Microsoft's Outage Likely Hit 450 Million. Headlines Said Thousands.

Robert wrote this post with the assistance of Claude Opus 5.5.

I was walking back to my office at 1:00 PM on Thursday. As I passed a member of our IT team, he mentioned that Microsoft 365 was having issues.

It was, and the issues ran well into Friday. Email stopped for state governments, hospital systems, universities, banks and airlines on four continents while Microsoft’s engineers worked through one fix after another. Microsoft didn’t report every customer restored until early Saturday morning.

By Thursday evening the outage had a number, and the number was everywhere: 130,000. It came from Downdetector, the crowdsourced outage tracker, and it went out on television, over the wires and in newspapers as a count of the people affected: 130,000 users who “lost access,” tens of thousands “affected.”

It was never that. Downdetector counts complaints that people choose to submit, and 130,000 was one reading from one page, Outlook in the United States, at 1:20 PM Eastern. By Friday morning, Downdetector’s owner said the site had received more than 7 million reports about Microsoft’s services from more than 90 countries. The one party able to count the people who actually lost service is Microsoft, and it hasn’t said.

This post is about the distance between those numbers. It’s also about a question the coverage raised for me: if the press gets my own field this wrong, what is it getting wrong in the fields I don’t know?

What Happened to Microsoft 365 on November 12 and 13

The outage began in Exchange Online, the service that carries Outlook mail for Microsoft 365’s business customers, and spread to everything that depends on it. Microsoft dates the start of its incident, EX1550543, to 11:17 AM Eastern on Thursday, and Downdetector’s Outlook reports began climbing around 12:25 PM. Microsoft’s first public post, at 12:54 PM, described trouble “in the Americas.” Thirteen minutes later it said the issue “may affect Microsoft 365 service connectivity for users in any region,” and at 1:26 PM it opened a broader incident, MO1550821, for the services that lean on Exchange.

What followed, in Microsoft’s own posts, was an afternoon and evening of fixes that didn’t hold (all times Eastern):

2:12 PM: “We’ve paused all deployments.”
2:31 PM: The load “isn’t driven by an increase in traffic, but by a number of servers being unexpectedly unavailable.”
3:19 PM: “We’ve not identified any recent change correlated with the start of impact.”
5:09 PM: Microsoft said “our actions taken thus far have not provided the anticipated relief” and that “a core process on affected servers is becoming overwhelmed.”
7:06 PM: “The servers we returned to service can become unresponsive again once they receive traffic.”
Evening: A process “responsible for establishing secure connections becomes overwhelmed,” and Microsoft had “begun reverting a recent network configuration change that correlates with the start of impact.”
10:13 PM: The changes applied to improve serviceability were “not taking as expected.”
Friday, 4:24 AM: Microsoft extended its email timeout to 48 hours “to help preserve delayed mail.”
Friday, 11:40 AM: Microsoft apologized and gave its first account of the cause.
Saturday, 1:09 AM: Microsoft said it had “restored service to the remaining impacted customers and regions.”

Friday’s statement is the only cause Microsoft has offered: “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.” Microsoft says it “will provide a full post-incident report once our review is complete,” and none had appeared publicly as of Saturday morning. When it does, it has three statements to reconcile: that no recent change correlated with the start of impact, that a recent network change did, and that the real fault came from an older upgrade.

The outage ran long enough to meet each region’s workday in turn. Ookla, which owns Downdetector, found that reports stayed high for about 20 hours, twice as long as any other large outage in the past year. Japan’s reports peaked around 9:00 AM local time on Friday, Singapore’s around 9:30, India’s around 10:15 and most of Europe’s between 8:30 and 9:15. From Microsoft’s start time to its final all-clear was about 38 hours.

Who the Microsoft 365 Outage Reached

Anything that depended on Exchange Online failed, for business and consumer customers alike. Microsoft’s status page listed Exchange Online, Teams, SharePoint Online, Microsoft 365 Copilot, Defender XDR, Purview and the admin center, and said features of other Microsoft 365 services that “interact with Exchange Online and email services are impacted.” Joining Teams calls and meetings and sending chats still worked; Teams calendar features didn’t. The consumer status card said the issue “impacts Outlook.com, OneDrive, OneNote, Microsoft To Do, Teams, and Microsoft 365 for the web.”

Governments lost email on three continents. Massachusetts texted its employees that “The Commonwealth’s email and other Microsoft services are currently unavailable.” Michigan’s state technology department posted that state government was “currently experiencing technical issues w/ email & other Microsoft apps.” Singapore’s Government Technology Agency said government email was down. Japan’s Digital Agency couldn’t receive email, Immigration New Zealand paused visa processing, the U.S. House of Representatives alerted its staff, and in Australia the federal police and the intelligence agency ASIO were affected.

