Recent data from the National Bureau of Statistics reveals that computing power demand in the first half of the year surged, driving electronic industry profits up by 96.9%. This growth contributed 8.5 percentage points to the overall profit increase of industrial enterprises above a designated size, with integrated circuit manufacturing profits skyrocketing by an astounding 2,579.5%.

On the same day, memory leader Changxin Technology (SH: 688825) debuted on the STAR Market, opening 471% higher and closing up 465.82%, achieving a market capitalization of 3.28 trillion yuan. This milestone surpassed Industrial and Commercial Bank of China to become the largest company by market value on the A-share market.

Behind these staggering numbers lies a crucial question: what is the real situation in the computing power market, and how does this rapid expansion affect ordinary people? At a recent industry conference in Zhengzhou, a reporter conducted field research to find answers.

Server Contract Prices Valid for Just One Week, Even Circuit Boards in Short Supply

In mid-July, Zhengzhou's temperature approached 40 degrees Celsius, yet the venue was packed with attendees. Inside the exhibition hall, a two-person-high cabinet housing China's first 100,000-card AI computing cluster drew crowds. The density of people was so high that a staff member noted the air conditioning was struggling to keep up.

Exhibits were divided between server component manufacturers and AI application firms. The hall buzzed with activity as people exchanged contact information and discussed technical details. One observer remarked the scene felt reminiscent of the early days of the internet in the 1990s.

According to official statistics, computer manufacturing and peripheral equipment profits grew by 689.3% and 305.8% respectively in the first half of the year. Integrated circuit manufacturing and semiconductor discrete device manufacturing saw profit increases of 2,579.5% and 31.2%, while electronic component and materials manufacturing rose by 209.7% and 26.9%.

A product manager from Inspur Computer stated, "Even circuit boards are in short supply now. Memory prices have tripled from last year. A memory module costing 10,000 yuan last year now costs 30,000 yuan, and you still can't find one." He explained that these circuit boards are the foundation of servers, connecting CPUs, GPUs, and storage components.

An exhibitor from a GPU company shared a similar story: "We signed a contract to sell servers, but when we tried to source them, prices changed daily. It started at 200,000 yuan, then jumped to 500,000 yuan, and finally to 800,000 yuan. We couldn't fulfill the order." The product manager added, "For a period, contract prices were valid for just one week. After that, we had to renegotiate, and the price would likely be different." This supply squeeze has persisted since late last year.

How Ordinary People Can Access Computing Power

An AI product department representative explained that when ordinary people use large models, they typically rely on cloud-based inference, which uses cloud computing resources rather than their own devices. For private applications where data must stay local, local deployment is necessary. This involves purchasing a server, as the model itself is open-source and can be downloaded for free.

When asked about the computing power required for local deployment, the representative clarified it depends on the goal. For example, running the DeepSeek model locally requires one unit of resources. Fine-tuning it requires 4 to 8 times that amount, and full training requires over 20 times. Most organizations fall into the fine-tuning category.

Another product manager illustrated the training process using an example: to build a model identifying people, vehicles, animals, and air quality at an intersection, you first set up fields for recognition. Then, you feed it data—like various vehicle models—to refine its accuracy. The more data, the better the model.

For simpler tasks, such as maintaining character consistency in AI-generated videos, local deployment isn't necessary. Users can purchase tenant isolation services from major model providers.

As computing power prices soar, is it becoming more accessible or more exclusive? A senior executive from Dawning Information Industry believes it is becoming closer to ordinary people. "Take Zhengzhou's Dawn 8000 as an example. It used to serve only scientists at research institutes, but now a large portion of its resources is connected to mobile app backends, providing token services. When you open an app on your phone, the computing power behind it might come from Zhengzhou," he said.

The Growing Importance of CPUs and Storage

To the untrained eye, a server resembles a desktop computer, comprising CPUs, GPUs, storage, and cooling. However, an industry expert likened the comparison to a toy car versus a rocket. "They differ completely in design complexity, structural precision, and purpose. Developing a server system can cost tens of millions of yuan," he explained.

As computing demands evolve, the ratio of components within servers changes. The executive from Dawning Information Industry noted that storage is playing an increasingly important role in long-context inference. A manager from Hygon Information Technology added that with the rise of AI agents, many tasks—like task scheduling, tool invocation, and memory management—are actually handled by CPUs, while GPUs focus primarily on model inference.

This shift has altered the CPU-to-GPU ratio. In traditional dual-GPU servers, common configurations included one CPU with eight GPUs or two CPUs with eight GPUs, a ratio of 1:4 to 1:8. In current data center scenarios, due to increased workloads like sandboxing, storage, databases, and vector computing, this ratio has evolved to 1:1.

To make servers smaller, one expert suggested reliance on mature liquid cooling technology, which reduces physical footprint.

Computing Power Will Eventually Be Like Water and Electricity, Available on Demand

Is there structural overcapacity in computing power? The executive from Dawning Information Industry acknowledged fluctuations between shortages and surpluses but noted that AI-driven demand, especially for high-end computing, continues to grow. "Historically, the cost of achieving equal performance or results tends to decrease over time. In the long run, computing power will become more inclusive," he said.

He cited face recognition as an example: before the large model boom, it was cutting-edge technology. Now, its cost and technical barrier have dropped significantly. "Large models will likely follow the same path. Model capabilities will improve, inference efficiency will increase, and the cost per task will decline. Eventually, large models will become a basic utility like water and electricity."

On the ultimate form of computing power, he predicted a shift toward cloudification, with edge devices equipped with lightweight computing for real-time responses. Individuals won't need to buy bulky local devices; they can access powerful AI through apps.

A vice president from Hygon Information Technology echoed this view, stating that computing power will eventually be available on demand, whether through local deployment or cloud access. "All businesses will deeply integrate AI capabilities. Computing power will be everywhere, and people will use it without even noticing its existence."

He added that while large-scale computing centers will continue to handle the most demanding tasks, application scenarios will diversify. For example, a programmer might need a trillion-parameter model for coding, but a surveillance camera only needs a lightweight model for facial recognition. Similarly, embedded devices for power waveform detection or a child's math homework at home can be handled locally without cloud involvement.

Therefore, at a mature stage, AI will split into specialized forms based on application characteristics and industry needs. Future chip layouts won't be limited to data center AI chips; phones, watches, smart glasses, and all terminal devices will have corresponding AI processing power. Smart glasses can locally record meetings and generate minutes, which is essential in environments without network connectivity.

While large computing centers will continue to handle top-tier tasks requiring immense power, the variety of application scenarios will expand significantly.