DeepSeek Enters the AI Infrastructure Race in Wulanbuhe, Inner Mongolia

After going through multiple layers of security and having their smartphone cameras covered with privacy film, Phoenix Tech entered this data center located in Ulanqab, Inner Mongolia. This facility supports large-scale intelligent computing services for cloud providers and AI companies, serving as a microcosm of China’s AI industry’s competition for computing infrastructure.

No dusty chaos as imagined; instead, a low hum—cold air rushing through the containers of the intelligent computing center. Large orange-and-white cabinets are neatly arranged, still awaiting the insertion of cards.

Today, we can say that no card is idle. Delivery is productivity—the faster the delivery, the greater the customer value,” said Wang Zhaoyang, General Manager of Alibaba Cloud’s Global Data Centers.

This year, as DeepSeek As new AI giants announce the construction or co-construction of their own data centers, the industry is quietly shifting from a lightweight model of renting server rooms and purchasing computing power to gaining deep control over computing infrastructure. Companies’ strategy has evolved from merely buying GPUs to hoarding strategic assets, with tens of thousands or even hundreds of thousands of GPUs becoming the new baseline threshold.

Ulanqab has become a key battleground for AI giants competing for computing power dominance. From self-built data centers with proprietary equipment to modular "containers" delivered in just 100 days via rented facilities, a new "Foxconn" model for the AI era is emerging.

Ulanqab, a grassland city in central Inner Mongolia with an annual average temperature of just 4.3 degrees Celsius. This mid-year, DeepSeek a job posting to bring it into the AI industry's spotlight.

In the past, it was known as "China's Potato Hometown." In 2013, as mobile internet was just emerging, Ulanqab introduced Huawei to build its first cloud data center, officially ushering in the era of big data. In the years that followed, the rise and fall of the internet quietly unfolded here in symbols invisible to the human eye, as Huawei, Alibaba, Apple, Kuaishou, Century Link, and Guoshu Data successively established operations in Ulanqab.

According to data released by the local government, as of the end of 2025, Ulanqab has signed 84 data center projects, including 81 intelligent computing centers, with a total investment exceeding RMB 500 billion.

In 2026, a new protagonist takes the stage, DeepSeek The intelligent computing center in Ulanqab has begun large-scale recruitment and plans to build a super-large intelligent computing center with a total power capacity of 1 gigawatt (GW), while simultaneously hiring IDC design and planning engineers.

Following the entire AI industry's attention, a previously hidden ultra-large-scale intelligent computing center cluster has been unveiled, and Phoenix Tech visited Ulanqab this August.

One step west from Beijing brings you to Ulanqab. Outside the high-speed rail station, besides the refreshing coolness, you’re greeted by prominent red promotional banners.

350 kilometers from Beijing, reachable by high-speed rail in under two hours, with a network latency of just 4 milliseconds. The region has low annual temperatures, strong winds, dry conditions, and is not located on a seismic fault line... Most importantly, it is a “low electricity price zone,” where 90% of the power comes from green sources, with electricity costing only RMB 0.32–0.35 per kWh—placing the same data center in Ulanqab can save 5 billion yuan annually compared to neighboring cities.

Combining all these advantages, Ulanqab has become the largest intelligent computing cluster in the country, surpassing all East Data West Computing nodes, and has emerged as a true "Token Capital."

Alibaba Cloud's data center campus is also located in Ulanqab, which not only houses data centers built by Alibaba in its early years but also a newly constructed 5.0 modular data center, completed in just 100 days—this facility now supports 80% of Alibaba Cloud’s intelligent computing services.

Left image: The "yurt"-style additions on top of Alibaba's data center building for insulation; Right image: Exterior view of the data center

Here, we find that building a data center is not merely a matter of "constructing a building," just as DeepSeek We are hiring IDC design and planning engineers because this field is full of intricacies. For example, Alibaba’s self-built 2.0 architecture data center in Ulanqab represents the pinnacle of the previous generation of technology. To save energy, Alibaba piloted its proprietary Panama power supply here, drastically minimizing the voltage conversion process; to conserve water, the data center employs a closed-loop system and even uses waste heat from servers to warm equipment rooms during winter.

In the distribution room, the guide told us, “In traditional data centers, the first floor is entirely infrastructure, and the second floor houses the servers. But in our new 5.0 architecture, everything has been ‘flattened.’”

“What ‘flatting’ essentially means is turning engineering into a factory, replacing engineering with products,” explained Wang Chaoyang. From 10-kilovolt medium-voltage distribution to our proprietary Panama power supply, lithium battery backup, and liquid-cooled IT cabinets—all are pre-fabricated within containers. On-site, all that’s needed is to lift and position the containers like building with LEGO bricks and connect the cables.

In the past, building a data center required thousands of workers on-site. Now, it’s just a team of cranes pushing containers into place. Our goal is to go from ground zero to delivery in just four and a half months.” The narrator revealed that, although current supply constraints due to pulsed shortages of upstream raw materials and components sometimes make it difficult to perfectly hit the ideal 30-day prefabrication window within the 100-day timeline, the assembly process in between has been optimized to the maximum.

Over the past few years, competition among large model companies has focused primarily on algorithms, data, and model parameters.

But as we enter the era of training with tens of thousands or even hundreds of thousands of GPUs, computing power has transformed from a purchased resource into a strategic asset that determines a company’s survival.

DeepSeek The shortage is a microcosm of this trend. A harsh reality faces all participants in the AI race: data centers with available computing power have already been snapped up. The traditional construction cycle of 12 months or more simply cannot keep pace with the exponential surge in demand for tokens. In the context of the massive surge in AI demand, whoever can rapidly secure a large-scale, high-density, low-cost dedicated computing infrastructure holds the key to entering the next round of competition.

