DeepSeek Went Down for Five Hours. Its Refund Plan Is the Story
On the morning of October 15, users of DeepSeek’s API began noticing something was wrong. Between 8:11 and 13:04 Beijing time, the company’s service system failed, and for roughly five hours developers building on DeepSeek’s models could not reach the API at all. DeepSeek, the Hangzhou-based AI lab whose low-cost models turned it into one of China’s best-known artificial intelligence companies, confirmed the outage in a service notice and apologized for the disruption.
The more interesting part came afterward. Rather than issuing a flat credit or a free month, DeepSeek said it would calculate each account’s spending over the seven days from October 8 to October 14 and return that full amount as bonus credit. The money lands automatically in user accounts, no claim form required. In other words, a heavy user who burned through a large API bill in the week before the crash gets a proportionally larger refund than someone who barely touched the service.
The company also said it had launched a full review of service stability and was stripping out single points of dependency in its infrastructure, promising to prevent similar failures. That language is familiar to anyone who follows cloud platforms, but it matters more for an AI provider, where a five-hour gap can stall a startup’s entire product.
What makes this a talking point is the design of the compensation itself. API pricing is the quiet battleground of the AI industry, and Chinese model providers have spent the past two years competing aggressively on cost per token. A refund tied to recent consumption effectively treats reliability as part of the price: if the service is down, the meter runs backward. Some users on Chinese discussion platforms praised the simplicity, noting that it rewards the customers who were most exposed to the outage. Others asked why the formula looked backward at the prior week rather than the hours actually lost.
There is a practical angle for global developers too. Many teams outside China now route some traffic through Chinese models because the per-token prices are low, and an outage is the moment that calculation gets tested. A five-hour window with no service is annoying for a hobby project and potentially serious for a business running customer-facing features on top of it. DeepSeek’s answer, essentially, is that the bill shrinks when the service fails.
For a company that built its reputation on doing more with less, the episode is a reminder that cheap intelligence still has to be available intelligence. Model quality gets the headlines, but uptime and the fine print around it are what turn a clever demo into infrastructure. DeepSeek has now put a number on that promise. The next outage, if there is one, will be judged against it.