When an AI tool stops responding, freezes mid-answer, or shows a message like “we are experiencing high demand,” it is easy to assume the problem is on your end. Most of the time, it is not. AI services go down for a handful of well-understood reasons, and the same patterns show up again and again across every major provider. Understanding those causes will not bring a service back faster, but it will tell you whether to keep waiting, switch tools, or simply step away for a few minutes.
This guide walks through the real reasons AI tools go offline, written in plain language. For each cause, we explain why it happens, roughly how long that type of outage tends to last, and what you can reasonably do about it. If you just want to know whether a tool is down right now, the live dashboard on our home page tracks the major services in real time.
Demand spikes and capacity limits
The single most common reason an AI tool slows down or refuses requests is simply that too many people are using it at once. You have probably seen this as an “at capacity” message, a spinning cursor that never resolves, or replies that suddenly become much slower. The service is not broken in the usual sense. It has run out of room to serve everyone.
AI tools are unusually prone to this for a few reasons. Running a large model is expensive in a way that ordinary websites are not. A normal web page can be served to millions of people from cheap, plentiful servers. An AI model, by contrast, needs specialized chips (often called GPUs) to generate each answer, and those chips are costly, in limited supply, and cannot be spun up instantly. When a provider runs out of available capacity, there is no quick way to add more in the moment.
Demand for AI is also extremely “spiky.” A viral post, a new feature launch, a popular use case during business hours, or a school deadline can all send usage surging far above normal within minutes. Providers plan for typical demand plus a buffer, but a large enough spike will still exceed what they have ready. When that happens, the service has to protect itself by slowing responses, queuing requests, or temporarily turning some users away so the whole system does not collapse.
How long it lasts: Capacity issues are usually among the shortest outages. They often ease within minutes to an hour as the surge passes or the provider shifts resources around. Paid plans frequently keep working even while free access is throttled.
What to do: Wait a few minutes and try again, ideally with a shorter or simpler request. If you rely on a tool for work, having a second service ready helps. You can compare the current state of the major options, for example on our ChatGPT and Claude status pages, and use whichever one is healthy.
Dependence on upstream cloud and infrastructure providers
Almost no AI company runs all of its own physical hardware. Instead, they rent computing power, storage, and networking from a small number of large cloud platforms. This is efficient and normal, but it creates a hidden dependency: if the underlying cloud provider has a problem, every AI tool built on top of it can go down at the same time, through no fault of the AI company itself.
These upstream services are layered. An AI tool might depend on a cloud region for computing, a separate service for user logins, a content delivery network to move data quickly around the world, and a database to store conversations. A failure in any one of those layers can take the whole product offline or break specific features (for example, you can log in but cannot send a message). Because so many companies share the same handful of providers, a single large cloud incident can appear to break “the whole internet” at once, including several unrelated AI tools.
How long it lasts: This varies widely. Minor infrastructure hiccups resolve in minutes. A serious failure at a major cloud provider can last several hours, and recovery is often gradual rather than instant, because services have to come back online in the right order.
What to do: If several unrelated apps are misbehaving at the same time, an upstream provider is a likely cause, and there is nothing you can do to speed it up. Switch to a tool hosted elsewhere if you have one, and check back periodically. Our full guide on what to do when an AI tool is down covers how to tell a personal problem apart from a provider-wide one.
Failed or risky deployments and config changes
AI products are updated constantly. Teams ship new features, security patches, and performance improvements many times a week. Most of these changes are invisible and harmless. Occasionally, one of them introduces a bug, behaves unexpectedly under real traffic, or interacts badly with another part of the system. When that happens, a routine update can knock the service offline.
A related and surprisingly frequent culprit is a configuration change: a small adjustment to a setting, a limit, or a routing rule rather than the software itself. These changes can be deployed quickly and can have outsized effects. A single mistaken value can route traffic to the wrong place or block valid requests across the entire platform. Because config changes are easy to make and easy to get wrong, they cause a meaningful share of real-world outages.
The good news is that mature providers are well practiced at handling these. They roll changes out gradually, watch for problems, and keep the ability to “roll back” to the previous working version. So while a bad deployment can cause a sharp, sudden outage, it is also one of the most quickly reversible kinds.
How long it lasts: Often short, from a few minutes to around an hour, because the fix is usually to undo the change. Outages that need a genuine code fix rather than a rollback can take longer.
What to do: Wait and retry. There is nothing on your side to fix, and the provider is almost certainly already aware. Avoid the temptation to repeatedly resubmit a large task, which can add load while the system is recovering.
Regional and network issues
The internet is not one single network. It is a vast collection of regional networks, undersea cables, internet providers, and data centers stitched together. An AI tool can be perfectly healthy overall while being unreachable for people in one country, on one internet provider, or in one part of the world, because of a problem along the path between you and the service.
These problems take several forms. A data center serving one geographic region might have trouble while others are fine. A networking issue at your own internet provider can make a service feel “down” when it is working for everyone else. Routing problems, where data takes a wrong or broken path across the internet, can cut off whole areas. Even the content delivery networks that speed up access can have localized faults.
Because these issues are regional, two people can strongly disagree about whether a tool is “down.” That is not confusion: it genuinely is down for one of them and up for the other. This is also why a status checker that tests the service independently is more reliable than your single experience.
How long it lasts: Highly variable. A local internet provider glitch may clear in minutes. A regional data center or cable problem can take hours to reroute or repair.
