Estimated read time: 8 minutes
On Wednesday, September 3, a great many people found out at roughly the same moment how much of their working day runs through a single text box.
ChatGPT started throwing errors. Google’s Gemini went wobbly. Grok fell over. Claude had already had its own bad afternoon the day before. For a couple of hours, the tools that a lot of small businesses have quietly built into their daily operations were all somewhere between degraded and useless, and Bloomberg was running a headline about it.
Nothing catastrophic happened. Everything was back within hours. But if your reaction to that morning was to stare at an error message and realize you had no idea what to do next, that reaction is the story, not the outage.
Table of Contents
What actually happened
The clearest record comes from OpenAI, which posts unusually detailed status updates. At around 10:58 AM UTC on September 3, it detected elevated error rates hitting 15 components of ChatGPT and 4 of Codex. Engineers had mitigations in place in roughly 24 minutes and called the incident closed at 11:22 AM. A separate, shorter disruption to ChatGPT’s Work Mode had already come and gone earlier that morning in about five minutes.
Google’s Gemini had the longest day of the three. User reports on September 3 lined up with at least one detected issue that took around two hours to fully clear.
Anthropic’s Claude was technically the day before. Claude Sonnet 5 threw elevated errors on September 2 for roughly 25 minutes before it was resolved.
Downdetector, which measures how loudly people complain rather than anything technical, logged more than 35,000 US reports for ChatGPT, about 1,400 for Claude, and about 1,200 for Grok. That ratio tells you more about market share than about severity.
The disruptions spanned consumer chat interfaces, developer APIs, image generation, and account creation. So depending on what you had built, you either noticed nothing or lost your morning.
It looked coordinated. It probably was not.
Within an hour the speculation was predictable: shared infrastructure, or an attack. Microsoft Azure was also seeing a bump in outage reports, and all three platforms lean on major cloud providers, so the theory had a shape to it.
It is worth being clear that no evidence has established that these incidents were connected. Three companies having overlapping bad windows is genuinely possible, especially when all three are pushing inference infrastructure hard against demand that keeps climbing. Monitoring services logged dozens of similar outages across AI providers during 2026, most of them traced to scaling pressure rather than anything dramatic.
The reason this matters for you is not forensic. It is that both explanations lead to the same place. If it was a shared dependency, then your “backup provider” may not be a backup at all, because it sits on the same cloud. If it was coincidence, then you have just learned that the base rate of AI downtime is higher than the marketing implies. Either way, planning for a single vendor to be reliably available was always the weak assumption.
The 99.9% number is doing a lot of work
Independent uptime monitors put the major AI providers at or above 99.9% availability across 2026. That sounds like a wall you could lean on.
Run the arithmetic. 99.9% uptime allows for just under nine hours of downtime a year. If those nine hours arrive as nine separate one-hour outages scattered across random Tuesday mornings, you will barely notice. If they arrive as one bad afternoon during your busiest week, you will notice a great deal.
Availability percentages describe the vendor’s year. They say nothing about the shape of your exposure, and the shape is the part that hurts.
What a few hours actually costs you
Be honest about your own numbers here, because the answer varies enormously and most people guess wrong in one direction or the other.
If you use AI the way most small businesses do, as a drafting and thinking aid, a two-hour outage costs you some annoyance and a slower afternoon. You write the email yourself. This is genuinely fine. Do not build a resilience architecture for this.
The exposure is different if AI sits inside something customer-facing or time-bound. An AI support agent answering tickets on your site. A scheduled process that generates listings, captions, or summaries on a deadline. A chatbot that qualifies leads. An internal workflow that fifteen people are blocked on. In those cases a two-hour outage is real money, and worse, it is often silent: the process fails, nobody gets an alert, and you find out from a customer.
The dividing line is not how much you use AI. It is whether anything breaks without a human noticing.
The backup plan, in four steps
This is the part that usually gets written for enterprises with platform teams. Here is the version that fits a business with under twenty people.
One: write down where AI is load-bearing. Not everywhere you use it. Just the places where, if the model returned an error for three hours, something would break that a customer or a deadline cares about. For most small businesses this list has between zero and four items on it. If it has zero items, you are done, and you can stop reading. That is a legitimate outcome and most owners never bother to check.
