By 2026, China’s computing power market has entered a period of sharp divergence. While top-tier firms enjoy record-breaking profits and order books filled through 2028, smaller regional players are drowning in debt with utilization rates hovering below 70 percent. This landscape has prompted a sharp warning from the leadership of one of China’s most critical tech hubs: the greatest danger to the industry is not a lack of hardware, but the transformation of GPUs into a speculative real estate-style bubble.
Hu Rong, chairwoman of the Wuhu Big Data Construction Investment and Operation Co., is the chief architect of the computing strategy in Wuhu, a key node in the national 'East-to-West Computing' project. Wuhu carries over 70 percent of Anhui Province’s intelligent computing tasks, hosting giants like Huawei and ByteDance. Yet, Hu is pivoting the city’s state-owned platform away from building 'machine room' infrastructure and toward a more sophisticated model of asset-light orchestration.
The most alarming signal, according to Hu, is 'heavy-asset financialization.' In a trend reminiscent of the property market’s boom years, some enterprises are using high-end GPUs as collateral for high-leverage financing, effectively 'flipping' computing power rather than using it for innovation. This 'computing real estate' model detaches asset valuations from actual commercial utility, creating a fragile system where any delay in technical delivery or AI monetization could trigger a systemic collapse.
Technological bottlenecks also persist in the push for cross-regional computing. Beyond the physical constraints of network latency, the industry suffers from a lack of unified standards. Different clusters and vendors utilize varying billing methods and APIs, forcing enterprises to engage in costly adaptation work. These frictions make the dream of computing power flowing as seamlessly as water and electricity a difficult goal to reach in the short term.
On the front of domestic self-reliance, the gap with global leaders like NVIDIA remains centered on the software ecosystem rather than just raw silicon. While Chinese-made chips are becoming competitive in 'inference' tasks—the process of running a trained model—they still lag in ultra-large-scale 'training' scenarios. Hu estimates a three-year horizon for domestic chips to find their footing in vertical industries, while a full five-year cycle will be required to challenge NVIDIA’s dominance in high-end model training.
To navigate these risks, Wuhu is prioritizing a 'state builds the stage, private firms play the lead' approach. By focusing on a provincial-level scheduling platform, the state-owned enterprise aims to lower the barrier to entry for AI firms while avoiding the trap of 'building for the sake of building.' This strategy seeks to cultivate a genuine industrial ecosystem where value is created through service and tokenization rather than simple hardware arbitrage.