Claude crashed. Why did the global AI "fuse" out?

On March 2, 2026, at 19:49 Beijing time, Anthropic’s AI assistant Claude suddenly experienced widespread service outages worldwide. The claude.ai web portal, developer console, AI programming tool Claude Code, and mobile apps all turned red almost simultaneously. Thousands of users flooded Downdetector to report issues, with peak reports reaching several thousand. Users attempting to log in saw HTTP 500 and 529 error codes or a brief message: “Claude will return soon.”

For the millions of developers, content creators, and enterprise users who have deeply integrated Claude into their daily workflows, this outage felt more like a massive blackout.

On social media, some joked, “Now I can only write prompts, what do I do”; developers said they were suddenly cut off mid-task and had to switch to ChatGPT or Gemini for emergencies; others teased in groups: “Today, AI-native companies might as well go team-building.”

01 “Whack-a-Mole” Outage

Anthropic has not yet provided a detailed official explanation for the outage, but a series of events over the past week may shed some light.

On February 28, Anthropic lost its contract with the U.S. Pentagon for refusing to allow Claude to be used for large-scale domestic surveillance and fully autonomous weapons systems. President Trump immediately criticized Anthropic on social media as a “left-wing lunatic” and ordered all federal agencies to cease using Claude. OpenAI quickly stepped in, announcing a partnership with the Pentagon.

This caused a dramatic reversal among global users. A “QuitGPT” movement spread rapidly on Reddit, Instagram, and X.com. A Reddit post calling for the cancellation of ChatGPT received 30,000 likes; the Instagram account “quitGPT” gained over 78,000 followers in a short time.

According to Tom’s Guide, about 700,000 users shifted from ChatGPT to other platforms. Anthropic became the biggest beneficiary of this digital migration.

Official data from Anthropic shows that since January 2026, the number of free Claude users increased by over 60%, with daily new registrations tripling compared to November 2025, and paid subscriptions doubling within the year. Before the Super Bowl LX, Claude ranked 42nd on the US App Store; by February 28, it reached the top free app, pushing ChatGPT to second place.

This surge was overwhelming. Sensor Tower data indicates Claude was in rapid growth throughout February, but the last few days saw user influxes far exceeding Anthropic’s infrastructure capacity.

Media reports quote Anthropic as saying they have been dealing with “unprecedented demand” over the past week.

From the timeline on Anthropic’s official status page, the evolution of the outage shows a “whack-a-mole” pattern.

At 11:49 UTC (19:49 Beijing time), the team began investigating, initially focusing on login and logout issues on claude.ai.

At 12:21 UTC (20:21 Beijing time), Anthropic claimed core API functions were normal, with problems limited to the web interface.

At 13:37 UTC (21:37 Beijing time), the situation worsened, with some API methods also reporting errors.

Subsequently, anomalies appeared in Claude Opus 4.6 at 17:09 UTC, followed by Claude Haiku 4.5 at 17:56 UTC. The cycle of fixes, re-emergence, and re-fixes lasted for hours.

It wasn’t until around 23:47 UTC (07:47 Beijing time on March 3) that major services gradually recovered. However, Opus 4.6 experienced multiple short-lived elevated errors, including a period extending to about 21:16 UTC (March 3, 05:16 Beijing time).

Just a few hours later, at 03:15 UTC (11:15 Beijing time on March 3), a new wave of failures occurred, affecting Claude Code and Cowork. As of this report, the issue remains under investigation.

Regarding the cause, media reports suggest that a Middle Eastern AWS data center may have been attacked by an unidentified object, causing a fire and power outage, impacting AWS’s compute pool. Since Claude heavily relies on these compute nodes, it lost support.

Why might a geopolitical conflict in the Middle East trigger a large-scale outage in U.S. AI companies?

Currently, AI service supply chains are highly globalized with few “choke points”: If regional conflicts damage submarine cables in the Red Sea, Bab el-Mandeb Strait, or Suez Canal, or cause disruptions in cloud data centers, power facilities, cross-border backbone networks, or cable landing stations around the Persian Gulf/Arabian Peninsula, it could lead to increased network latency, routing anomalies, authentication, billing, and control access failures, as well as delays in cross-region replication and failover.

Large models’ inference and training depend heavily on bandwidth, low latency, and cloud control planes. Disruptions to these “underlying elements” can cause distributed cloud services to cascade into systemic outages affecting users globally.

Notably, on the day of Claude’s outage, xAI’s status page showed Grok (Web/iOS/Android) also experienced about 40 minutes of “temporarily unavailable” events around UTC 23:00. Whether there is a shared upstream or causal link remains unconfirmed.

If true, this suggests the outage was not just a front-end authentication issue but also involved physical vulnerabilities in the underlying cloud infrastructure.

In the cyber realm, large models with immense compute power are particularly fragile when faced with physical attacks.

02 Chain Reaction in the Downstream Ecosystem

The reason Claude’s outage drew such widespread attention is that AI has evolved from a chatbot into a critical node in an entire AI-native productivity chain.

