1. Chinese banks are finding a new lending opportunity in AI data centers as the property sector declines; projects are sprouting in windy, arid regions like Ulanqab and Zhongwei, attracting banks, insurers, private equity and local governments [para. 1][para. 2]. One bank branch official estimated computing centers could represent a third of the branch’s planned lending in 2027, reflecting AI’s computing demand and a lack of attractive alternatives [para. 3][para. 4].
2. The business is like specialized commercial real estate but with huge electricity consumption; AI facilities require dense GPU clusters and specialized cooling, and demand has surged since late 2022 [para. 8][para. 11][para. 12]. Approvals for large power loads are hard to get, especially outside the eight national computing hubs created in 2022 under “East Data, West Computing,” which steer workloads to western regions with abundant renewable energy [para. 13][para. 14][para. 15]. Location decisions favor low power prices and policy support; electricity is a controllable cost, with power prices around 0.32–0.35 yuan/kWh in Ulanqab potentially saving 5 billion yuan per gigawatt-year compared with alternative sites [para. 16][para. 17][para. 18][para. 20]. Scarce approvals make built projects more valuable [para. 21].
3. The telecom operators are expanding aggressively, changing the industry balance among independent providers, internet/cloud firms and the state-owned carriers [para. 22][para. 24]. In Q3 2026, the three operators launched more than 225 data center tenders worth about 36.6 billion yuan, 90% aimed at national hubs; China Mobile alone began three gigawatt-scale campuses and planned two more in 2027 [para. 26][para. 27]. Their advantages in approvals, financing costs and subsidies let them undercut independent operators, though some see subsidies as helping deploy domestic chips [para. 29][para. 30][para. 31].
4. Internet and cloud companies’ spending is rising fast: Tencent’s first-half 2026 capex was 84.7 billion yuan, and Alibaba’s Q2 capex rose 75% year-on-year to 67.7 billion yuan [para. 33][para. 35]. Financing has followed, with bank loans, syndicated loans and finance leases for servers; a China Construction Bank-led consortium provided a 4.68 billion yuan loan for a Range Intelligent data center [para. 37][para. 39][para. 42]. Banks have introduced “token loans” using AI activity metrics, and China Development Bank supported a Chindata project with a 136 million yuan policy loan [para. 43][para. 44][para. 46].
5. Investors are mostly optimistic: AI applications are just beginning, and IDC projects global AI investment of $815.9 billion in 2028, with China exceeding $100 billion; daily token usage in China jumped from 100 billion in early 2024 to over 140 trillion in March 2026 [para. 47][para. 48][para. 49][para. 50]. Operators expect robots and virtual agents to add demand, and models with roughly 5 trillion parameters will require much larger clusters [para. 52][para. 53].
6. But the central risk is overcapacity. The big question remains whether AI can sustain expansion or whether China is setting the stage for another infrastructure glut [para. 7][para. 55]. One investor said slower commercialization or tech shifts could leave facilities underused; China already suffered excess conventional capacity in 2020–2023, and the probability of a repeat is “not small” [para. 56][para. 57][para. 58]. Physical infrastructure offers steadier cash flows and clearer exits—GDS and Range listed data center REITs in 2025—while GPU-based businesses face enormous capital needs and depreciation uncertainty, with a gigawatt project possibly costing 30 billion yuan for infrastructure and over 100 billion for GPUs [para. 59][para. 60][para. 61][para. 62]. Some investors remain cautious about tenant concentration and long-term viability [para. 64].
AI generated, for reference only
Sign in with Google
Sign in with Facebook
Sign in with 财新