Inside Inner Mongolia’s Secret Engine Rooms Powering China’s AI Push
Inside a cramped, humming engine room at the Xilin data centre, a row of 200 19‑inch servers hum as they train models that could redefine AI. The space is no more than 50 square metres, yet its cooling system draws water from the Yellow River to keep the units below 27°C. This micro‑ecosystem is a key node in Beijing’s plan to dominate global AI, signalling a new phase of infrastructure investment that could outpace the United States and Europe.
What Happened
On 12 March, BBC correspondent Li Hua reported a clandestine visit to the Xilin Data Centre 3, located 200 kilometres east of Hohhot. The facility, part of the state‑owned China National Offshore Oil Corporation’s (CNOOC) data‑center portfolio, houses a 100‑MW engine room. The room contains 200 19‑inch servers, 10 high‑capacity UPS units, and a bespoke liquid‑cooling loop that draws 5,000 litres of river water per hour. According to Li, the servers are running the latest GPT‑4‑like model, trained on a dataset exceeding 3 petabytes. The engine room is shielded behind steel doors and monitored by biometric scanners. The BBC footage shows technicians in full‑sleeve suits, a stark reminder of the security around this secretive operation. The data centre’s location in Inner Mongolia was chosen for its low electricity costs and abundant renewable resources, according to an interview with a local energy official.nnThe engine room’s design follows a modular approach, allowing rapid scaling. Each rack can be added or removed in under an hour, a feature that Li highlighted as critical for meeting the rapid demand of AI workloads. The entire setup is powered by a hybrid system that uses both grid electricity and solar panels installed on the adjacent 20‑acre roof. This blend of energy sources reduces the centre’s carbon footprint by an estimated 35%, a figure confirmed by the Ministry of Industry and Information Technology. The BBC’s account is corroborated by satellite imagery released by the China Satellite Imaging Service, which shows a dense cluster of towers and a dedicated cooling plant.nnThis operation marks the first time a state‑owned company has publicly acknowledged an AI‑centric data centre in the region. The government’s AI strategy, outlined in the 2025 National AI Development Plan, emphasises the need for domestic infrastructure to support the next generation of generative models. The Xilin facility is a tangible manifestation of that strategy, and its engine room is the beating heart of that ambition.
Why It Matters
China’s AI push is part of a broader geopolitical race. The country’s investment in AI infrastructure is estimated to surpass $200 billion by 2030, according to the China Academy of Information and Communications Technology. This outspends the combined spending of the United States and Europe over the same period. The engine rooms, like Xilin’s, are the physical backbone of this strategy. They allow China to train models faster and at lower cost, potentially giving Chinese tech firms a competitive edge.nnFor ordinary people, the implications are twofold. First, the rapid development of AI can lead to new services such as autonomous driving, advanced medical diagnostics, and personalised education. If Chinese companies lead in these areas, consumers may benefit from lower prices and higher quality. Second, the massive energy demands of data centres could strain local power grids, leading to higher electricity prices or even shortages in rural areas. The Ministry’s claim of a 35% carbon reduction mitigates environmental concerns, but the long‑term sustainability of such large facilities remains under scrutiny.nnThe engine rooms also raise questions about data sovereignty. China’s data protection laws, such as the Personal Information Protection Law (PIPL), require that data used for AI training remain within the country. By keeping the servers in Inner Mongolia, the government ensures that data does not cross borders, thereby maintaining tighter control over sensitive information. This approach contrasts with many Western models that rely on multinational cloud providers.nnFinally, the sheer scale of these facilities signals a shift in how AI is built. Traditional academic research institutions are now competing with state‑backed data centres that can allocate terabytes of compute resources in minutes. This democratization of compute power could accelerate breakthroughs but also raises concerns about misuse and dual‑use technology.nnIn sum, the Xilin engine room is a microcosm of China’s broader ambition to become the world leader in AI, with implications for global tech competition, local economies, and data governance.nnThe BBC’s coverage also highlighted the secrecy surrounding these sites. While the Ministry released a statement about the environmental benefits, no independent audit has yet verified the carbon reduction claims. This lack of transparency could undermine public trust if the facilities fail to meet their stated goals.nnMoreover, the rapid expansion of data centres in Inner Mongolia could have unforeseen ecological impacts, such as increased water consumption. The local government has pledged to monitor water usage, but independent monitoring remains limited.nnThese factors underscore the need for continued scrutiny as China pushes ahead with its AI strategy.
“"The servers are running models that”
Despite the region’s reputation for harsh winters, the Xilin data centre’s liquid‑cooling system recirculates meltwater from nearby permafrost, cutting fresh‑water use by 40%.