Hospitals felt it in ways no outage chart can show. ChristianaCare in Delaware warned that the outage was “impacting email and other communications such as faxes and pagers that rely on email addresses.” Unity Health Toronto declared a Code Grey, its alert for a loss of infrastructure. An NHS spokesperson said local services in England were “using other IT systems or paper records where possible,” and some UK hospitals rescheduled procedures. Michigan Medicine posted that the outage was “impacting access to Outlook and other Microsoft applications,” and Santé Québec said appointment bookings, test results and prescriptions could have been affected before contingency plans took hold. For those of us in healthcare IT, ChristianaCare’s notice is the one to remember: fax and paging systems that run through email addresses are easy to forget until the mail stops.

The outage also made good bait. By early Saturday, Microsoft was warning of a phishing campaign whose messages posed as outage notices, with subject lines such as “Outage” followed by a six-character code.

How Many People Use Microsoft 365 and Outlook.com

Microsoft last gave a total in January: “over 450 million” paid Microsoft 365 commercial seats. Amy Hood, its chief financial officer, gave the figure on the January 28 earnings call, and Microsoft has reported year-over-year seat growth of 5 to 6 percent in each quarter since. The other official figures:

Measure	Latest official figure	Source
Microsoft 365 paid commercial seats	Over 450 million	Amy Hood, earnings call, January 28, 2026
Microsoft 365 consumer subscribers	Nearly 95 million	Satya Nadella, earnings call, April 29, 2026
Outlook.com	Over 400 million active email accounts	Microsoft advertising documentation, unchanged since at least 2023
Teams monthly active users	More than 320 million	Nadella, October 2023, repeated by Microsoft in August 2025

These figures overlap, so they can’t simply be added. And the scale figure most stories used was out of date. AP’s “Over 430 million people use Microsoft 365 apps, Microsoft said on its website” traces to a Microsoft blog post whose footnote cites the April 2025 earnings call, where 430 million was the paid-seat count. Microsoft has reported more than 450 million since January.

Microsoft hasn’t named a single region, plan or kind of customer that escaped, and its own status pages listed business and consumer services alike. Until it says otherwise, it’s far more reasonable to assume the outage reached most of its 450 million-plus paid seats, and consumer accounts besides, than anything close to 130,000.

The only independent estimate I’ve found comes from Anderson Economic Group, a Michigan think tank. Its president, Patrick Anderson, told CNN that as many as 26 million users on Thursday and 51 million on Friday may have had their operations disrupted in ways that caused an economic loss, and the firm put the lost productivity at about $7 billion. That’s a narrower measure than everyone who lost email, and it’s still hundreds of times the number on Thursday’s screens.

What Downdetector Counts, and What It Doesn’t

Downdetector counts reports, not people. Its methodology page says it shows a possible issue “when user reports are significantly higher than typical levels for that service,” judged against “the average number of reports for that given time of day,” and that it limits “how often repeated reports from the same user are included in the public report totals.”

It’s a smoke detector, and a good one, built for newsrooms. In 2012, two employees of IDG, the publisher of Computerworld and CIO, started it because reporters calling companies about outages would often get no response or wait hours for one. “We thought, wouldn’t there be ways to automate this so we didn’t have to check with the press office and we could get the data directly ourselves?” co-founder Sander van de Graaf told Bloomberg.

A smoke detector tells you there’s a fire. It can’t tell you how many rooms are burning. Every report requires someone to notice a problem, know the site exists, find it and click a button. I doubt a significant share of the public knows what Downdetector is. The people who do, IT staff like me, check it in the first five minutes of an outage to confirm the problem isn’t ours, often without registering that we’ve been affected too. Inside a company, people with no email call the help desk. They don’t file reports with a website.

The reports scatter, too. Ookla found that GoDaddy, which resells Microsoft 365 email to small businesses, drew more than 60 times its usual volume of reports, and Gmail more than 40 times, possibly from senders whose mail to Outlook users bounced. None of those count toward Microsoft.

And the number on the chart isn’t a running total. It rises and falls with the pace of new reports, which is how Thursday afternoon’s 130,000 became “around 50,000” by evening. A falling line means fewer new complaints, not fewer people without email.

Even by Downdetector’s own data, 130,000 was a sliver. Ookla, part of Accenture since June, counted more than 7 million reports across Outlook, Microsoft 365, Teams, Copilot and the Microsoft Store by 11:00 AM Friday, the most for any outage in the past year, and Reuters later reported more than 7.4 million. The figure that led Thursday’s coverage was under 2 percent of that.