What to do: Test whether the problem is yours by trying another website, switching from Wi-Fi to mobile data, or restarting your router. If only one AI tool is affected for you while a checker shows it healthy elsewhere, the issue is likely local to your network or region rather than the service itself.
Scheduled maintenance
Not every outage is an accident. Providers sometimes take a service, or part of it, offline on purpose to perform upgrades, migrate data, replace hardware, or apply changes that cannot be done safely while the system is live. This is called scheduled or planned maintenance, and it is a sign of a well-run service rather than a failing one.
Responsible providers try to schedule this work for quiet hours and announce it in advance on their official status pages. During the window, you might see a clear maintenance notice, reduced functionality, or a temporary inability to sign in. Unlike a surprise outage, planned maintenance has a known end time, so it is the most predictable category of all.
How long it lasts: As long as the announced window, which is usually defined ahead of time. Most maintenance is measured in minutes to a couple of hours, and the service often returns earlier than the stated end.
What to do: If you see a maintenance message, simply wait for the window to end; retrying will not help and the timing is already known. If the work routinely falls during your peak hours, it is worth keeping a backup tool available for those windows.
Security events and rate limiting
The final category covers outages that are caused, directly or indirectly, by protecting the service. AI tools are high-profile targets. They face attempts to overload them, abuse them, or break in. The defenses against those threats can themselves make the service unavailable to ordinary users for a time.
One common example is a flood of malicious traffic designed to overwhelm a service so that real users cannot get through. Defending against this can slow everyone down or trigger protective measures that temporarily block legitimate visitors along with the bad traffic. Another example is rate limiting: providers cap how many requests a single user, app, or account can make in a given period, both to share capacity fairly and to stop abuse. If you hit that cap, you may see errors or be temporarily locked out even though the service is otherwise fine. This often feels like an outage but is really a deliberate limit.
Occasionally, a provider will also take a feature or the whole service offline as a precaution while investigating a genuine security concern. These events are less frequent but can last longer, because safety comes before speed of recovery.
How long it lasts: Rate limits typically reset within seconds to minutes, or at the start of the next billing or usage period. Outages tied to attacks or precautionary shutdowns can range from minutes to several hours depending on severity.
What to do: If you suspect a rate limit, slow down, space out your requests, and avoid sending many large tasks back to back. If the whole service is affected by a security event, there is nothing to do but wait and follow the provider’s official updates.
How long different outages last
Outage length depends almost entirely on the cause. The table below gives rough, general guidance. Real incidents vary, and these are not promises, but they help set expectations so you know whether it is worth waiting.
| Cause | Typical duration | Best response |
|---|---|---|
| Demand spike / “at capacity” | Minutes to about an hour | Wait and retry; use a backup tool |
| Upstream cloud or infrastructure failure | Minutes to several hours | Switch to a differently hosted tool; check back |
| Failed deployment or config change | Minutes to about an hour | Wait; a rollback usually fixes it |
| Regional or network issue | Minutes to hours | Test your own connection first |
| Scheduled maintenance | The announced window | Wait for the stated end time |
| Security event or rate limit | Seconds (limits) to hours (incidents) | Slow down, or wait for official updates |
A useful rule of thumb: most AI outages that you will personally notice are short, on the order of minutes, and resolve without you doing anything. The longer, multi-hour events are rarer and usually trace back to a serious upstream or infrastructure problem. If a tool has been completely unreachable for more than an hour, it is worth checking an independent status source rather than retrying endlessly.
How we detect outages
We monitor AI services independently and continuously, rather than relying only on what each company reports about itself. Official status pages are valuable, but they sometimes lag behind reality or describe a problem in vague terms. By checking the services ourselves from a single, independent monitoring location, we can often confirm an outage, or rule one out, before it is formally acknowledged.
In practice, that means regularly testing whether each service responds, how quickly it responds, and whether key features behave normally, then comparing results across locations so we can tell a genuine outage apart from a local or regional glitch. When enough signals point the same way, we flag the service as degraded or down. You can read the full details of how this works, including what we test and how we decide, on our methodology page.
If you are unsure whether a tool is down right now, start with the live dashboard, check the specific service page you depend on, and if it really is down, follow our step-by-step guide for getting your work done in the meantime. Knowing the cause will not always speed things up, but it will almost always tell you whether to wait it out or move on.
Frequently asked questions
Why do AI tools say "at capacity" so often?
Running an AI model takes expensive, limited specialized chips, and demand for AI is very spiky. When too many people use a tool at once, the provider runs out of room and temporarily turns some requests away. These outages are usually short, often clearing within minutes to an hour.
Is the AI tool down for everyone, or just me?
It can be either. Regional and network problems mean a service can be down for you while working fine for others. Test another website, switch networks, or check an independent monitor. Our live dashboard at / shows whether the major services are down more broadly.
How long do AI outages usually last?
Most outages people notice are short, on the order of minutes, and resolve on their own. Capacity, deployment, and rate-limit issues recover quickly, while serious cloud or infrastructure failures can last several hours. If a tool is unreachable for over an hour, check an independent status source.
What should I do while an AI tool is down?
Wait a few minutes and retry with a simpler request, avoid resubmitting large tasks repeatedly, and switch to a backup tool if you have one. For a full walkthrough, see what to do when an AI tool is down.
How do you know an AI service is down?
We test each service ourselves, continuously from a single monitoring location, checking whether it responds, how fast, and whether key features work. We compare results to separate real outages from local glitches. The full process is on our our methodology page.