More Than Email Went Dark for 24 Hours: the Microsoft 365 Outage Traced to a Capacity Upgrade
At noon on Nov 12, 2026, Microsoft 365 began failing in North America and spread to 74 countries and territories. Servers failed health checks one after another, and restarted servers fell over again. Microsoft blames a fault from a capacity upgrade — not a cyberattack, not AI — but the full report has still not arrived.

Just after noon Eastern Time on Thursday, November 12, 2026, users in the United States found that Outlook would not open. Then Teams calendars went, then Copilot's answers, then the collaboration behind SharePoint — and overnight the failure spread from North America to 74 countries and territories, forcing hospitals, government offices, and airlines back onto phone lines, paper records, and, in some offices, fax machines. By the time Microsoft declared everything restored on the morning of Saturday, November 14, about 36 hours had passed. And the cause Microsoft eventually gave was a single sentence: a problem introduced during an earlier network capacity upgrade, which surfaced following a more recent network change. Not a cyberattack. Not an AI running out of control. Its own upgrade. This piece walks through those 36 hours in order: what broke first, why the fix kept failing, what Microsoft said — and what it has still not answered.
From one mailbox to 74 countries
The failure started small: users could not open their Exchange Online mailboxes. According to Microsoft's own record, at 5:54 pm GMT on November 12 it acknowledged the problem, still limited to North America at the time; just 13 minutes later, it warned that users in other regions could be affected too. That speed alone said how serious it was — this was not a small fault in one data center, but a systemic problem spreading in real time. By later that day the list of broken things was long: Outlook on the web, desktop, and phone could not reach mailboxes; Teams chats and meetings mostly kept working, but the features that depend on mailboxes and calendars — scheduling, creating new chats — started failing; everything in Copilot, SharePoint, Planner, To Do, and Power Apps that depends on a mailbox or calendar would not load; and parts of Defender, Purview, and Intune, the security and admin tools, were affected too. The one thing that mostly still worked was the files themselves — what was stored in SharePoint could still be opened. The files were not lost. The road to them was.
Downdetector, the outage-tracking site, logged roughly 130,000 to 140,000 Outlook problem reports worldwide at the peak, across 74 countries and territories. Those numbers should be read for what they are: reports filed by users themselves, not a count of affected accounts — the real number of people affected is certainly higher. The most visible casualties were the industries that cannot work without email: in India, IndiGo and Akasa Air publicly warned passengers that email support might not be able to receive their messages; hospitals, courts, and government offices had routine operations disrupted. Personal services were hit too — Outlook.com, OneDrive, OneNote, and Microsoft To Do were all on the affected list. This was not a failure aimed only at business customers; it hit offices and individual users at the same time.
One detail is easy to miss: before the main incident, there had already been a smaller one the same evening. At 8:15 pm India time on November 12, the Americas saw a shorter Teams problem — delayed and failed chats — resolved by about 9:25 pm. The main incident began at 9:47 pm, barely twenty minutes later. Microsoft has not said whether the two were connected, but the timing is worth recording: the system had already wobbled once before the main failure, like a small tremor before an earthquake. Nobody treated it as a warning at the time; in hindsight, it may have been the first link in the same chain.
A day of fixes that did not hold: why servers kept falling over
Microsoft's incident updates during the outage are the most valuable record of what happened, because they honestly wrote down each failed step. This is how it unfolded: in the early hours of November 13, Microsoft said something crucial — the extra load was not driven by more traffic, but by "a large number of servers unexpectedly becoming unavailable." In other words, it was not a surge of users crushing the system; the system's own servers were dropping out one after another, and the ones still running had to carry all the work, and then began to struggle too.
The reason for the dropouts, in Microsoft's account, was that servers were failing automated health checks. A health check is a routine test that decides whether a server is still fit to receive traffic; servers that fail are automatically pulled out of service. The problem was the chain reaction: the more servers were pulled, the harder the remaining ones had to work, and the harder they worked, the more likely they were to fail their own health checks, so more servers were pulled. Meanwhile, engineers found that servers they restarted would become unresponsive again as soon as real traffic reached them — fixed, back online, down again. In an update at 3:30 am India time on November 13, Microsoft admitted: "our actions taken thus far have not provided the anticipated relief." In plain language: nothing we have tried is working.
