A 'Token Factory' on the Loess Plateau—How a Small Western City is Reshaping the Computing Map in the AI Era
In the morning, a cross-border seller in Southeast Asia taps on their phone to generate product copy with AI. In the afternoon, a programmer in the Yangtze River Delta uses a large model to quickly write code. In the evening, a designer in the Greater Bay Area uses AI for one-click image editing and output... These seemingly unrelated moments in life and work may all be connected to the same place: Qingyang, Gansu. Intelligent application instructions from around the world are instantly broken down into countless tokens, traversing mountains, rivers, and seas, processed at a computing center on the Loess Plateau, and then precisely delivered back to users' fingertips. Every seamless interaction represents the most authentic pulse of the "token economy."
A token is the smallest unit for AI to understand the world and the fundamental unit for model inference. Every query, every generation, essentially consumes tokens. As AI moves from the lab into industry, tokens are no longer just jargon among technicians but have become a new yardstick for measuring the vitality of the intelligent economy. By March of this year, China's average daily token call volume had exceeded 140 trillion, representing growth of over 1000 times compared to the 100 billion at the beginning of 2024 and an increase of more than 40% in just three months from the 100 trillion at the end of 2025.
Amid this transformation, Qingyang, leveraging the national "East Data West Computing" strategy, abundant green power resources, and a high proportion of intelligent computing, has risen from a small western city to a global core supply base for green tokens. A clear transformation path from being an "electronic warehouse" to a "token factory" is unfolding.
From "Electronic Warehouse" to "Token Factory": A New Logic for Computing Power Economy Takes Root in Qingyang
"Token is the new commodity." At the 2026 GTC conference, NVIDIA founder and CEO Jensen Huang introduced the concept of a "token factory," clearly stating that the current AI industry competition has evolved from the "era of large models" to the "era of token production capacity." Future data centers will no longer be warehouses for storing files but "factories" that produce tokens. This judgment marks a fundamental shift in the role of data centers.
Traditionally, data centers focused primarily on static data storage and network exchange, offering limited value creation and weak industrial linkage. However, with the rapid proliferation of AI agents—machines that autonomously execute tasks, process logic, and generate content—data centers are fully transitioning into new production units that continuously ingest data, produce tokens in real-time, and output intelligent services. Thus, the "token factory" has become the mainstream form of the computing power economy.
The explosive growth in token calls has directly given rise to a token industry centered on standardization, scale, and low cost. AI services are becoming like water, electricity, and the internet, forming a commercial loop that can be precisely measured, continuously scheduled, and traded on a large scale. The core of the token economy is making intelligent services measurable, priceable, and tradable—every time a user calls an AI, the token consumption, power usage, and cost incurred are clear and calculable.
Facing the new opportunity of the token economy, Qingyang, located in the west, stands at the forefront of the industry by virtue of its unique resource endowment and strategic positioning. Among the eight national hub nodes of the "East Data West Computing" project, Qingyang started the latest, lacking first-mover advantage. However, it did not follow the old path of developing general-purpose computing power. Instead, it targeted the intelligent computing track from the outset, charting a new course for overtaking. This decision was well-timed. As the AI wave surged, with the concentrated emergence of domestic large models like DeepSeek and a critical window of rapid growth in national intelligent computing demand, Qingyang's early planning perfectly met this wave of demand.
By the end of 2025, the intelligent computing scale of the Qingyang data center cluster had reached 114,000 PFlops, making it the national hub node and data center cluster with the largest incremental growth, the fastest growth rate, and the highest proportion of intelligent computing. Currently, Qingyang has 21 data centers planned for construction, with completed and under-construction standard racks reaching 240,000. It is estimated that by the end of this year, the computing power scale will exceed 300,000 PFlops. Research teams from Beijing, Shanghai, Guangzhou, and Shenzhen continue to visit, and computing power demand orders are pouring in from all over the country. A green "token factory" serving the nation and radiating to Southeast Asia has already risen from the Loess Plateau.
From "Selling Racks" to "Selling Tokens": Transforming Resource Endowment into Core Token Advantage
In the world of AI, electricity is the "fuel tank" determining how far intelligence can run. According to calculations by the China Academy of Information and Communications Technology, electricity costs account for 50% to 70% of data center operating costs, representing the largest rigid expenditure for computing power. In other words, whoever can save on electricity bills gains the initiative in pricing for "selling tokens." Qingyang's advantage lies precisely in the wind and sunlight of this Loess Plateau.
As part of the Longdong ten-million-kilowatt-level energy base, Qingyang is exceptionally endowed with wind and solar resources, boasting abundant and stable green power supply. Through innovative green power aggregation models, Qingyang achieves local consumption of clean energy dedicated to computing power parks, resulting in significant electricity price advantages. Compared to large data centers in eastern China, Qingyang's electricity can be 0.2 to 0.3 yuan cheaper per kilowatt-hour. This seemingly small price difference translates into a staggering cost advantage when applied to large-scale computing clusters.
Green power supply brings not only ecological benefits but also effectively reduces production costs for computing enterprises. Coupled with the cost advantages of lower land prices in the park and supporting infrastructure like the "five horizontal, four vertical" road network, Qingyang continues to attract various computing demands. Climatic conditions further amplify Qingyang's advantages. Located in the heart of the Loess Plateau, Qingyang has a low average annual temperature and a cool climate, allowing for long-term use of natural cooling methods. This significantly reduces cooling power consumption, keeping its PUE (Power Usage Effectiveness) consistently better than the national average.