First impacted are developers. Claude Code is one of the most relied-upon AI programming tools globally. Previous reports estimate its annual revenue at around $200 million. Boris Cherny, founder of Claude Code, revealed on a podcast that he hasn’t manually edited a line of code since November 2025.

When Claude Code becomes completely unavailable, media reports indicate that the developer community generally reverts to pre-generative AI habits—coding manually.

Professional developers are forced to switch mid-workflow to GitHub Copilot or ChatGPT’s coding features, which entails efficiency losses and context breaks. For companies deeply integrated with Claude API, the impact is even more direct.

Content creation sectors are similarly affected. Teams relying on Claude for copywriting, report generation, and data analysis had to pause work. customer service bots went silent, and support tickets piled up.

Deployflow’s analysis estimates that for a 25-person engineering team, a 4-hour outage at a rate of £90 per hour could mean over £9,000 in productivity loss, not counting downstream delays.

The deeper issue concerns trust. Analysis from ainvest indicates that repeated service interruptions are eroding user confidence in platform reliability, especially for developers and enterprises building on Claude. Continuous uptime is a fundamental requirement.

What makes Claude so indispensable to enterprises?

It’s the ongoing infrastructure built around Claude’s “Agent” foundation.

According to official data disclosed in July 2025, four months after Claude Code’s release, it attracted 115,000 developers, processed 195 million lines of code weekly, and had 3 million weekly downloads.

In January 2026, the newly launched Claude Cowork, a desktop agent capable of clicking, managing files, and executing tasks across software, was more aggressive: equipped with 11 industry-specific plugins covering legal, sales, finance, and more, it acts as a “digital employee” handling knowledge work.

At the core, Anthropic’s MCP protocol is becoming the de facto standard for connecting AI to external tools. Both OpenAI and Google have announced support, creating an ecosystem with over 500 commercial application connectors.

Claude is no longer just an API; it’s an “AI operating system” composed of the model (brain), Code/Cowork (execution), and MCP (connectivity).

Claude’s deep integration into developer and enterprise ecosystems has created systemic dependencies at the infrastructure level. Yet, the reliability of this infrastructure still falls short of expectations.

03 Fragility of AI Infrastructure

This outage is not an isolated incident. Forrester’s “2026 Predictions: Cloud Computing” report predicts at least two major, multi-day cloud service disruptions in 2026 triggered by upgrades to AI data centers. The logic is that AWS, Azure, and Google Cloud are shifting investment from traditional x86 and ARM environments to GPU-centric AI data centers, making aging infrastructure more vulnerable amid increasing complexity.

Forrester also forecasts that at least 15% of enterprises will adopt private AI deployments on private clouds in 2026 to mitigate rising costs, data lock-in, and operational risks.

Warnings emerged in 2025: AWS experienced over 17 million Downdetector reports and outages lasting more than 15 hours, affecting services like Netflix and Snapchat. In November 2025, Cloudflare outages caused many sites, including Claude, Shopify, and X, to go down. In December, Amazon’s in-house AI programming tool Kiro autonomously decided to delete and rebuild an environment while repairing a customer-facing system, causing a 13-hour AWS Cost Explorer outage. Single points of failure and chain reactions are becoming the most dangerous systemic risks in the AI era.

The industry’s lessons are multi-dimensional. First, multi-model redundancy is no longer optional but essential. Companies with pre-deployed multi-LLM fault-tolerance—such as switching to Gemini or GPT models when Claude is unavailable—are less affected. Future AI infrastructure must incorporate model redundancy as a core design, similar to multi-cloud strategies.

Second, observability is critical. Deployflow’s analysis shows that token latency tracking and error rate alerts are early warning signals for service collapse, enabling teams to switch before total AI access is lost.

Third, physical infrastructure security is severely underestimated. If the causal chain of the Middle Eastern data center attack holds, AI infrastructure faces threats beyond software—geopolitical risks, physical attacks, and natural disasters.

Forrester also highlights a trend: “Neoclouds” like CoreWeave, Lambda, and Nebius, focusing on high-performance GPUs, are projected to generate $20 billion in revenue in 2026, challenging the dominance of hyperscale cloud providers in generative AI.

These providers are building GPU-first architectures from scratch rather than retrofitting old data centers, potentially offering new solutions for AI infrastructure resilience.

For enterprises and platforms building AI infrastructure, the clear lesson is: don’t put all your eggs in one basket, and don’t assume any single provider can guarantee 100% uptime.

Before AI becomes truly “water, electricity, and coal,” its infrastructure must reach the same level of reliability. Otherwise, every outage will be a stress test for the entire ecosystem.

As of this writing, Claude services still experience intermittent issues, and Anthropic continues investigations.

Source: Tencent Technology

Risk Disclaimer and Terms

The market carries risks; investments should be cautious. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should evaluate whether any opinions, viewpoints, or conclusions herein are suitable for their circumstances. Invest accordingly at their own risk.