DeepSeek's 7-Hour Outage: The AI Vendor Risk Lesson

Seven hours of silence

At 9:35 PM on March 29, the chatbot that roughly 355 million people use every month stopped working. Logins failed. Prompts hung. Developer workflows that depended on the DeepSeek API returned errors all night. The outage lasted seven hours and thirteen minutes — and a separate round of performance problems dragged into the next morning.

It was DeepSeek’s longest downtime since launch. The company’s previous record was under two hours, so this was an order of magnitude worse than anything users had planned for. DeepSeek still hasn’t publicly explained what caused it.

If you’re a small business using AI tools, this is the story you should have pinned to the fridge. Not because DeepSeek is unusually fragile — it isn’t — but because every AI provider will eventually have a day like this, and AI vendor risk is a problem most small businesses haven’t thought through yet.

What actually happened

DeepSeek’s status page logged the incident as a “major outage” starting late on March 29. Users across regions reported three symptoms: failed logins, requests that timed out without a response, and sessions that expired mid-conversation. The company marked the problem resolved late that night, but users reported continued degraded performance through the following morning.

A few facts to anchor the story:

The outage hit both the consumer chatbot and the developer API.
It affected users worldwide — not a regional incident.
No official root cause has been published. Industry speculation covers the usual suspects: a bad deployment, capacity pressure from a new model rollout, or an infrastructure fault.
DeepSeek processed the recovery in stages, so some users got service back hours before others.

None of those details are unusual. Every major cloud provider has outages that look like this. Azure, AWS, Google Cloud, OpenAI, Anthropic — they all have public status pages full of incidents. The uncomfortable part is how much of your business runs on top of them.

Why a 7-hour outage is a small-business problem

When an enterprise’s AI vendor goes down, they have a procurement team, a fallback contract, and an engineer on call to switch traffic. When a small business’s AI vendor goes down, three things happen at once:

The customer-facing chatbot stops answering. Visitors get either a generic error or dead silence. You lose leads for the duration of the outage, and some of them don’t come back.
Internal workflows halt. If your content calendar, intake forms, or scheduling uses an AI model in the middle, every step after that node stalls until you manually intervene.
Nobody on staff knows how to swap providers. A developer with API experience could route traffic to a different model in an hour. Most small businesses don’t have that developer, so they sit and wait for the vendor to come back.

That last point is the real AI vendor risk for a two-person contractor’s office or a rural restaurant. The cost of downtime isn’t just seven lost hours — it’s the discovery that the business has no lever to pull when the cloud gets rainy.

The lesson: build for the day your vendor fails

The goal isn’t to avoid DeepSeek, OpenAI, or any specific provider. The goal is to make sure a seven-hour outage at any one of them doesn’t freeze your operations. Four practical moves make a real difference:

Write down which AI dependencies are load-bearing. Walk through your customer-facing and internal workflows and mark every step that calls an AI model. You usually find one or two that, if they fail, take the whole flow with them. That’s your short list.
Choose one backup model per critical path. You don’t need five. One is enough. For text generation, that often means pairing an OpenAI or DeepSeek model with an Anthropic or open-source Llama model — different companies, different infrastructure. For voice and transcription, Whisper plus a commercial provider covers most failure modes.
Keep a graceful fallback, not a blank screen. If your intake widget can’t reach its AI model, it should still collect name, phone, and a free-text message, then queue the work for human follow-up. A working form is better than a broken chatbot. Our Hollr intake widget defaults to this pattern for exactly this reason.
Own your data and prompts. If your content, reviews, appointments, or customer history live inside a single vendor’s account, moving elsewhere takes days. If they live in your own database and you call the AI as a stateless service, you can swap providers in an afternoon.

None of these require a platform team. They require an hour of honest conversation about what stops when the vendor stops.

What small businesses should do this month

Even without an engineering background, a small business owner can take three concrete steps in the next few weeks:

Make a one-page AI inventory. List every tool that uses AI under the hood — chatbots, email assistants, scheduling, content generators, voice agents. For each, note the vendor, what happens if it’s down for a day, and who on your team knows how to turn it off.
Test one outage. Pick your most important AI-powered workflow and intentionally disable it for an hour. Watch what breaks, then fix the most embarrassing thing you find. This is the cheapest resilience exercise you’ll ever run.
Ask your vendors the uncomfortable question. What’s their uptime SLA? What’s their recent incident history? What’s the escalation path if something breaks at 8 PM on a Friday? If the answer is “we don’t publish that,” treat it as information.

We’ve written before about why AI talent churn translates into product risk and why single-provider dependency gets worse as the AI market consolidates. DeepSeek’s outage is a concrete, recent example of the same underlying principle: the plumbing matters, and the plumbing occasionally leaks.

The bottom line

A seven-hour AI outage isn’t an emergency for a business that has a backup plan. It’s a bad afternoon. For a business that doesn’t, it’s a day of dead air, lost leads, and a conversation with staff that should have happened six months earlier.

DeepSeek will fix whatever broke on March 29. The next outage — at some other provider, for some other reason — is already on the calendar. You just don’t know when.

If you’d like help mapping your AI dependencies and designing a fallback plan that actually works for a small team, our consulting practice does this with Appalachian businesses every week. We also help teams set up the AI infrastructure — multi-model routing, graceful fallbacks, and the monitoring needed to catch an outage before a customer does.

The rain is coming. Build the roof before the storm.