The "Token Factory" on the Steppe: How Ulanqab Became the Heart of China's AI Computing Power

As multiple U.S. states grapple with power anxiety due to excessive investment in AI data centers, Ulanqab in Inner Mongolia, China, is rewriting the geography of computing power with a different logic—here, there is no shortage of wind or solar resources, nor any lack of ambition to transform them into Tokens.

The rise of Ulanqab has even caught the attention of Wall Street, with Goldman Sachs releasing a special report titled “Ulanqab Emerges as One of the Largest and Fastest-Growing AI Computing Clusters in the Asia-Pacific.” By June 2026, Ulanqab had secured total capacity commitments of 12.5 GW from data center operators and internet companies (including operational and planned projects), nearly quadrupling from 3.3 GW in July 2025 in less than a year. This represents an expansion of more than ten times the approximately 1.2 GW of online capacity available in 2025.

In communication with Tencent Finance, Zheng Zihao, General Manager of AIDC at Envision Group, stated that Ulanqab did not “suddenly become popular,” but rather that this “chosen land” was finally rediscovered in the AI era. Although the general perception is that the U.S. lacks electricity but not chips, while China lacks chips but not electricity, it is true that China does not lack electricity. However, it still lacks the corresponding power transmission infrastructure, and Ulanqab is one of the starting points for this infrastructure upgrade.

Why Ulanqab?

The rise of Ulanqab stems from its natural advantages.

The first layer is location and latency. Located about 240 kilometers in a straight line from Beijing, the fiber-optic round-trip latency is only 4.2 milliseconds. This makes it the “low-latency backyard” for the computing demands of the Beijing-Tianjin-Hebei region, while avoiding the ceiling on electricity prices and energy consumption indicators in first-tier cities.

The second layer is energy endowment and policy. Ulanqab’s effective wind farm area accounts for one-tenth of the national total, with annual sunshine exceeding 3,200 hours. New energy installed capacity has surpassed 20 GW, with green energy accounting for about 67% of the entire region. The terminal electricity price for data centers ranges from 0.33 to 0.36 yuan per kWh, approximately 50% lower than in the Beijing-Tianjin-Hebei region. More importantly, the Western Inner Mongolia Grid supports “direct green power connection”—companies can build their own wind farms and deliver power directly to server rooms via dedicated lines, without relying on public grid intermediaries. The National Development and Reform Commission has clearly stated that new intelligent computing centers must have a green energy proportion exceeding 80%, while actual green energy consumption in Ulanqab’s new projects has steadily exceeded 86%.

The third layer is a market-oriented ecosystem. Emerging model providers such as DeepSeek and Kimi have settled in, while top-tier clients like Tencent and ByteDance have deepened their presence. The local government is advancing along a three-stage path: “Steppe Cloud Valley – Intelligent Computing City – Token Capital,” with resource acquisition following market trends. “Only top-tier tech companies are coming here; pricing and construction are highly market-oriented,” said Zheng Zihao.

Low temperatures are also a hidden bonus. With an average annual temperature of 4.3°C, natural cooling sources can be utilized for about 10 months of the year, significantly reducing basic heat dissipation costs beyond liquid cooling.

Demand-side data confirms that the investment boom is not a bubble. In 2025, the utilized capacity of data centers in Ulanqab grew by over 70% year-on-year to approximately 800 MW. Online computing power reached 165 EFLOPS by June 2026, more than doubling year-on-year, accounting for 5% and 7% of the nation’s total utilized data center capacity and intelligent computing power, respectively.

In terms of electricity consumption, data center usage in the first half of 2026 increased by 90% year-on-year to 3.3 TWh, accounting for 7% of Ulanqab’s total citywide electricity consumption (compared to 4.3% in the same period last year).

Direct Green Power Supply Becomes a New Model

In Ulanqab, the most critical infrastructure providers are undoubtedly traditional Internet Data Centers (IDCs), such as GDS Holdings and VNET Group, which have gained significant attention due to AI in recent years.

Their essence is that of data center developers and operators: acquiring land, building server rooms, and renting racks, with electricity purchased directly from the State Grid. Their core revenue comes from rack rentals.

According to Tencent Finance, GDS Holdings is primarily tied to Alibaba (as a major enterprise client) and fits the “landlord” model within traditional domestic IDCs, catering largely to cloud service providers. VNET Group serves both retail and enterprise clients; historically, it had a broad retail client base (including Microsoft Azure China operations, IBM, Momo, and TouchPal), while ByteDance has been its fastest-growing enterprise client in the past two years.

