June 23, 2026 · 8 min read

What Happens When the AI Goes Dark

At 9:32 yesterday morning, the internet blinked. A single misconfiguration at a shared infrastructure provider took down X, Microsoft Teams, Zoom, Robinhood, Reddit, and Discord at the same time. Trades went unmade. Meetings collapsed. By noon most of it was back, and most people moved on. You should not move on.

Because June 22 was not an isolated bad morning. Earlier this month, Microsoft Copilot went dark for five hours. The day after, Claude had a global outage that took down the model, the API, and the coding tools built on it. Ookla now counts millions of AI outage reports, and finds that agentic systems fail more than twice as often as the simpler tools they replaced. Forrester is openly predicting multi-day hyperscaler outages this year as the AI buildout pulls money away from the boring infrastructure that used to keep the lights on.

Here is the question that should be on your desk this morning. You have spent two years wiring your operations around AI that is always available. What is your plan for the hours it is not?

You Optimized for a System That Isn't Always There

This is what efficient execution does to you, and it is worth naming plainly. You found a capability that worked, you built your process tight around it, and you removed the slack. The manual fallback nobody used anymore got retired. The people who knew the old way moved on. The redundancy looked like waste, so it got cut. Quarter after quarter, you optimized, and every round of optimization made you a little more dependent on the assumption that the assistant would always answer.

That assumption is now wrong several times a month. And the more precisely you tuned your operation around the AI, the harder it slams into a wall when the AI is down. The most efficient companies are not the most resilient ones. Very often they are the most fragile, because they spent the buffer that resilience is made of.

Real execution in the age of autonomous systems is not just about how fast you move when everything works. It is about whether you keep moving when a piece of it fails. That tension, precision and resilience held together instead of traded against each other, is the whole subject of Geometric Precision: Strategic Execution in the Age of Autonomous Systems.

Concentration Is the Risk Nobody Priced

Look closely at yesterday and you see the real exposure. It was not that one company failed. It was that dozens of unrelated services all failed together, because they quietly depended on the same piece of shared plumbing underneath. Nobody chose that concentration. It accumulated, one sensible vendor decision at a time, until a single point of failure sat under half the things you use.

Your AI stack is concentrating the same way, and most leaders have no map of it. How many of your critical workflows route back to one model provider? When that provider has a bad day, how many of your teams stop at the same instant? With downtime now costing serious operations thousands of dollars a minute, this is not an IT footnote. It is a strategic dependency hiding in plain sight on your own org chart, and the first time you map it honestly is usually the first time you realize how much you bet on one vendor staying up.

Build the Fallback Before You Need It

I am not telling you to rip out the AI or slow down. That would be cowardice dressed as caution, and your competitors would eat you. I am telling you to do the unglamorous work that separates an operation that survives an outage from one that simply stops: know your single points of failure, keep a path that does not route through them, and make sure at least a few humans still remember how to run the critical process by hand.

The companies that handled yesterday well were not the ones with no AI. They were the ones who had asked, in advance and on a calm day, what they would do when it went dark. Resilience is a decision you make before the outage, never during it. After the alarm goes off, all you get to do is find out what you decided months ago.