Claude crashed: Why did the global AI industry "break down"?

Author: Xiaojing, Tencent Technology Reporter

At 7:49 PM Beijing time on March 2, 2026, Anthropic's AI assistant Claude suddenly experienced a widespread service outage globally. The claude.ai webpage, developer console, AI programming tool Claude Code, and mobile application almost simultaneously displayed red warnings, with thousands of users flooding Downdetector to report the fault, reaching thousands of reports at peak times.

Figure: Claude fault report spike curve in Downdetector

From the timeline on the Anthropic official status page, the evolution of the fault exhibits a "whack-a-mole" pattern.

At UTC 11:49 (Beijing time 19:49), the team began investigating the issue, initially determining that the fault was concentrated in the login and logout paths of claude.ai.

At 12:21 UTC (20:21 Beijing time), Anthropic announced that its core APIs were functioning normally, and the problem was limited to the web interface. At 13:37 UTC (21:37 Beijing time), the situation worsened, and some API methods began to report errors. Subsequently, the Claude Opus 4.6 model experienced an anomaly at 17:09 UTC, followed by Claude Haiku 4.5 at 17:56 UTC. The cycle of fixing, recurring, and fixing continued for several hours.

Main services were gradually restored around 15:47 UTC (23:47 Beijing time). Subsequently, Opus 4.6 experienced several short periods of elevated errors (including one that lasted until approximately 21:16 UTC, 5:16 Beijing time on March 3rd). Just a few hours later, at 3:15 AM UTC on March 3 (11:15 AM Beijing time), a new round of outages occurred, affecting Claude Code and Cowork. As of press time, the issue is still under investigation. Regarding the cause of the outage, media reports also indicate that AWS data centers in the Middle East were suspected of being attacked by an "unidentified object," causing a fire and power outage. This impacted the AWS computing pool, and the Claude model, which heavily relies on these computing nodes, lost support.

Image: Incident Notice from AWS Official Status Page: A power outage (a data center was struck by a foreign object, causing a spark/fire, and the fire department cut off the power) has resulted in increased error rates and service disruptions for several services and EC2 network-related APIs in an availability zone in the UAE (UAE) region (ME-CENTRAL-1). Why could geopolitical conflicts in the Middle East potentially cause large-scale outages for US AI companies? Currently, the critical links in AI services are highly globalized and have a few "choke points": If geopolitical conflicts lead to damage to submarine fiber optic cables in the Red Sea-Mandeb Strait-Suez Canal area, restricting regional networks, or disruptions and congestion in cloud data centers and power facilities, cross-border backbone networks, and submarine cable landing stations around the Persian Gulf/Arabian Peninsula, it could trigger surges in cross-regional network latency, abnormal routing convergence, authentication/billing/control plane access failures, and obstacles to cross-regional replication and failover. Furthermore, large-scale model inference and training are even more dependent on bandwidth, low latency, and the cloud control plane. Once these "underlying elements" are disturbed, the originally distributed cloud services can amplify the failures in a chain reaction, resulting in systemic outages affecting global users.

Image: The "convergence point" of submarine optical cables in the Persian Gulf: TeleGeography's submarine optical cable map shows that the UAE-Iran cable between Fujairah, UAE, and Jask, Iran, is just one segment. A dense network of optical cables in the surrounding area connects the Middle East to the backbone networks of Europe and Asia; the vulnerability of global data links would be amplified should geopolitical conflicts or infrastructure disruptions occur in the region.

More noteworthy is that Claude...