The turning point came at about 7:20 am India time on November 13, when Microsoft started reverting "a recent network configuration change that correlates with the start of impact." Then came patches and server reboots; the midday update reported only "marginal success," and it was not until 12:59 pm that Microsoft could report "positive results for server stability." At 2:40 pm the same day, Microsoft extended the email timeout window to 48 hours — messages sent during the outage would keep retrying for two days instead of bouncing. Asia Pacific was the first region to recover, at 4:34 pm, followed by Europe and North America; by the afternoon GMT, Microsoft said email queues were "significantly reduced and approaching normal levels." Microsoft records the end of impact at 9:40 pm India time on November 13, just under 24 hours after it began — but it kept the incident open until the morning of November 14, when it confirmed the delayed mail had been processed, and posted its final update: the services "have fully recovered." From the first alarm to that final update, about 36 hours.
The recovery was uneven, and Microsoft's updates say so plainly. Teams chats and meetings came back relatively early, but calendar functions that depend on Exchange Online — scheduling, checking availability — lagged behind for longer; clearing the email backlog took longer still, and at 1:32 am on November 14 Microsoft was still writing "we do not have an ETA for full recovery yet," only posting its final confirmation between 9:20 and 9:45 am. That state of "mostly back, but parts still dragging" is exactly how large cloud outages normally end: it is not a single switch that turns everything back on, but dozens of interlocking subsystems lighting up one by one, not necessarily in the order you need most. The gap between "it feels fixed" for users and "fully recovered" on Microsoft's status page is often several hours of queued mail.
What Microsoft said: an apology, a cause, and a deliberate exclusion
On the evening of November 13, Microsoft posted its apology on X, and the wording is worth keeping in full: "We are deeply sorry for the disruption this incident has caused our customers and partners. We know the trust they place in our services and take that responsibility extremely seriously." The same post gave the cause: "This was a technical issue not related to cyber or AI / agents. The incident was introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change. We have addressed the underlying issue and continue to restore service to remaining customers and regions." Microsoft also explained why the recovery was so slow: given the scale of demand, services were restored gradually, "validating stability at each stage" — slow, but steady. And it promised a full post-incident report once its review is complete.
The most notable line in that statement is the deliberate exclusion: "not related to cyber or AI / agents." In 2026, when any large company's systems go down, the first two theories are always the same: was it hacked, or did an AI agent run out of control? Microsoft closed off both routes in its very first apology. That itself says something about how the times have changed: a few years ago, the standard explanation for an outage only had to rule out a cyberattack; now it has to rule out AI as well. The other side of the statement is the four questions Microsoft did not answer: what exactly went wrong in the capacity upgrade, which network change triggered it, whether the health-check system itself made things worse, and whether customers will receive service credits. The full post-incident report has not been published, and no date has been given — until it appears, the public explanation is that one sentence.
Two unconfirmed reports should be flagged clearly as such. The first comes from The Verge, which, citing sources, linked the outage to configuration changes in Azure virtual networks; Microsoft has not confirmed this in its public statements, and its incident record does not list Azure as an affected service — Azure's own status page showed no issue through the whole outage, even though complaint trackers showed a spike in Azure reports too. The second is a coincidence of timing: the outage began two days after Microsoft's monthly security updates, and Microsoft has not linked the two, with no evidence that they are connected. Both can only count as "unconfirmed," not as conclusions; but they show one thing — until the full report appears, outside accounts of "which wire went wrong" will keep changing, and the only statement that stands is the one Microsoft itself admitted: a hidden fault left by a capacity upgrade, brought to the surface by a network change.
Microsoft's last word came on November 14, through the Microsoft 365 Status account: "We have completed our efforts to address the underlying issue of the incident and restored service to the remaining impacted customers and regions. Further updates can be found at MO1550821 in the Microsoft 365 admin center." The wording is careful — it speaks of completing the effort to address the underlying issue, not of having explained the root cause. The distinction matters: the first says the service is back, the second says the problem has been explained. As of writing, the full post-incident report Microsoft promised has not been published, and no date has been given. Restoring service and explaining the cause are two different jobs; the first is done, the second is still waiting.
Even the AI tools went down with it: the second layer of impact
The outage had a second layer of impact that is easy to overlook: it did not only interrupt Microsoft's own services, it interrupted the AI tools built on top of them. OpenAI reported trouble connecting ChatGPT to Microsoft 365 during the outage; Anthropic said that mail, calendar, and search requests in its Microsoft 365 connector and Claude for Outlook returned errors for more than 24 hours. In other words, that day it was not only people waiting for email — the AI assistants were waiting for email too. Assistants sold as able to "read your inbox and arrange your schedule" could do nothing while the inbox itself was down.
