Five-hour DeepSeek outage raises doubts about China's computing power

Chinese AI service DeepSeek was hit by an outage lasting almost five hours on Thursday, Oct. 15. According to Caixin Global, the problems began at 8:11 a.m. China time. Chat was completely down from 9:08 a.m., and all services were restored by 1:04 p.m.

The outage hit both chat and the API (the access developers and companies use to connect to the models) for DeepSeek-V4-Pro and DeepSeek-V4.1-Flash. The website and the app were also down.

The cause has not been disclosed

In its notice, DeepSeek said it had found the cause and restored the services, but the company did not say what the cause was. According to Chinese media, however, the company's chatbot told users in the chat after service was restored that the servers "could not handle so many users." That was the chatbot's answer, not an official explanation.

According to BigGo Finance, experts speculated that the failure stemmed from computing power scheduling rather than coding errors, because the fix took so long. Liu Xingliang of the DCCI Internet Research Institute said that this kind of irregularity is typically caused by traffic spikes, infrastructure failures or configuration problems.

DeepSeek gave users a refund of the entire week's API usage as account credit. The amount depended on usage and was up to 3,000 yuan (about $449) for heavy users.

A pattern of outages

According to Caixin, it was DeepSeek's fifth service disruption in October. There was also a long outage in March. Sources disagree on its length: two outlets report more than seven hours, while Bloomingbit says about 12 hours. Sources also disagree on the number of major outages in 2026: BigGo counts at least 10, while other outlets report at least 18.

Chip shortage in the background

Bloomingbit, citing other media outlets, reports that China's demand for AI chips this year is well above 5 million units, while supply is expected to reach 3.5 million at most. That leaves a shortfall of 1.5 million units. The figures have not been checked against a primary report.

A leaked and unconfirmed transcript of an investor meeting on May 20 quotes DeepSeek founder Liang Wenfeng as saying that Huawei has provided capacity for about 16,000 Ascend 950 cards, while training the very largest models would require about 200,000. The transcript was produced with speech recognition and may contain errors. DeepSeek has not confirmed it.

According to Caixin, DeepSeek has previously responded to capacity constraints by adjusting its API prices and working with Huawei. The company is also planning a 1-gigawatt data center expansion in Ulanqab.

For ordinary users, this means that popular Chinese AI services can be unstable when demand is high and computing power is scarce.