China's interest in artificial intelligence has resulted in underutilized data centers and a decline in GPU rentals
Overbuilt data centers in China face low usage and GPU rental drop.
In the aftermath of ChatGPT's debut in late 2022, China's artificial intelligence sector witnessed an explosive surge in both interest and investment efforts, propelling a wave of hastily built data centers strung across the country. Xiao Li, a former real estate contractor who had pivoted to AI infrastructure in 2023, recounted an intensely competitive GPU trading landscape where Nvidia's H100 chips once fetched up to 200,000 yuan on the black market, despite US export restrictions. Yet, present narratives reflect a starker reality where Nvidia H100 server rentals have plummeted to 75,000 yuan per month amid construction oversupply and a lack of buyers, coupled with ongoing offloading of surplus GPUs by fractured project leaders aiming to mitigate financial strain.
Competitive pressures from entities like DeepSeek, whose open-source model rivaled ChatGPT in performance but commanded only a fraction of the cost, shifted the AI industry's focus from building large language models to deploying practical applications. Commenting on the changing landscape, Hancheng Cao from Emory University highlighted a poignant industry pivot from robustness in model development to adeptness in application. "The burning question shifted from 'Who can make the best large language model?' to 'Who can use them better?'" Cao observed. This initial exhilaration paved the way for stark revelations of many data centers being unsuited for low-latency inference tasks vital for real-time reasoning, ultimately dulling the allure of AI enterprises in locales with cheaper land and electricity yet insufficient infrastructure.
As GPU rental prices sharply dipped, stakeholders like Jimmy Goodrich of the RAND Corporation, argued that the downturn stemmed from ill-conceived projects by inexperienced investors and officials championing short-lived economic growth. This zeal, bolstered by China's political mechanisms favoring expedient economic ventures, encouraged rapid data center development as local leaders sought to lift their political profiles while simultaneously revitalizing economies stifled by the pandemic. However, these projects were often steered by figures lacking deep AI technical insights, leading to results that failed to meet evolving industry standards.
Nevertheless, the allure of AI advancement persists, catalyzed by a central government prioritizing self-sufficiency in AI technologies and major corporations such as Alibaba and ByteDance committing to cloud computing and AI infrastructure investments. Despite prevailing challenges, a particular fondness for Nvidia's forthcoming H20 chip, tailored for the Chinese market, suggests a continual demand underpinned by long-term potential. Yet, for individuals like Fang Cunbao, a data center project manager, the market volatility has necessitated a career recalibration. "The market is too chaotic. The early adopters profited, but now it's just people chasing policy loopholes," Fang shared, signaling his pivot toward AI education – a reflection of the recalibrated ecosystem echoing broad industry quandaries.
The broader context encapsulates a dichotomy where groups are seemingly at an inflection point; a bureaucratic resolve aligns expansion with burgeoning AI capabilities even as other entities and businesses reckon with the practicality and consequences of unchecked infrastructure development. The anticipations versus reality present a critical juncture for AI's footprint in China, elucidating lessons in strategic coherence and adaptive capacity amid rapid technological migratory patterns and inversely proportional market dynamics.
Sources: MIT Technology Review, TechSpot, RAND Corporation