Envision is a new entrant in the data center sector, with major firms like Tencent already as cooperative clients. In August 2026, Envision’s “Galaxy Base” in Ulanqab was officially powered on, featuring a 120,000-square-meter super monolithic structure standing on the Gobi steppe. It is reported that the Envision Ulanqab Galaxy Base plans for a 2 GW capacity and will possess parallel computing capabilities for millions of cards, bringing the “direct green power supply + ultra-large-scale interconnection” computing-power synergy paradigm into the spotlight.

What is new? Zheng Zihao stated that the extremely high proportion of direct green power supply is the comparative advantage of Envision’s AI computing centers.

This company, which started in new energy, is now integrating computing centers from the energy side—offering integrated solutions for land acquisition, server room construction, and direct green power supply. In other words, Envision delivers not just server rooms, but acts as a service provider for full-stack computing-power synergy solutions.

Traditional IDCs operate on the model of “building a structure and drawing power from the grid, either from the State Grid or thermal power.” However, AI training and inference consume massive amounts of electricity with extreme volatility. The instantaneous impact load when GPU clusters start up is something traditional grids and UPS architectures simply cannot handle. When electricity costs account for 60%–80% of AIDC operating costs (excluding chip costs), and the state mandates that new intelligent computing centers must have a green energy proportion of ≥80%, questions of “whether there is enough power, whether it is green enough, and whether it is stable enough” become core competitive barriers.

Thus, Envision chose to enter the fray itself.

Building its own wind farms, direct green power connections, energy storage for peak shaving, and source-grid-load-storage integration—these are things Envision has been doing for over a decade, and now this capability is being transferred to AI data centers.

China Does Not Lack Electricity, But Infrastructure Is the Core Issue

The general perception among various sectors is that China lacks chips but not electricity, while the U.S. lacks electricity but not chips. However, Zheng Zihao believes the reality is more complex: “China certainly does not lack electricity; what is needed is the corresponding power transmission infrastructure.”

Data centers naturally develop in clusters. Once they concentrate in a specific area (such as Ulanqab or Hebei), the construction speed of local grids, substations, and transmission lines cannot keep up with the pace of rack expansion. “China’s resource mobilization capability ensures that this problem will not be as severe as in the U.S., but as computing scale rapidly leaps forward, there is still room for optimization in computing-power synergy.”

Energy storage is the key to breaking through this bottleneck. Zheng Zihao compares energy storage to a “sponge”—absorbing electricity when supply is abundant and releasing it when supply is scarce, making the power delivered to GPUs smoother and more stable. When GPU clusters start computing simultaneously, the instantaneous load impact is huge, and energy storage can absorb this shock to the grid, forming a two-way buffer.

Unlike traditional data centers that rely on UPS systems and diesel generators, proprietary power supply can enhance power predictability, optimize internal power architecture, and reduce redundant configurations. Currently, some of Envision’s computing centers have begun deploying user-side energy storage to replace diesel generators. However, to further improve the regulation capability of energy storage, cooperation from power electronic equipment such as solid-state transformers is required, and such equipment is not yet widely adopted on a large scale.

How Many GW Can Actually Be Delivered?

The ultimate amount of GW that Ulanqab can deliver remains a focus of market attention.

As of June 2026, Ulanqab had secured total capacity commitments of 12.5 GW from data center operators, internet, and AI enterprises (including operational and planned projects). Goldman Sachs stated that the 12.5 GW committed capacity is about three times that of Johor, Malaysia. The institution predicts that China’s online data center capacity will increase from 28 GW in 2025 to 53 GW in 2030, nearly doubling.

For comparison, data provided by an industry insider to Tencent Finance shows that the scheduled new data center capacity in the U.S. for 2026 is approximately 13.6 GW, with 2.2 GW actually put into operation in the first quarter, and an expected risk-adjusted addition of 11.5 GW from the second to the fourth quarters. The original schedule for 2027 reaches 36.3 GW, and calculated at a fulfillment rate of 50%–60%, the actual new capacity will be approximately 18–22 GW.

However, the actual delivery pace in Ulanqab may not be as fast as the planned figures suggest. Zheng Zihao estimates that the large-scale release of Ulanqab’s data center capacity may not occur until after 2028, considering that electricity, computing power, and supporting infrastructure still need time to be perfected.

At present, there are still differences in the pace of AI infrastructure construction between China and the U.S. Capital expenditure by U.S. tech giants may reach approximately 1 trillion yuan in 2027, while China’s investment growth is relatively slower due to restrictions on key computing resources such as GPUs. However, with the continuous development of domestic computing power, the accelerated construction of power infrastructure, and the ongoing release of AI application demand, China’s AI data center market remains in a long-term expansion cycle. While the release of Ulanqab’s capacity may take time, from a longer-term perspective, the growth space for China’s AI infrastructure construction is still worth anticipating.

Risk Warning and Disclaimer

The market carries risks; investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Responsibility for investments made based on this content lies solely with the investor.