China turns farmland into AI data hub with $37B investment

China has begun transforming farmland into a major tech hub, aiming to strengthen its role in artificial ￰0￱ a 760-acre island on the Yangtze River, vast rice fields in the city of Wuhu are being cleared for server ￰1￱ executive linked with a supplier for one of these projects described the effort as building the “Stargate of China,” referencing a $500 billion ￰2￱ center plan by Oracle, OpenAI, and ￰3￱ smaller in scale, it supports Beijing’s strategy to centralize scattered data ￰4￱ data centers will train LLMs while server farms near cities will handle inference In March, Beijing set out a plan to utilize existing data centers in remote areas to train ￰5￱ comparison, the newly built server farms are set up closer to big population ￰6￱ will handle “inference” with proximity to speed up apps for end ￰7￱ Fedasiuk, former state department adviser on China, said, “China is starting to triage scarce compute for maximum economic output.” One example is Wuhu’s “Data Island”, which hosts four AI data centers run by China Telecom, Huawei, China Mobile and China ￰8￱ there, the cluster is expected to serve cities in the Yangtze River Delta, including Shanghai, Nanjing Hangzhou, and ￰9￱ south, Guizhou will supply Guangzhou, while Qingyang in central Gansu will serve Chongqing and ￰10￱ to a city notice, 15 companies so far have put up data centers across Wuhu, with a combined investment of $37 ￰11￱ local government is offering subsidies to cover AI chip procurement ￰12￱ push for tighter coordination is also meant to soften China’s weaknesses against its global ￰13￱ States export restrictions have cut Chinese groups off from modern processors and systems made by ￰14￱ chipmakers, including Cambricon and Huawei , have faced challenges to fill the gap, in part because China’s manufacturing capacity is ￰15￱ has also restricted Samsung and TSMC from manufacturing advanced artificial intelligence chips for Chinese customers.

Meanwhile, ￰16￱ leaders are racing ahead with hardware orders. Google, Meta, and xAI are deploying 10 of 1000 of Nvidia’s newest ￰17￱ aims to link data centers to offset chip limitations Many Chinese artificial intelligence data centers rely on weaker chips or build advanced systems via the black ￰18￱ people familiar with the trade say a system of intermediaries has grown across China to source Nvidia GPUs restricted for export to the ￰19￱ is also trying to tap the already existing idle ￰20￱ AI boom from 2022 onward concentrated data centers in energy-rich provinces such as Inner Mongolia and ￰21￱ shortages of skilled staff and limited local demand left valuable processors sitting unused even as need spiked ￰22￱ chip purchases have been financed by the local government in many ￰23￱ are reluctant to let go of those assets because the equipment feeds into local ￰24￱ the servers is also costly and slow.

Hence, “a technical solution has to be ￰25￱ is connecting data centers,” said Edison Lee, an analyst at ￰26￱ has ordered the use of networking equipment from Huawei and China Telecom to link together processors at multiple sites and form unified computing pools. China’s telecom groups are using the same mix of switches, routers, transponders, and software to shift computing power from western regions to eastern demand ￰27￱ is also working on a fix for the efficiency ￰28￱ company is using its telecom and AI hardware expertise on a new networking approach called UB-Mesh, which it says can increase the training efficiency of LLMs over numerous computing clusters by assigning tasks more effectively over the ￰29￱ you're reading this, you’re already ￰30￱ there with our newsletter .