How China’s New AI Super‑Unit Powers the Million‑Card Era

Nvidia CEO Jensen Huang predicted that China will win the AI race, emphasizing that the real bottleneck for massive AI compute lies not in chips but in supporting infrastructure such as network, power, storage and cooling.

Recent setbacks at OpenAI—pausing the Stargate UK project due to high energy costs and regulatory uncertainty, shelving a Norwegian data centre, and abandoning an expansion in Texas—illustrate how energy constraints can halt AI deployments.

Why a "Super‑Unit" Is Needed at the Million‑Card Scale

When AI clusters grow from tens of thousands of GPUs to hundreds of thousands or even a million cards, physical constraints move from the silicon to the surrounding systems. Stable, high‑capacity power, efficient heat removal, and ultra‑fast interconnects become decisive for real‑world compute throughput.

Yuanjing’s Ulanqab Star River Base

Located in Inner Mongolia’s Ulanqab—an area with stable geology, cool climate, abundant wind power and 67% green electricity—the Yuanjing Star River Base was recently commissioned as the world’s largest AI‑specific super‑unit.

The facility covers more than 120,000 m² (equivalent to 20 football fields), houses a 2 GW power capacity, and can run a million‑card AI cluster delivering petaflop‑level performance. It is described as the strongest token‑output AI data centre globally.

Concept of an AI Super‑Unit

Unlike traditional data centres that merely host servers, an AI super‑unit integrates compute, networking, power, storage and cooling into a single, highly coordinated system. As Yuanjing AIDC General Manager Zheng Zihao analogizes, it solves the logistics of a “large‑scale chip army”: providing a massive deployment base, ample power “food”, and a clear supply line.

AI Power System Design

To turn variable wind energy into stable compute power, Yuanjing employs a two‑step modelling approach:

Yuanjing Tianji Meteorological Large Model fuses satellite, radar, ground stations and device data, adding atmospheric dynamics and terrain constraints to forecast wind‑solar output across multiple time scales.

Yuanjing Tianhu Energy Large Model uses these forecasts together with real‑time device status to continuously optimise generation, storage and load balancing, reserving capacity when wind is strong and pre‑charging when it weakens.

The EnOS IoT operating system unifies wind turbines, storage units, power equipment and compute racks under a single control plane, creating a closed loop from perception to decision to act.

Power delivery adopts an SST + BESS 800 V DC architecture with solid‑state transformers that convert medium‑voltage AC directly to DC, reducing conversion losses and embedding battery storage within the power path for instantaneous support.

Cooling Integrated with Power

High‑density compute generates massive heat; any bottleneck directly reduces chip utilisation. Leveraging Ulanqab’s low‑temperature climate, the super‑unit combines wind‑cooling, liquid‑cooling and other high‑density solutions to minimise non‑compute energy consumption.

From Traditional AIDC to Energy‑Centric Delivery

Historically, data centres were built by telecom operators, specialised developers and internet giants. As AI workloads reach GW‑scale, energy acquisition, distribution, storage and cooling now dictate project timelines and efficiency, shifting AIDC focus toward energy infrastructure.

Yuanjing’s model assigns customers the responsibility of deploying and managing their own servers and chips, while Yuanjing supplies the entire power‑grid, green‑energy, storage and cooling infrastructure, effectively turning the data centre into a ready‑to‑run AI compute platform.

Replication Prospects

Yuanjing’s "Mission Gobi" aims to construct 5 GW of green AI compute capacity across global desert regions by 2030, using the same "energy‑rich zone + super‑unit + AI power system" blueprint.

The key question is whether this integrated model can be duplicated in other renewable‑rich locations, transforming the "million‑card era" into a new AI infrastructure paradigm where compute follows energy, and system‑level efficiency unlocks the value of every chip and every kilowatt‑hour.

In this view, the AI super‑unit is not the end of the compute race but the beginning of a shift from mere hardware stacking to holistic system collaboration.