Moreover, the rise of agents has led to a surge in demand for reasoning. According to China's National Bureau of Statistics, the daily average number of token requests nationwide exceeded 140 trillion in March 2026, requiring substantial expansion of data center infrastructure.

Huge demand for tokens has led to the rapid construction of data centers; Phoenix Tech observed numerous ongoing projects in Ulanqab. In areas such as Yiwutang, Bayin, and Chayuqianqi, alongside internet giants like Huawei, Kuaishou, and Alibaba, the largest portions of land are occupied by third-party data service providers such as Century Link and China Gold Data.

How to make tokens operate more efficiently has become a new competitive priority in the AI era.

Wang Chaoyang recalled that when the modular solution was first proposed, there was strong internal opposition due to concerns about high costs. “But once the product is iterated and optimized, the costs will inevitably come down.” Even when seeking partners, resistance remained. “One partner disagreed, saying the cost was too high. But after reviewing the numbers, we found they had overestimated the cost per kilowatt by more than a third—they later regretted it, realizing they had made a calculation error.” Wang Chaoyang said that now, nearly all partners have fully accepted the solution.

What truly signaled a shift in industry direction was the reaction from competitors. “One of the largest competitors spent three months internally learning the system. When they heard we were developing the next architecture, they became very concerned. Other partners are also asking when our new standard will be released.” It’s not just in China—overseas operators are following suit as well. In his view, this solution has evolved from Alibaba’s own exploration into a path being collectively adopted by the industry.

Alibaba Cloud has even more ambitious plans for data center construction—“to become the Foxconn of this industry.” Through the ODM model, it enables the supply chain to mass-produce according to Alibaba’s proprietary standards.

Inside the modular unit, even a power module is highly standardized. The guide used the on-site air conditioning as an example: “Standard air conditioners on the market use AC power, but converting AC to DC incurs energy loss. To support this architecture, Alibaba specifically partnered with smaller manufacturers to customize DC air conditioners.” Wang Zhaoyang later added that within this modular system, “each module has been optimally tuned,” and China’s supply chain fully supports it: “We can build anywhere in any major base, and these manufacturers are willing to set up factories alongside us locally.”

Wang Chaoyang emphasized that modularization is useless if it doesn’t reach over 90%. “If you only achieve 30% or 20%, you’re merely addressing partial product issues—you can’t solve overall cost challenges, and you certainly can’t resolve overall delivery timeline issues.” He also acknowledged the practical challenge that once internal designs are finalized, they become difficult to change. “Fortunately, we can continuously iterate—modularization must be backed by versioning and standardization.”

The ability to turn data centers into standardized products could also reshape the competitive landscape of the U.S.-China AI race. Wang Chaoyang also noted that China’s manufacturing capacity is “incredible”—simply picking a few companies in Zhejiang would be enough to produce such containerized data centers. In an overseas environment where industrial labor shortages are common and delivery times can stretch to 30 months, modular products supplied by China’s supply chain would offer a decisive advantage. “When our performance catches up to our competitors’, our token costs will be utterly unbeatable.”

The limit of computing power is electricity. The ultimate goal of all extreme efficiency and cost control is electricity.

“The same data center, if located in Ulanqab versus a neighboring city, could save or cost 5 billion yuan in electricity annually,” said Wang Chaoyang. When data center scale exceeds tenfold compared to the past, and rack power consumption frequently reaches 1,000 kilowatts (equivalent to the heat generated by over 10,000 people), electricity prices become the only critical factor beyond the chips themselves.

However, cheap electricity doesn’t just appear. As AI enters the era of explosive growth in inference, computing demand is undergoing a “mutual convergence”—training workloads moving westward in a “three-tier depth” to seek low-cost green power, while inference workloads shift eastward in a “three-tier下沉” to get closer to economic hubs.

This hits the biggest pain point today: power and computing coordination. Wang Chaoyang observes that many so-called source-grid-load-storage systems are still primarily focused on absorbing renewable energy, not true computing coordination. “Power delivery takes years—often three to five—but computing delivery has been compressed to just 100 days. As tokens experience near-exponential growth, coordinated planning is critical.”

Caption: Buildings under construction are everywhere.

He determined that the gap between the current planning scale and actual demand is so large that it forces the industry to deliver explosive solutions within two years. “Power companies are now excited but don’t know how to get involved. In the future, data centers cannot remain rigid loads—they must become flexible, self-sufficient systems where computing power adapts to electricity to achieve dynamic balance.”

To achieve balance in water-scarce Inner Mongolia, Alibaba is also aggressively tackling water consumption. In Ulanqab, all data centers use reclaimed water, with the WUE of benchmark projects reaching as low as 0.088—nearly water-free. Wang Chaoyang admitted: “Using more water can save electricity, and because water is priced too low, manufacturers prefer to use it. But when building ultra-large-scale clusters, social costs must be considered to find the delicate balance between cost efficiency and sustainability.”

From the "Capital of Potatoes" to the "Capital of Computing Power," Ulanqab is witnessing a rapid surge in computing infrastructure development. As Alibaba Cloud and others transform data centers from traditional construction projects into mass-producible "products" on factory assembly lines, when DeepSeek As new entrants begin building their own computing infrastructure along this path, a new paradigm for AI infrastructure supported by Chinese manufacturing is emerging.

In the orange-and-white modular units on the grasslands, tokens are being produced around the clock. The next chapter of this computing power surge may well be the true arrival of AI for all, as these "Chinese solutions" flow globally through the supply chain.

This article is from the WeChat official account "Phoenix Tech," authored by Phoenix Tech.