August 14, 2026 (InvestinChina.asia) — Computing power has become the scarcest strategic resource of the artificial-intelligence era. From the grasslands of Inner Mongolia to the highlands of Qinghai, from the old industrial base of Heilongjiang to the green-power frontier of the west, a nationwide race to build AI computing infrastructure is underway. After visiting computing centers in Ulanqab, Haidong and Harbin, reporters found projects advancing rapidly across all three regions, with training-oriented capacity moving decisively westward. A new computing-power map spanning east and west, north and south, is taking shape at accelerating speed.
“Before the data center is even built, the orders have already arrived,” said Ji Xinhua, chairman and CEO of UCloud, describing the appetite for high-density computing space in Ulanqab.
Ulanqab: Where Grasslands Meet Gigawatts
In Ulanqab, Inner Mongolia, data-center projects from Huawei, Apple, Alibaba, Kuaishou and Tencent have successively landed, forming one of the country’s most important AI-computing corridors. DeepSeek is reportedly planning a large AI data center in Ulanqab, adding roughly 1 GW of computing capacity, with partial operations targeted for late 2027 or early 2028. Industry sources confirmed to reporters that DeepSeek had already been leasing data-center space in Ulanqab.
UCloud’s Ulanqab intelligent computing center, a flagship facility for the cloud provider, covers about 140,000 square meters and is designed to house some 12,000 cabinets of varying power specifications, serving large-model training, inference and smart-terminal applications. “Our road is called UCloud Avenue; next door is Alibaba Avenue,” Ji said, gesturing at the cluster of neighboring tech campuses. He revealed that existing data centers run at very high utilization rates and that new facilities are attracting orders before construction even begins.
Ji summarized four reasons for choosing Ulanqab at scale: cheap electricity, a cold climate that aids heat dissipation, proximity to Beijing, and the ability to satisfy international clients such as Apple with strict 100% green-power requirements. “In the AI era, the core output of computing power is the token, and the ultimate foundation of the token is electricity. Cheaper electricity means cheaper tokens — and that is a huge advantage in Ulanqab,” he said.
Alibaba Cloud counts Ulanqab as one of its eight domestic data-center nodes, with green electricity accounting for roughly 45% of its local consumption. Wang Chaoyang, general manager of Alibaba Cloud’s global data-center business, said power price was the decisive factor: “The local tariff is about 0.32 to 0.35 yuan per kilowatt-hour, while in eastern and southern China it is generally 0.6 to 0.9 yuan. For a gigawatt-class data center, siting it in Ulanqab versus the east means a difference of 5 billion yuan in annual electricity costs.”
Qinghai: Highland Province Becomes Nation’s First Green-Power-Computing Pilot
Roughly 2,000 kilometers west of Ulanqab, Qinghai is racing ahead on the strength of its clean-energy endowment and cool climate. The province has become the country’s first pilot zone for green-power-computing coordination. By the first half of 2026, Qinghai had built 49,760 standard racks, with total installed computing power reaching 28,116 PFLOPS, forming an integrated supply system combining general-purpose, intelligent and supercomputing capacity.
At China Mobile’s Plateau Big Data Center in Haidong — the earliest and largest green data center on the Qinghai-Tibet Plateau, commissioned in 2015 — 7,153 standard racks deliver 18 MW of IT power and host more than 300 AI large models via the operator’s western MOMA platform node. Since 2025, China Mobile Qinghai has invested over 2 billion yuan to advance its “2+8+X” tiered computing-infrastructure layout. In 2026, the center’s Phase II and the first stage of the Xining Green-Power Intelligent-Computing Integration Demonstration Base were completed, while the Qaidam green micro-grid computing demonstration project broke ground simultaneously.
The Phase II project, with total investment of 2.5 billion yuan, will erect three data-building blocks and two power plants in three stages. The first stage, with 907 million yuan invested, is topped out and in final equipment commissioning; it will deliver 14,000 standard racks, 35 MW of IT power and computing capacity of 32,400 P.
Heilongjiang: The Northeast Catches Up Fast
Heilongjiang started later but is moving quickly. In 2024, the first phase of Harbin’s Digital Longjiang Intelligent Computing Center and China Mobile’s Harbin node — a cluster of over 10,000 cards — entered service, marking the province’s zero-to-one breakthrough in computing power. Projects now under construction, including China Mobile Harbin Data Center Phase III and the Jixi Xiansi intelligent computing center, exceed 8,000 P in combined scale.
