DeepSeek's 5-Hour Global Outage: Full Credit Compensation and the Coming Era of AI SLAs

At 8:11 a.m. Beijing time on October 15, AI assistants everywhere stopped responding. DeepSeek's status page turned red: the chat service, the V4 Pro API and the V4.1 Flash API were all affected. The global interruption lasted roughly five hours from onset to full recovery — the longest in DeepSeek's public-service history.

Anatomy of a five-hour outage

According to the official incident log, the fault was confirmed at 08:11 and moved through "investigating", "identified" and "fixing" states before gradual recovery in the afternoon. Third-party monitor StatusGator logged 86 reports within 24 hours and counting. Users in Luxembourg saw no server response, users in Poland got no replies, and one New York user's verdict was blunt: "Too slow."

Full credit: an industry-level precedent

After recovery, DeepSeek apologized and announced that every account's consumption over the previous seven days would be fully refunded as credits. "Full compensation by consumption" is standard practice in cloud SLA regimes, but executing it at scale for an LLM API is essentially a first. As the V4-series APIs take on more production workloads, availability promises are moving from word of mouth to contractual terms. The fact that V4.1 Flash had launched shortly before the incident also spawned the quip that "it crashed right after launch."

LLM availability is turning from a perk into a contract clause.

Two signals beyond the outage

First, no root cause has been disclosed, yet community chatter jumped from "is it my network?" to "is this pre-launch maintenance?" — guessing a new model at every outage is itself a measure of how closely watched the lab is. Second, on the very same day, an account named "DeepSeekHarness" quietly appeared on Bilibili and quickly passed ten thousand followers. Harness is DeepSeek's open-source agent runtime: it lets models call tools, edit files and run commands, and splits complex tasks across sub-agents. Its official presence is widely read as the prelude to promoting the agent stack to Chinese users.

For enterprises that depend on a single model API, the five-hour pause was an early "availability bill": multi-model routing, cross-vendor fallback and private deployment backups are no longer optional in architecture reviews. At NineZenith, model-layer decoupling and failover are foundational to the Zenith-Act platform — the stronger the vendor, the more exit routes your architecture should keep.

(Compiled from public reports by QbitAI, Observer.com, The Paper, Phoenix Technology, Sohu Technology, etc.)