TL;DR
- The large cloud outages of 2026 were long. Several ran 14 to 22 hours, and Microsoft 365's outage on Nov. 12 was still not fully resolved a day later.
- Most of them did not start with a software bug. They started with power, cooling, fiber cuts and, in one case, physical attacks on data centers. Physical problems take longer to fix than a bad update.
- Electricity is the new pressure point. The grid regulator for North America issued its highest-level alert this year about large data centers suddenly dropping off the grid.
- For a small business, the question is not which provider never fails. It is whether the business can keep taking calls, orders and payments for one bad day, and limp through three.
- Governments in Europe, the United Kingdom and the United States already tell households to be ready for 72 hours on their own. The same time frame is a sensible test for a business.
What Changed in 2026
For years, the typical cloud outage looked like this: a provider pushed a bad configuration change, a large part of the internet broke for two or three hours, and the provider rolled the change back. Those still happen. A Cloudflare cleanup task in February withdrew a large share of customer network routes for about six hours.
The longer outages this year had a different shape. Public incident reports from the major providers describe a chiller failure that overheated part of Amazon Web Services' largest region in Northern Virginia in May, with recovery taking roughly 14 hours. Microsoft Azure's West US 2 region was down for about 22 hours at the end of May after a power and cooling problem. In July, a brief voltage drop from the local utility tripped both power feeds at a Google Cloud site in the Netherlands, a backup unit failed, and cooling followed. That incident ran close to 15 hours. In March, drone strikes damaged Amazon data centers in the Middle East, an event no failover plan for a single region was designed to absorb.
The pattern matters for planning. A software rollback takes minutes once the cause is found. Replacing a chiller, restoring power to a building or rerouting around cut fiber takes hours or days. When an outage begins with something physical, the reasonable assumption is that it will be measured in hours, not minutes.
Why the Big Providers Fail Together
Most of the internet runs on a small number of companies. Amazon, Microsoft and Google host a large share of the world's websites and business software, and Cloudflare sits in front of millions of sites as a security and delivery layer. Inside each provider, a few regions carry far more than their share. Northern Virginia alone hosts more data center capacity than most countries.
This concentration is why a single failure shows up everywhere at once. A business that has never heard of a particular region still loses its point-of-sale reports, its scheduling app and its email when that region fails, because its vendors chose the same place.
There is also a hidden layer. Large platforms rely on a few shared services, such as sign-in systems and the internet's address book, that can pull other regions down when they fail. A vendor that advertises "multi-region" may still depend on one of those shared pieces. That claim is worth asking about directly before relying on it.
The Grid Is Now Part of the Story
The newer risk sits upstream of the data centers. AI computing has made data centers some of the largest and most volatile electricity customers on the grid. Their power draw can swing sharply within seconds.
In May, the North American Electric Reliability Corporation, the body responsible for grid reliability across the United States and Canada, issued a Level 3 alert, its most serious category, after large computing loads dropped off the system or oscillated within seconds. Reporting in July described a fault in Northern Virginia that knocked roughly 3.8 gigawatts of data center load offline in under 30 seconds. The grid held. The operator of the largest grid in the country, PJM, has since worked on emergency rules that would let it cut power to very large data centers to protect everyone else.
None of this means a blackout is coming. Some energy analysts argue the official forecasts overstate the risk, partly because not every planned data center will be built. It does mean two things for a business owner. Cloud providers now face power problems they did not face five years ago, and winter peaks, when heating demand is highest, are when the grid has the least room to spare. In the North Country, a winter storm that takes down local power lines remains a far more likely event than any national failure, and the planning is the same.
Think in Days, Not Hours
The 72-hour figure comes from household emergency planning. FEMA's long-standing guidance asks households to be self-sufficient for at least three days. The European Union adopted the same minimum in its 2025 preparedness strategy, and the United Kingdom launched a campaign in October asking households to keep three days of food and water. Germany advises up to ten days.
The reasoning transfers well to a business. The first 24 hours of a large outage are confusing, and help goes to the most urgent cases first. A plan that only covers a two-hour disruption fails exactly when it is needed.
A practical way to use the time frame is to walk through it in three stages.
Hour one: know what is down. Many hours are lost restarting routers and calling the internet provider when the problem is a cloud vendor several states away. One person should know where to check: the vendor's status page, an independent outage tracker, or a quick call to a neighboring business. If the vendor is down, local troubleshooting is wasted effort.
Day one: keep serving customers. The essentials are taking calls, taking orders and taking payment. Each depends on something: a phone system, a booking app, a card reader. For each one, the plan names the fallback. A card reader that can store transactions offline. A paper order form. A cell phone that receives forwarded calls. A pre-written note for the website and the business's online listings that says email is delayed and gives the phone number. The earlier piece on the Microsoft 365 outage covers the email side of this in detail.
Days two and three: work without the cloud. If the outage continues, the questions change. Can payroll run? Can invoices go out? Are the day's appointments, customer phone numbers and price list available anywhere except the cloud service that is down? A printed or offline copy of the next week's schedule and the top customer contacts, refreshed weekly, covers most of this at almost no cost.
A One-Page Plan Someone Owns
Most small businesses do not need a continuity binder. They need one page, kept in two places (one of them paper), with a named owner. A workable version answers six questions:
- Who decides that the business is in outage mode, and who is the backup if that person is away?
- Where does the team check whether the problem is local or a vendor's?
- How will staff reach each other if the usual chat or email is down?
- How will customers be told, and through which channels?
- What is the fallback for calls, orders and payments?
- Where are the offline copies of schedules, contacts and prices, and when were they last refreshed?
The page is only useful if it has been tried. A short exercise twice a year, such as pretending email and the booking system are down for a morning, will show which answers were wishful. The most common finding is that the backup channel was never set up, or that the only copy of the plan lived in the system that failed.
What the Noise Gets Wrong
Online commentary about outages tends to swing between two extremes: that the internet is about to collapse, or that the cloud is so reliable that planning is unnecessary. Neither fits the record.
The large providers are, on most days, more reliable than anything a small business could run itself. They also fail, sometimes for a full working day, and the failures of 2026 were longer and more physical than those of earlier years. Claims of a coming "global internet shutdown" on a specific date have circulated for years and have never happened. A long outage at one major vendor, on the other hand, happens several times a year.
The useful position sits between those extremes. Keep the cloud services. Assume one of them will be unavailable for a day at an inconvenient time. Decide in advance what the business does when that happens, and write it on one page.