Ookla has made the users-for-reports mistake itself, and fixed it. Its analysis of the 2024 CrowdStrike outage now ends with an editor’s note: “This article was updated August 6 from 5 million users to 5 million reports to better reflect how Downdetector tracks the number of outage reports submitted.”

One big outage lets us check a Downdetector count against a real one. When AT&T’s wireless network failed on February 22, 2024, Downdetector’s reports peaked at 73,502, and Ookla wrote that the outage left “thousands of users” without service. The FCC’s investigation found that it affected “more than 125 million registered devices,” blocked more than 92 million calls and stopped more than 25,000 calls to 911. That’s about 1,700 devices for every report at the peak. The FCC got its number from AT&T, because a regulator asked.

Systems built on voluntary reports usually say what they can’t do. The FDA’s adverse-event dashboard puts it plainly: “Rates of occurrence cannot be established with reports.” Downdetector’s methodology page has no such sentence.

How the Media Reported the Microsoft 365 Outage

Across about 70 stories from more than 55 outlets, much of Thursday’s coverage turned Downdetector reports into people. The errors fall into a few patterns, and some outlets appear in more than one.

Reports became people who lost access.

NBC News NOW, on air: “Some 130,000 users lost access to such things as their email,” and later, “around 50,000 users still remain without access.”
NBC Chicago: “Reports peaked at just over 130,000 people experiencing issues.”
NBC DFW: “More than 120,000 Microsoft users could not access email and other components.”
USA Today: the outage “affected tens of thousands of users across the United States.”
Windows Central: “tens of thousands of users are affected by the outage for each Microsoft service.”
CBC: “Thousands of Canadians were affected by the outage,” on the strength of “more than 20,000 reports.”
NDTV: “Thousands of users across the world have been unable to access Outlook.”

Headlines said “thousands.” Reuters ran “Microsoft 365, Outlook down for thousands of US users, Downdetector shows” over a story counting “more than 97,000 incidents of people reporting issues,” with the caveat “The actual number of affected users may vary.” It has used the same headline formula for X in February, Spotify in September and Microsoft 365 in October 2025. Tom’s Guide (“down for thousands”), How-To Geek (“leaves thousands struggling”), SFGATE (“thousands of users still affected”), The Detroit News (“impacts thousands”) and Al Jazeera’s first headline (“hits tens of thousands of users”) said the same.

“Users reported”: the wire version. AP’s Thursday story said “Nearly 130,000 users on outage tracker Downdetector reported problems accessing Outlook just after 1 p.m. EST.” Nothing in that sentence is false, but with no word on how many people use Outlook, it reads as a head count. AP’s reach carried it to GBH in Boston, the CBC, India’s Economic Times, the Sydney Morning Herald, the Idaho State Journal and Alaska’s Daily Sitka Sentinel. Euronews and The Wall Street Journal (“hundreds of thousands of users in more than 70 countries reported problems”) followed the same pattern, and the Times of India made the U.S. Outlook figure “worldwide.”

By Friday, the scale changed but the record didn’t. NBC’s Morning News NOW reported “about 450 million users still struggling to access their email,” a day after the same network told viewers that “some 130,000 users lost access.” No NBC report I found reconciled the two figures, and Thursday’s was never corrected. The New York Times described the outage as “affecting millions of users, according to Downdetector,” crediting a count of users to a site that counts reports.

Some got it right, on deadline. CNN wrote that Downdetector’s 130,000 “understated the scope of the problem, since most people do not report their experiences,” and that “As many as 430 million users were reportedly affected at some point.” TechRepublic called the figures “user-submitted reports rather than confirmed counts of affected individuals,” and Orlando’s WFTV said they were “rather than a confirmed count of affected customers.” NPR wrote, “While Microsoft did not say how many users were affected, Microsoft 365 apps have around 430 million users globally.” The BBC even flagged possible “under-reporting in some countries, where Downdetector is among the websites blocked due to censorship.” On Friday, Al Jazeera added: “The true number is likely far higher, since the figure only counts people who visit the site to report an issue.”

By Friday morning, AP had rewritten its main story: “The total number of people affected is not clear, but the company said in January that it had more than 450 million paid Microsoft 365 commercial subscriptions.” That’s the sentence the coverage should have started with.

Two things were missing across the whole catalog. None of the stories reported asking Microsoft how many customers lost service and getting an answer, though several noted that Microsoft hadn’t said. And none of the revisions came with a correction notice: AP, USA Today, Al Jazeera, Euronews and Reuters all changed their wording between versions without telling readers.