Liu Wei, maintenance supervisor at China Mobile Harbin Computing Center, said the under-construction B04 and B05 buildings represent 1.2 billion yuan of investment, 44,500 square meters of floor area, 4,278 cabinets and 61.25 MW of IT load, and will be commissioned in 2026. A proposed B06 facility has been filed with the National Development and Reform Commission for window guidance, which would further expand the cluster’s capacity.
Power Is the New Battleground; Domestic Chips Move From “Usable” to “Reliable”
As AI-era data centers multiply in scale and density, power supply has become the decisive factor in site selection. Wang Chaoyang noted that cabinet power density in the AIDC era has reached 100 kW, and that racks of 1,000 kW will appear within two years. Ji Xinhua observed a sharp structural divide in the domestic market: low-power data centers sit idle while high-power facilities are overwhelmed with demand, often securing orders the moment a project breaks ground. “We believe demand for high-power data centers will keep growing, and we will increase our investment in them accordingly,” he said.
The bottleneck is migrating upstream from chips. Wang described the cascading constraints: “First GPU capacity fell short, then storage became the bottleneck, then electricity supply, then fiber optics. China produces 600 million kilometers of optical fiber a year, yet bottlenecks still emerge — a single 100,000-card cluster requires hundreds of thousands of kilometers of fiber. And every optical module becomes a constraint in turn.”
On the chip front, domestic hardware is advancing from “available” to “good enough.” Alibaba’s Pingtouge Zhenwu AI chip had shipped 560,000 units by April, and Wang expects domestic chip production capacity to keep climbing, accounting for an ever-larger share of Alibaba’s computing base within three years. China Mobile Harbin has achieved 100% localization of AI chips, building a high-speed, lossless two-tier domestic network architecture with self-developed SDN and GSE technologies. China Mobile Qinghai is onboarding ecosystem partners including ZhongHao CoreTech, Enflame, Biren and Moore Threads, with the center’s Phase I already hosting the Nanjing Intelligent Computing Center Qinghai node and a Sugon digital-analog training cluster, totaling 3,385 P.
Ji believes domestic GPU performance is improving rapidly under policy support and market pull. Products have reached the “usable” stage, though still trailing the highest-end foreign offerings in overall performance. “It won’t take long for the domestic GPU industry to accelerate comprehensively,” he said.
Training Goes West, Inference Sinks East
The defining trend of China’s new computing map can be captured in six words: training goes west, inference sinks east. Clusters of 1,000 and 10,000 cards initially concentrated in the Beijing-Tianjin-Hebei region, the Yangtze River Delta and the Greater Bay Area, but clusters of 10,000 to 100,000 cards have essentially migrated west of the Hu Huanyong Line — the demographic boundary dividing China’s dense east from its sparse west.
Conversely, inference-oriented computing is moving east, driven by the fact that eastern regions are the primary demand centers for AI inference. Wang expects the Yangtze River Delta, the Greater Bay Area and the Beijing-Tianjin-Hebei triangle to see a flood of inference-focused computing clusters, gradually forming industry-group coverage. “Industries with solid digital foundations will inevitably become the gathering places for token usage,” he said.
Ji Xinhua argued that the core constraint on agent applications today is not latency but the sheer scarcity of computing resources. Users tolerate several-second response times for image generation and tens of seconds for video generation, so latency is not the main obstacle; insufficient supply is. Dispersing computing power to the edge today would fragment idle capacity that cannot be efficiently pooled or rescheduled. Hence the industry currently favors large-scale centralized deployments for unified scheduling and reuse. Looking forward, as edge computing matures, caching or lightweight inference preprocessing at the edge may become a viable way to optimize the overall architecture.
The optimization of the entire system, Ji noted, is essentially a recurring cycle: identifying bottlenecks, concentrating breakthroughs, discovering new bottlenecks and breaking through again — beginning with chips, extending to memory, then to inter-card networking, and now to optical communications and interconnect.
The result is a computing-power map unlike anything China has built before: training clusters anchored in the green-power-rich west, inference clusters clustered along the industrialized eastern seaboard, and a nationwide infrastructure race where the winning variable is no longer silicon alone, but silicon multiplied by electrons — cheap, clean and reliably delivered.