difficult as computing power does today.

Compared with algorithms, when ChatGPT first appeared, although we could not accurately predict its future, we could roughly forecast development directions based on technology accumulation, model development routes, and barriers to replication.

But today's AI computing power industry has bound too many things to this chariot, interpreting it in a highly complex manner. Today it is capital conspiracy, tomorrow it is political maneuvering, and the day after, it becomes the romantic gossip of leading figures.

Cutting through these intricate discussions, let us return to the most basic questions of the industry: Who is using this computing power? Who is willing to pay for it? And how many can continuously afford these costs?

Everyone can find some answers to these questions in their daily lives.

For ordinary consumers, the use of AI is becoming increasingly frequent, but this has not automatically translated into revenue for model vendors.Some domestic large models still need to set call frequency and quota limits for free users, or restrict access during peak times. Because every invocation incurs corresponding inference costs for the vendor. Currently, the account between commercial revenue and computing power costs does not balance.

The enterprise side is similar. Local governments can use subsidies to "help them mount the horse and send them off," but cannot forever pay the bill for enterprises.When policy subsidies withdraw, will enterprises still be willing to spend real money to purchase computing power services? If AI enters actual production without bringing sufficient operational benefits, enterprise demand for computing power will stop at just grabbing subsidies.

These bottlenecks are more real and worthy of inquiry than Jensen Huang's romantic scandal. From this perspective, the Jensen Huang incident merely lifted a corner of the veil covering the AI computing power industry. This reminds us not to be dragged along by every hot news item.

Intelligent computing centers have not yet entered garbage time, but we must say NO to duplicate construction of low-quality intelligent computing centers, and NO to computing power investments lacking commercial return support.