Summary in Plain Language
There’s a bizarre phenomenon occurring in the domestic computing power sector: on one hand, there are over 500 intelligent computing centers built with substantial investments, yet on average, 70% of the cabinets remain unused and accumulate dust. Many computing projects in western regions and even small counties are losing tens of millions of yuan annually. The city investment platforms that funded these projects are already heavily in debt, and the financial gaps are only growing larger. On the other hand, tech giants like Alibaba, Tencent, and ByteDance are competing fiercely for computing power. Alibaba Cloud doesn’t even have a single idle AI chip available. To meet their needs, these companies are not only flocking to reliable computing clusters like Ulanqab but are also spending billions of dollars to build overseas centers in Southeast Asia. Essentially, this isn’t a true surplus of computing power; rather, it’s a mismatch between the local development priorities and the actual demand from enterprises. Additionally, the production capacity of domestically made high-end AI chips cannot keep up, leading to a ridiculous waste of resources where there’s no demand for the available capacity and a shortage of capacity where it’s needed.
---
Detailed Explanation
1. Understanding the Contrast
It might be hard to grasp the extent of this contrast. Let’s put it in terms that are more relatable: out of the over 500 intelligent computing centers in use nationwide, the average utilization rate is only 30%. That’s like having 10 internet cafes with 7 computers that no one uses all day. In many newly built computing parks in the west, only 2 out of 10 cabinets have functioning servers, resulting in annual losses of several million to tens of millions of yuan. Even small counties with populations of less than 400,000, such as Dong’e in Shandong and Kazuo in Liaoning, have planned computing centers costing billions of yuan, which often end up being for show only.
On the demand side, the situation is completely different. Alibaba plans to invest 200 billion yuan in computing power by 2026, while Tencent and ByteDance will invest a combined 480 billion yuan. The amount these three companies spend on computing power this year alone is equivalent to the annual public budget of several provinces. With a shortage of computing power domestically, these giants are turning to Southeast Asia. In the state of Johor in Malaysia alone, Chinese-funded projects amount to 49 billion US dollars, with ByteDance contributing 2.1 billion. There are already 290 data centers in operation in Southeast Asia, and another 135 are under construction, most of which are backed by Chinese companies.
In short, the so-called “surplus of computing power” is a false one. The unused capacity is essentially useless, while the companies are competing for high-performance computing power needed to run large models.
2. The Reason for the Gap
The low utilization rate and resulting losses might seem puzzling. Why do local governments continue to invest heavily in computing centers? It’s not that the implementation has gone awry; rather, it’s that the goals are fundamentally misaligned. The initial goal from the central government was to achieve “independent control over computing power,” but this is an intangible goal. Local governments needed tangible indicators to report on and use for evaluation, and intelligent computing centers seemed like the perfect choice: the investment amount is clear, the number of racks is quantifiable, and the opening ceremony can make news, meeting all the criteria for “deliverable and demonstrable” achievements.
As a result, two completely different evaluation systems emerged. Local governments focus on “how much investment they’ve made in computing power this year, how many new racks they’ve added, and whether they’ve exceeded their targets.” However, they don’t consider whether companies will use the facilities or whether the cost of using the computing power is viable. Meanwhile, companies focus on “how much it costs to use their services, whether the performance is stable, and whether it can handle large model training.” There’s no connection between these two sets of priorities, leading to a situation where local governments are building vigorously, but companies are reluctant to use the facilities.
3. Why Do Companies Choose Overseas Computing Power?
Companies don’t use the idle domestic facilities. There are two main reasons:
- The production capacity of high-end AI chips is severely insufficient. By 2026, the domestic demand for AI chips will be at least 4.2 million, but the annual production capacity is only 2.6 million, a 40% gap. These scarce chips are mostly snapped up by leading companies, leaving most of the idle local centers without the necessary high-performance chips. Even if the racks were available, they couldn’t run large models, and the cost of using them would be higher than building overseas centers.
- Moving computing power overseas takes advantage of current regulatory loopholes. Chips can be installed directly in overseas data centers without being imported into China, avoiding chip import restrictions. Engineers can write code domestically and remotely use overseas computing power to train models, which complies with regulations and has been a key factor in the rapid development of Chinese large models in the past two years, helping them catch up with Silicon Valley.
However, not all domestic computing power is unused. Places like Ulanqab, which are close to Beijing, have low electricity costs, and whose governments don’t focus on superficial projects, have become some of the fastest-growing AI computing hubs in the Asia-Pacific region. Their investment scale has increased tenfold in two years, demonstrating a real match between supply and demand.
4. The Hidden Truth Behind the Chaos
While many criticize the waste of money on building computing centers, this “rush to develop” has its benefits. Without the local investments, China’s computing power wouldn’t have grown to second in the world in just a few years, providing the foundational infrastructure for large model companies. Chinese open-source models, from DeepSeek to Kimi, have closed the gap with those from Silicon Valley to within half a year. Domestic applications like DouBao and AI tools have also developed their own localized paths. However, the current bubble is concerning. Most of the investments in computing centers come from local government investment platforms that are already heavily in debt. If this continues, the resulting losses will fall on the local governments, potentially exacerbating debt issues and hindering the AI industry’s progress.
The next step is to shift the evaluation of computing power from a focus on the number of racks to the actual business outcomes and cost savings for companies. We need to stop building useless facilities for the sake of meeting numerical targets and instead invest in domestically made chips and high-performance computing power. Only then can China’s AI industry overcome the current challenges and make significant progress.