Update, July 2026: the model came back online on June 30, 2026, about three weeks after it went dark. I've left this as written during the outage — it is the same prep worth doing before the next one, because there will be a next one.
Last week the U.S. government ordered the most capable AI model on earth switched off. Three days after it launched, gone, and as I write this, still dark. I wrote about what that means for anyone building on these tools in an essay called Rented Intelligence. The reply I got most was a fair one. Fine, but what do I actually do about it?
Here is the weekend version. None of it is hard. Most of it is the kind of thing you mean to do and never quite get to, right up until the week a model you depend on disappears.
Find out how exposed you are
You cannot protect a dependency you have never mapped. Open a blank doc and list every place a model touches your business. The assistant on your site. The tool that drafts your emails. The workflow that tags your leads. The thing that summarizes your calls. Next to each one, write which model it runs on, and what happens to that task if the model is gone for 48 hours.
Most operators have never done this. They learn their exposure the hard way, at the worst possible time. Twenty minutes with a blank page beats that.
Make the model a setting, not a foundation
The teams that got hurt this week had a single model name baked into everything. Swapping it meant touching ten places at once.
Build so the model is one line you can change. If you work in Zapier or n8n, keep the model as a field, not a hardcoded step. If you call an API, route through one place that names the model, so a swap is a single edit. The target is a business where "use a different model" is a config change, not a project.
Wire the backup before you need it
A fallback you have never tested is not a fallback. It is a hope.
Pick a second provider for each task that matters, and run a real job through it this weekend. Watch where the output drifts, fix the prompt, and leave it wired and ready. When Fable went dark my work shifted to Opus in minutes, because the path was already there. The move was boring. Boring is the whole point.
Own the parts no one can switch off
The model is rented. These are not:
- Your prompts. Keep them in a file you control and version them. A prompt that lives only inside a vendor's chat window is one product decision away from breaking, and you will not get a say in it.
- Your data. Export it on a schedule. If leaving a provider means losing your history, you are renting your business too.
- Your workflows and your judgment. Write down how the work actually gets done, so it survives a model swap and survives you having an off day.
These compound. Every month you operate, they get more valuable while the model underneath them gets more replaceable.
Keep a manual mode for anything a customer sees
You do not need a permanent analog plan. You need a bridge. For each customer-facing thing the AI does, know how you would deliver it by hand for two days. Slower, uglier, fine. The businesses that panicked last week had no answer to a simple question. What if it is just off for a while?
Stop running everything on the top model
Match the model to the job. Most of your work does not need the frontier model, and the frontier model is the one most likely to get restricted, repriced, or pulled. Run the routine work on something smaller and cheaper. Save the expensive, exposed model for the few tasks that genuinely need it. You lower your cost and your risk in one move.
The point
A model going dark should be a Tuesday annoyance, not an existential event.
The whole difference lives in how you built before it happened. You do not need anything I sell to do the list above. You need a free weekend and the willingness to treat the smartest tool in your business as the most replaceable part of it.
If you run on AI and you work through this, I would like to hear what your dependency map turned up. Reach out, or tell me what I left off the list.