The Questions an Undercount Lets Microsoft Skip

Microsoft’s own contract measures outage downtime by the number of users affected. Under its service level agreement for online services, Exchange Online downtime is counted in user-minutes: “the sum of the length (in minutes) of each Incident that occurs during that Applicable Period multiplied by the number of users impacted by that Incident.” Microsoft’s contract depends on the very number nobody reported.

I don’t think reporters set out to protect Microsoft, and the effect doesn’t depend on intent. A public record that says “thousands” spares the company the questions that “hundreds of millions” would force:

How many tenants and users lost service, and for how long? Microsoft has the telemetry, and its own agreement needs the answer.
What were the “earlier network capacity upgrade” and the “recent network change”? And how do they square with Thursday afternoon’s statement that no recent change correlated with the start of impact?
Why did a component that checks server health take so many servers out of service, and why did restored servers fail again once traffic returned?
Will Microsoft credit affected customers without making each one file? Today it doesn’t.

The credits are modest. Below 99.9 percent monthly uptime, a customer gets back 25 percent of that month’s fee for the affected service; below 99 percent, 50 percent; below 95 percent, the whole month. In a 30-day month, 99 percent allows 7.2 hours of downtime, so a tenant that lost mail for most of a day lands in the second tier. But nothing happens unless the customer files a claim, by December 31 for a November outage, listing “the number and location(s) of affected users.” Organizations that buy Microsoft 365 through a reseller have to go through that reseller instead: Microsoft’s partner rules have the reseller file the claim, with an email from the customer’s own domain as proof. The credits are the “sole and exclusive remedy,” and they “will not be awarded to compensate for any other forms of loss, including but not limited to lost revenue, operational costs, or any indirect losses.” The one party asked to produce a head count is the customer.

AT&T’s real number became public because a regulator asked for it. For this outage, no story I found reported putting that question to Microsoft, and Microsoft hasn’t volunteered an answer.

If the Press Gets Tech Wrong, What Else? The Gell-Mann Amnesia Effect

This week left me with a question. If mainstream media misreport the one field I know professionally this badly, how do I know they aren’t doing the same with politics, the economy and everything else?

Put Thursday’s error in political terms. If all of Microsoft’s more than 450 million paid seats were hit, the 130,000 that led the coverage was about 0.03 percent of them, roughly one user in 3,500; the coverage missed 99.97 percent of the people affected. The same fraction of the 155,238,302 votes cast for president in 2024 is about 45,000. Calling 130,000 the number affected was like polling 45,000 people before that election and then announcing that 45,000 Americans voted, when more than 155 million actually did.

Michael Crichton called this the Gell-Mann Amnesia effect, after the physicist Murray Gell-Mann: you catch the errors in a story about your own field, then trust the rest of the paper anyway. “You turn the page, and forget what you know,” he said in a 2002 speech.

The errors are real. Since 1936, studies that asked the people named in news stories to check them have found errors in 41 to 61 percent of stories, and wrong numbers rank among the most common. But the “amnesia” is overstated. In a 2003 Gallup poll, 78 percent of people who had been part of a news story said the coverage of it was accurate, and across 23 European countries, the outlets the public trusts most are largely the ones media experts rate most accurate. Readers tell outlets apart.

The press doesn’t get everything wrong. It gets particular things wrong, and predictably: numbers, technical terms and causes, most of all in breaking news written fast by generalists. Thursday’s error belongs to the most common family, a number that measures one thing reported as if it measured another, and that family turns up on every beat. Crimes reported to police aren’t all crimes; in 2024, only about 48 percent of violent victimizations were reported. The question is always the one so few reporters asked on Thursday: what does this number actually count?

Is all this undermining journalism? Gallup found trust in the mass media at a record low of 28 percent last year, and silent fixes don’t help: a reader who saw Thursday’s version never learns that it changed. A 2023 study by Brendan Nyhan and colleagues found that published retractions make readers’ beliefs more accurate and cost the outlet only a little trust. The outlets that rewrote their stories this week without saying so passed up the cheapest credibility there is.

None of this takes a statistics degree. Call a report a report. Give the user base, so readers can see the scale. Ask the company how many customers lost service, and when it won’t say, print that. And when a number turns out wrong, say so where readers can see it.

Downdetector did its job on Thursday. Within minutes it told the world that something big had broken at Microsoft. It was never built to say how big, and the one company that can say hasn’t. Until it does, every story that settles for “thousands” does Microsoft’s work for it, whether anyone means it to or not.

Sources

Microsoft

Downdetector and Ookla

Impacts

Coverage cited in the catalog

Media accuracy and trust