Ask HN: Why were OpenAI, Claude, and Grok simultaneously down?

Hacker News Discussion ↗

Multiple major AI services — including OpenAI’s ChatGPT/Codex, Anthropic’s Claude, and xAI’s Grok — experienced overlapping outages, returning 404s and connection errors worldwide. Commenters float causes ranging from Cloudflare or cloud-provider issues to cascading overload as users and tools automatically fail over from one provider to another, with later reports pointing to an outage at a shared xAI compute center. The incident is seen as a reminder of how centralized and interdependent today’s AI infrastructure has become, and as an argument for having local or multi-provider fallbacks.

Outage symptoms and scope

  • Users report ChatGPT / Codex returning 404s on chatgpt.com/backend-api/codex/responses, often with Cloudflare cf-ray IDs ending in airport codes.
  • Many note it’s global: reports from Brazil, India, Europe, US, etc.
  • Similar instability seen for Claude and Grok (failed new sessions, “at capacity” messages, disconnects), while some existing sessions kept working.
  • Some see problems with Gemini and various cloud providers via Downdetector, though others say those services worked fine for them.
  • Services gradually recover; some users note Codex/ChatGPT and Claude coming back region by region.

Hypothesized causes (all unconfirmed in thread)

  • Shared infra issues:
    • Cloudflare (due to cf-ray and simultaneous spikes), DNS/BGP, or a “load‑bearing” third‑party dependency.
    • Major cloud providers: AWS, Azure, Google Cloud, or backbone/fiber providers.
  • Cascading overload:
    • When one LLM fails, users and tools auto‑failover to others, potentially DDoS‑ing alternatives.
  • Vendor‑specific:
    • Coincidence with OpenAI’s Astra release.
    • SpaceX/xAI: a later post cites an xAI statement about an outage at a Memphis compute center impacting Grok and “compute partners,” but it’s unclear how directly this ties to all outages.
  • Some users suspect a Cloudflare HTTP/3 / R2 incident; others point to Cloudflare leadership denying any service disruption.
  • More speculative suggestions (state actors, rogue AI, supply‑chain bugs like “left-pad”) are treated as jokes or speculation.

Status pages and telemetry

  • OpenAI, Anthropic, and xAI status pages eventually acknowledge elevated error rates and partial outages.
  • Users share incident links and note that official status pages lag and often under‑report issues.
  • Downdetector shows spikes for multiple AI and cloud providers, but commenters stress it’s based on user reports and can be misleading.

User reactions and takeaways

  • Many express dependency on LLMs for coding and work; some frame the outage as a “natural experiment” in productivity loss.
  • Others emphasize local/self‑hosted models (Qwen, Ollama, home GPUs) as a hedge against centralized failures.
  • Overall consensus: exact root cause is unclear; likely mix of infra issues and cascading cross‑provider load.