Two: give every item on that list a manual fallback, not a technical one. The instinct is to add a second API. Resist it for now. The cheaper and more reliable answer is usually a documented human procedure: if the AI support agent is down, tickets route to this inbox and this person answers them. If the caption generator fails, we post the three evergreen captions in this folder. Write it in a shared doc. A one-page runbook beats a clever failover you have never tested.
Three: add monitoring before you add redundancy. The expensive failure is not the outage, it is the four hours before anyone realized. If a process depends on an AI call, make it tell you when it fails. Even an email to yourself on error is a large improvement over silence. Most small business AI breakage is discovered by a customer, which is the worst possible detection system.
Four: only then consider a second provider, and know what you are buying. Multi-provider routing works, and it is not free. AI APIs are not drop-in interchangeable. A prompt tuned for one model does not reliably produce equivalent output from another, so real redundancy means maintaining and testing two sets of prompts, not just holding a spare API key. If you use a tool that already routes across providers for you, you have most of this benefit without the work. If you would be building it yourself, be sure the thing you are protecting is worth the ongoing maintenance.
The order matters. Most people skip straight to step four, build something half-finished, never test it, and end up with two things that can break instead of one.
What not to do about this
Do not conclude that AI is unreliable and retreat. A handful of hours a year is not a reason to abandon tools that are demonstrably saving you time. The correct response to a dependency is to manage it, not to pretend you do not have it.
Do not buy an “AI uptime” product because of one bad Wednesday. The category of vendors selling reassurance after a news cycle is well established and their timing is excellent.
Do not assume self-hosting solves this. Running an open model on your own infrastructure moves the failure from someone else’s on-call rotation to yours. For a business without an ops person, that is usually a downgrade in reliability, not an upgrade.
Do not skip the boring version. The single highest-value thing to come out of September 3 for most small businesses is a one-page document nobody will read until the day they need it. That is what good continuity planning looks like, and it is the same reason data backup and disaster recovery is unglamorous and non-negotiable.
The vendors will keep improving. Outages will keep happening anyway, because that is what infrastructure does at scale. The businesses that will shrug off the next one are the ones that spent twenty minutes this week writing down what they would do.
FAQ
Were the September 2026 outages actually connected? No connection has been established. All three companies published status updates and none pointed to a shared cause. Speculation focused on cloud infrastructure, because Azure reported elevated issues in the same window, but that remains speculation rather than a confirmed link.
How often do AI providers actually go down? More often than most people assume. Monitoring services recorded dozens of incidents across major AI providers during 2026, most brief and most attributed to inference infrastructure straining under demand. Headline availability stayed at or above 99.9%, which still permits nearly nine hours of downtime a year.
Should I pay for an enterprise or business tier to get better reliability? Sometimes, but read what you are buying. Higher tiers often provide priority capacity and an SLA with service credits. Service credits refund a fraction of your subscription; they do not compensate you for a lost day. Buy the tier for the capacity and support, not for the guarantee.
Is running two AI subscriptions a reasonable backup? For manual work, yes, and it is cheap insurance. If you draft in ChatGPT and it is down, opening Claude costs you nothing but a paste. For automated workflows, a second subscription is not the same as a second integration, and the integration is the expensive part.
What if my whole product depends on an AI API? Then this is a core engineering concern rather than a small business tip, and you want genuine multi-provider routing, graceful degradation, and a status page of your own. The important part is deciding deliberately, because a product built on one API without a fallback has made that decision by default.
How do I know if AI is load-bearing in my business? Ask what would still be broken at the end of the day if every AI tool you use returned an error all morning. If the honest answer is “some things would be slower,” you are fine. If the answer names a customer-facing process, that process needs a written fallback.
Related Coverage
- Best Data Backup and Disaster Recovery Tools for Small Business – the same continuity thinking, applied to the data you cannot lose
- SaaS Tools AI Disruption Audit – how to audit which vendors you are actually dependent on
- Best AI Agent Tools for Small Business – the tools most likely to be load-bearing when a provider stumbles
Faceted Media Magazine covers business, AI, and entrepreneurship for the people building what’s next.