ChatGPT, Claude and Grok all went down yesterday morning. Same window, about two hours, tens of thousands of people staring at error screens. Turned out three supposedly rival AI companies were all renting the same patch of Microsoft Azure, so when Azure's East US region wobbled, they wobbled together. Gemini stayed up. It runs on Google's cloud instead (the more you know, hey).
And I already know what's coming. By Monday the feed will be wall to wall with the same post. Concentration risk. Single point of failure. You need a multi-provider strategy, a fallback model, redundancy across clouds. A lot of serious advice telling you to go and rebuild your stack.
Here's my problem with all of it. Almost none of you need to do any of that.
Let me be blunt, because the pile-on already annoys me. Running two AI providers in parallel, ready to fail over the second one blinks, is enterprise advice being sold to businesses that aren't enterprises. It's expensive. It's a real engineering project. Two integrations to build, two to maintain, two sets of quirks, a switching layer holding them together, all so you can survive an outage that turns up maybe once a year and lasts an afternoon. Run the numbers on that. You'd spend more on the insurance than the thing it insures against will ever cost you.
Because what actually happened to most businesses yesterday? A copywriter couldn't draft product descriptions for a bit. A support bot threw errors for ninety minutes. An analyst waited until after lunch to run the report they were going to run anyway. Annoying. Not fatal. Nobody lost a customer because a landing page got written at 3pm instead of 1pm.
The thing the redundancy crowd skips is that near-perfect uptime is hard and costly, and you already lean on it everywhere without a backup. Your electricity has no fallback. Your office has one internet line. Stripe goes down and your checkout goes with it, and you don't keep a spare payment processor idling on standby. You accept all of that, and you're right to, because the cost of the odd outage is smaller than the cost of engineering it away. AI is no different. Pretending it's a special case is how consultants get paid.
So no, I don't think the lesson is diversify. I think it's smaller and less sellable than that. Go and find out which of your workflows genuinely can't survive two hours off. There'll be fewer than you fear. For most, the honest answer is "cope, it'll be back before the meeting." For the rare one that can't, the customer-facing bot that's the only way people can reach you, say, then build the fallback for that one thing and leave everything else alone.
The part actually worth keeping from yesterday has nothing to do with your setup. It's that three companies who sell themselves as competitors fell over in unison because they share plumbing none of them advertise. The independence was partly a story we told ourselves. Worth knowing, not so you panic-buy a second provider, but so "we use two AI vendors" stops meaning more to you than it should. Two vendors can still sit on one cloud.
I'll be straight about our side of it. We build on this infrastructure too. If Azure has a bad morning, we can have a bad morning with it, and I'm not going to sell you invulnerability I don't have. What I'll sell you instead is perspective, which is free. A rare short outage is what using extraordinary tools cheaply costs you. It's a bargain. And most of the people telling you to armour up against it are the ones selling armour.
Use the two hours to go outside. It'll be back.