Low-cost green power, natural cooling, and efficient heat dissipation—these three combined advantages give this Loess Plateau region the confidence of a "natural data center." More forward-looking is Qingyang's new path of "zero-carbon computing" through deep integration of green power and intelligent computing. In December 2025, the Qingyang East Data West Computing Industrial Park was included in the first batch of national-level zero-carbon park construction lists, becoming one of only six "overall" creators among 52 parks nationwide.
Currently, the green power proportion in the park's intelligent computing centers exceeds 80%, the computing power to end-user electricity price has dropped to 0.398 yuan per kilowatt-hour, and the data center PUE is below 1.2. This "low electricity price, low energy consumption, low PUE" significantly reduces the computing power cost consumed per unit token, making AI large model training and inference more economical and accessible. The value of resources on this northwestern land is being reassessed under the heat of the "token wave."
Now, a group of large intelligent computing centers, including those from China Mobile, China Telecom, China Unicom, and China Energy Engineering Group, have gathered in Qingyang. The siphon effect of computing enterprises is driving the formation of clusters for diverse businesses like large models, intelligent driving, and AI data annotation in Qingyang. Currently, models such as MiniMax, Kimi, Zhipu AI, and DeepSeek have all deployed training operations in Qingyang. A computing power landscape characterized by high density, high energy efficiency, and high intelligent computing has taken shape.
Efficiency is the key metric for measuring the competitiveness of this "token factory." One Petaflop (P) of computing power is roughly equivalent to the combined computing power of 500 high-performance computers. What does this mean? In processing massive astronomical data, analyzing data for 200,000 celestial bodies manually might take researchers 169 days. Relying on 1000 Petaflops of computing power, the same task can be completed in just 10 seconds. This exponential leap in computing efficiency is accelerating the transformation of Qingyang's resource endowment into a production advantage for tokens.
From "Building Good Computing Power" to "Using Computing Power Well": Qingyang's Solution for the Last Mile
If green power and computing power give Qingyang the capability for "token production," then computing-network synergy, low-latency transmission, and national scheduling enable Qingyang to truly realize the value of "token supply." After tokens are "produced" in the western data center, how do they traverse thousands of miles to be delivered across the country and finally appear seamlessly on users' mobile screens? An efficient computing power network is the key to solving this problem.
Industry consensus holds that only networks with latency under 20 milliseconds can truly support "intelligent computing." It is this hard indicator that led Qingyang to place network construction on par with computing power construction in importance. Qingyang is located near the national geographical center, within 2000 kilometers of first-tier cities like Beijing, Shanghai, Guangzhou, and Shenzhen. This gives it a natural locational advantage for radiating nationwide, minimizing the average total time for data transmission and processing from sender to receiver—commonly known as "latency."
To open the "highway" for "East Data West Computing," China Mobile built Gansu's first 400G OTN backbone optical network, reducing the latency between Lanzhou and Qingyang to under 3 milliseconds. This forms an ultra-low latency circle of "1-3-5" milliseconds within the province and "6-8-12-15" milliseconds between provinces, providing a high-speed channel for Qingyang's integration into the national computing power market.
With the network connected, computing power can truly "run." But having the "road" is not enough; efficient "scheduling" is also needed. Addressing the limitations of traditional Content Delivery Networks (CDN) in meeting the high-concurrency, low-latency demands of AI large models, Qingyang is driving a comprehensive upgrade of the network distribution system from static content caching to dynamic token real-time scheduling.
Traditional CDNs primarily serve the accelerated download of static resources like videos and images. Their architecture struggles to meet the stringent requirements of generative AI for real-time computing power scheduling—token generation requires millisecond-level response speeds and cross-regional computing power coordination. Qingyang has addressed this technical bottleneck by building a deterministic network. This network acts as a "highway" for token computing power flow, dynamically allocating computing resources based on real-time load to ensure the efficient operation of large-scale token applications, thereby supporting daily call volumes at the hundred-billion level.
This upgrade relies on the super-ten-thousand-Petaflop-scale intelligent computing cluster deployed in Qingyang and autonomous scheduling platforms from companies like Lingqiong Shunlian. By integrating resources from various computing centers within Gansu Province, it achieves "interconnected computing forming a network" and "unified scheduling by one network." This full-stack token production and scheduling system not only increases local computing power utilization to over 90% but also breaks geographical limitations. Currently, Qingyang has established computing power "pairing" agreements with 58 domestic cities, forming a scheduling network for "domestic circulation and global radiation." With the accelerated layout of出海 channels, it is becoming a key node for China's computing power going global. "Chinese Tokens" are setting out from the Loess Plateau, flowing towards the world.
The network is connected, computing power is "running." How can more enterprises "access and afford" it? Leveraging Qingyang's low-cost, highly reliable computing power foundation and targeting scenarios in the east like AI inference, model fine-tuning, and AI agent hosting, Qingyang is exploring service models billed by token, allowing SMEs to "pay for the computing power they use," significantly lowering the barrier to AI application. Currently, several internet companies have already migrated their inference services to Qingyang, with token call volume doubling month by month.
From "selling server cabinets" to "selling tokens," this represents a significant leap in the computing power value chain. From the Loess Plateau to users' fingertips, a seamless high-speed channel is rapidly taking shape. This path leads not only to the national computing power market but also to the vast blue ocean of the global intelligent economy.
Comments