DeepSeek Suffers Nearly 5-Hour Outage as China’s AI Boom Fuels Chip Shortage

China Faces a Shortfall of More Than 1.5 Million AI Chips

Even Nvidia Gaming GPUs Are Being Snapped Up

China’s artificial intelligence industry is growing rapidly, but the computing infrastructure needed to support it is failing to keep up with demand. Operational strains are also surfacing, with DeepSeek, one of the country’s leading AI companies, suffering a service outage that lasted nearly five hours.

China’s government is accelerating its push for semiconductor self-sufficiency, but domestic supplies of AI chips remain far below demand. Chinese technology companies are buying large volumes of Nvidia’s gaming graphics processing units, or GPUs, to secure scarce computing capacity. As the race to develop AI models expands into a scramble for data centers and semiconductors, computing infrastructure is emerging as a key factor that will determine how fast China’s AI industry can grow.

Computing Infrastructure Bottlenecks China’s AI Ambitions

According to the South China Morning Post and Chinese media reports on Oct. 16, DeepSeek’s chatbot and application programming interface, or API, services were disrupted for 4 hours and 50 minutes the previous morning. APIs for its V4 Pro and V4.1 Flash models were also suspended. Users around the world received error messages saying the server was temporarily unavailable. DeepSeek said it would refund all API charges incurred from Oct. 8 through Oct. 14 in the form of service credits.

The latest disruption was not DeepSeek’s first major outage. Its services were also down for about 12 hours in March. The company said it would eliminate single points of dependency in its system and improve stability. It did not disclose the specific cause of the latest outage. DeepSeek is also pursuing plans to build an AI data center in China’s Inner Mongolia Autonomous Region with about 1 gigawatt of capacity.

The biggest challenge facing China’s AI sector is a shortage of semiconductor supply, industry participants say. The Financial Times reported that China’s demand for AI semiconductors will exceed 5 million units this year. Supply, by contrast, is expected to total no more than 3.5 million units, including about 1 million produced overseas. Demand is forecast to top 10 million units next year.

China is ramping up chip production quickly, but demand for the computing power needed to develop AI models and run services is rising even faster.

There Aren’t Enough Chips to Run AI

Competition among Chinese technology companies is also driving demand higher. Combined capital expenditure this year by Alibaba, ByteDance and Tencent is forecast to reach about 700 billion yuan, or roughly $97.4 billion, more than double last year’s level.

Morgan Stanley projects China’s total spending on computing infrastructure will reach $193 billion this year. That amounts to about 21% of the US level. As AI models become more advanced and user numbers grow, massive computing resources are increasingly needed not only for training but also for inference, the process of delivering services to users.

China’s government and companies are responding by expanding domestic chip production. Huawei is on track to meet its target of producing 1.6 million AI chips this year. The chip units of Alibaba and Baidu are set to produce a combined 1 million units, while Cambricon is expected to make about 300,000. Alibaba has unveiled its self-developed Zhenwu V900 AI chip, and Huawei plans to launch the Ascend 960DT for AI training in the first quarter of next year. ByteDance also plans to test a prototype of its own chip later this year and move toward mass production in 2027.

AI companies are also stepping up data-center construction. DeepSeek and Zhipu AI are each building proprietary data centers using more than 100,000 Huawei chips. Huawei is also developing technology that links multiple semiconductors through high-speed networks to offset the performance limits of individual chips. Over the longer term, it aims to build a data center that can connect as many as 1 million chips.

For now, though, Chinese companies are turning to Nvidia to ease immediate supply shortages. The Financial Times reported that Chinese technology companies bought more than 1 million of Nvidia’s gaming and workstation GPUs this year. With US export restrictions making it harder to secure advanced AI chips, companies are using lower-performance general-purpose GPUs instead. Those products trail Huawei’s Ascend 910C in most areas, but they can still handle early-stage inference tasks such as reading and processing users’ queries.

Overseas data centers are also serving as an alternative. Tencent signed a deal with Oracle to lease 100,000 GPUs at data centers in Southeast Asia. ByteDance and Alibaba are also pursuing contracts to rent AI semiconductors from overseas data centers for extended periods.

Performance is not the only reason Chinese companies cannot easily replace Nvidia. Nvidia’s CUDA software platform is technology Chinese AI developers have used for years. Huawei’s in-house CANN platform is regarded as relatively difficult to use.

A Beijing-based industry official said that even if China makes progress in localizing semiconductor production, the supply shortage will be difficult to resolve anytime soon.

Kim Eun-jung, Beijing correspondent, Korea Economic Daily kej@hankyung.com