No internet, no intelligence. Four times the network went, and the assistant went with it.

The smartest assistant in the world is a blank screen when the request cannot reach it. Between your desk and a cloud model sit your internet provider, a content delivery network, the vendor’s cloud region and the grid that powers all of them. Each link failed, publicly, in the last twelve months.

18 November 2025: the network in the middle

Cloudflare carries a large share of the world’s web traffic. On 18 November 2025 a routine configuration change crashed a service underpinning its bot mitigation. TechCrunch reported that ChatGPT, Claude, Spotify, X and other major sites went down for roughly two hours. Cloudflare’s chief technology officer Dane Knecht explained: “A latent bug in a service underpinning our bot mitigation capability started to crash after a routine configuration change we made. That cascaded into a broad degradation to our network and other services.”

Note what failed. Not OpenAI, not Anthropic. A third company neither you nor the vendor chose for that morning, doing routine work. On 5 December 2025 it happened again: Cloudflare’s own post-mortem records a 25-minute outage from 08:47 UTC affecting “approximately 28% of all HTTP traffic served by Cloudflare”, triggered by a change made while mitigating a software vulnerability. In its words, “not caused, directly or indirectly, by a cyber attack”.

20 October 2025: the region

Amazon Web Services’ own summary of the US-EAST-1 disruption states that “the event started at 11:48 PM PDT on October 19 and ended at 2:20 PM PDT on October 20”, about fourteen and a half hours, “triggered by a latent defect within the service’s automated DNS management system”. Engadget listed what stopped: Snapchat, Reddit, Alexa, Venmo, Bank of America, Zoom, Lyft and dozens more.

The data was safe the whole time. As one academic told Engadget, “Amazon had the data safely stored, but nobody else could find it for several hours.” A model in that region would have been in the same position: intact, and unreachable.

28 April 2025: the grid

At 12:33 CEST on 28 April 2025 mainland Spain and Portugal lost power. ENTSO-E, the European association of grid operators, concluded that the blackout stemmed from “a combination of many interacting factors, including oscillations, gaps in voltage and reactive power control”, leading to “fast increases of voltage and cascading generation disconnections in Spain”. Hospitals ran on generators. Anything that needed a network outside the building ran on nothing.

Closer to home, on 31 January 2026 parts of Ukraine and Moldova, including both capitals, lost power after a grid malfunction. Radio Free Europe reported Ukrainian officials attributing it to “technical problems”. Romania was the exporting end of that line. A grid fault three borders away can be your afternoon.

Every link in the chain was run by competent people doing routine work. That is the point. Routine work fails, and the chain has four links you do not own.

What it means for a building that has to keep working

• A hospital’s clinical letters and a bank’s customer-facing staff do not get to wait for a status page. A tool that is part of the workflow has to be reachable from inside the building when the outside is not.
• A generator keeps the lights on. It does not restore an undersea cable or a Virginia data centre. On-premise AI is the only kind a generator helps.
• None of this is a legal argument. It is an availability argument, and DORA’s exit plan and business-continuity rules turn it into paperwork you have to sign.