That detail is worth pulling out on its own, because it exposes a reality that is still forming: the capability of an AI assistant is, to a large extent, bounded by the services it stands on. Copilot was itself one of the affected services that day, and the ChatGPT and Claude connectors to Microsoft 365 depend on the same mailboxes and calendars. When we talk about AI assistants, we talk about models, prompts, and permissions; this outage is a reminder that there is a more basic layer of dependency underneath — no matter how smart your assistant is, it is only as smart as the server under its feet. When email, calendars, documents, and AI assistants all come from the same supplier, a failure in any one link stops the whole chain.
The outage was not over when the phishing started
There is one more thing that happened after the failure, and it deserves a line in this piece. On November 14, Microsoft's threat intelligence team warned that phishing had already been detected on November 13, exploiting the outage: the scam emails cited the real incident number, EX1550543, and claimed held mail would be deleted within 24 hours, trying to get recipients to click links or hand over account details. It is an old routine: after every large outage, someone impersonates the official side with "restore your mailbox" or "claim compensation" messages. Microsoft will not ask for your password by email, and any message that cites the incident number and pressures you to act immediately should be treated as a scam first. Incident numbers are public information — scammers can get them, and so can you: EX1550543 and MO1550821 can both be looked up in full in Service health in the Microsoft 365 admin center.
For users and IT teams: do these things now
If you are an ordinary user, the incident is over, but three small things are worth doing right now. First, check the important messages sent between November 12 and 13 — meeting invites, payment confirmations, interview links — and make sure they actually reached the other side. Microsoft says the backlog has been delivered, but some senders' systems may have given up retrying before the outage ended. Second, look in your Outbox and Drafts for messages that never left. Third, treat any message claiming to be from Microsoft, mentioning this outage, and pushing you to click a link or hand over a password as fraud until proven otherwise.
If you run IT for an office, the list is longer: first confirm in Service health that EX1550543 (Exchange Online) and MO1550821 (Microsoft 365) show as restored for your tenant, and save the incident history — Service health keeps only 30 days of history. Then check mail flow for stuck or late messages and bounce notices, and if you run a hybrid setup, check the queues on your own servers too. Tell staff plainly that email is back, and ask them to check for replies and invites from Thursday and Friday that may have arrived late or not at all. When the post-incident report appears, download it and keep a copy. And if you plan to claim service credits under the uptime guarantee, write down your own start and end times of impact now, while they are fresh — in a few weeks, nobody will remember them exactly.
Back to the outage itself. Its real lesson is not in the technical details, but in a simpler question: when one supplier provides your email, chat, calendar, files, security tools, and AI assistant, convenience and fragility are two sides of the same coin. These 36 hours did not prove that "the cloud is unreliable" — quite the opposite: the service did come back, and no mail was lost. What they proved is that once centralization fails, the radius of impact is everything. You do not need a second email system to guard against one outage. But you do need a written answer to a simple question: the next time Microsoft 365 will not open, how do you reach your staff, and how do you reach your customers? That answer should be written now, and rehearsed once — not worked out during the next failure.
It does not need to be complicated — a few lines are enough: where the staff phone list lives, and whether you can reach it offline; which channel takes over for customers — phone calls, a WhatsApp broadcast, or a notice page on the company website; who has the authority to order the switch to the backup channel, and who sends that order; and most importantly, whether the plan has ever actually been rehearsed. Many companies discovered on the day itself that their backup contact details were stored inside the very mailbox that was locked, or that only one person knew where they were. During this outage, hospitals and airlines fell back on phones and paper — proof of one thing: the most old-fashioned backup plan is often the most reliable. Write it down, print it out, and put it in a drawer. Technology goes down. Paper does not.
Comments (0)
Sources
Related articles
On 20 November 2026, Signal said it will block in-app screenshots by default on iOS 27 and auto-enable screen security on Android when an AI assistant reads the screen, to stop OS-level Siri AI and Gemini Intelligence scanning the display. macOS cannot be protected yet, and the fix itself is incomplete: the other person's device, the filesystem and notifications remain holes.
On 20 November 2026, OpenAI published its anti-distillation defences for GPT-6.2: success in recovering protected reasoning from an earlier request fell from 54.0% to 0.2%, and from the current request from 22.9% to 0.9%. Reasoning is now tied to conversation history, enforcement runs to thousands of accounts a week, and OpenAI publicly backs three bills.
On 20 November 2026, two House committees published a joint report: Anysphere, Cursor's developer, told Congress it will explore phasing out Chinese open-weight models — while the same report states Cursor's flagship Composer 2 was built on Moonshot's Kimi K2.5. Airbnb and DoorDash are named too, with 12 policy recommendations.