Jakob Lange

← All insights

Microsoft's November outage: data residency is not resilience

On 12 and 13 November 2026 a dormant defect, a routine failover and a dip in capacity took Exchange Online down worldwide for almost a day, and the mechanisms meant to contain the failure spread it. European data never left Europe; Europe lost its email anyway. What decided who kept working was the failure domain.

At 01:07 UTC on Thursday, 12 November 2026, Microsoft’s automation moved Teams traffic onto the networking gateway that also sits in front of every Exchange Online mailbox. Microsoft calls this kind of failover normal operation, and for fifteen hours nothing happened. At 16:20 UTC, 17:20 in Frankfurt, Exchange Online went down, and within two hours the failure had spread across the gateway fleet worldwide. For most of the next day the NHS worked on paper, Euronext asked listed companies to phone in anything urgent, Norway’s members of parliament had no email, and Brazil’s Supreme Federal Court suspended procedural deadlines because lawyers could not receive the access codes for electronic filing. Germany’s federal administration, which runs Microsoft’s software on its own hardware, reported no major disruption.

Europe read the outage as a sovereignty story: one American company, one global infrastructure, a continent without email. Microsoft read it as a technical one and said that neither attackers nor AI agents were involved. Both readings are correct, and both miss the fact that matters for an architecture. The data of European customers never left Europe. Europe was without its email for nineteen hours anyway. Residency is a property of where data rests; resilience is a property of what fails together. This article reconstructs the outage from Microsoft’s preliminary post-incident report of 15 November, with the final report due on 20 November, and draws the consequences for organisations that depend on one suite.

Twenty-four hours without Exchange, in order

Twenty-four hours without Exchange Online, in UTC A timeline from Thursday 12 November 2026, 00:00 UTC, to Saturday 14 November, 06:00 UTC, with five lanes. Conditions: a defect in the gateway servers’ connection handling, latent since July, and about 14 percent of gateway capacity out of service since 11 November; at 01:07 Teams fails over onto the gateway it shares with Exchange Online. Gateway fleet: load rises through the day; at 16:20 Exchange Online goes down, the defect spreads across the fleet within two hours, and at the low point more than 80 percent of gateway servers do not respond; from 07:20 on Friday a staged recovery runs until 16:10, when customers are no longer actively affected. Microsoft’s response: all changes frozen at 19:00; servers restarted between 22:00 and midnight fail again under traffic; at 07:20 the July change is reverted. Europe: down from Thursday 17:20 to Friday 12:35 Central European Time, about nineteen hours. Mail backlog: at 12:00 on Friday the delivery timeout is extended to 48 hours, so that delayed mail is not returned as undeliverable; queued mail drains until the final update on Saturday at 04:15. Conditions Gatewayfleet Microsoft Europe Mailbacklog latent since July · ≈ 14 % of capacity out since 11 Nov 01:07 · Teams fails over to the gateway it shares with Exchange load rises from 01:07 16:20 · Exchange Online down · at worst > 80 % of servers staged recovery until 16:10 19:00 · all changes frozen 22:00–00:00 · restarted servers fall over again 07:20 · July change reverted Thu 17:20 – Fri 12:35 CET · about 19 hours 12:00 · delivery timeout extended to 48 h backlog drains · final update Sat 04:15 Thu 00:00 12:00 Fri 00:00 12:00 Sat 00:00 12–14 November 2026 · UTC (CET = UTC + 1)
Fig. 01 From a routine failover to a global outage and back. The conditions built up over months and one morning; Exchange Online was down from 16:20 on Thursday until a staged recovery that began with the revert at 07:20 on Friday; Europe lost Thursday evening and Friday morning; the mail backlog took until Saturday to drain. Times in UTC from Microsoft's preliminary post-incident report (MO1550821, EX1550543) as reported by the trade press; bars are drawn to the hour.

The preparation began in July, when Microsoft rolled out a change to the software that handles incoming connections on its Microsoft 365 networking gateway, the fleet that accepts users’ connections before they reach a mailbox. The change went out worldwide over several weeks and ran for more than three months without visible effect. On 11 November an unrelated issue took nearly 14 percent of the gateway’s capacity out of service. At 01:07 UTC on the 12th, Teams failed over onto the gateway it shares with Exchange Online, and load rose through the day.

At 16:20 Exchange Online raised its high-priority alert, the official start of the incident; Outlook, Outlook on the web, ActiveSync and the mail APIs were confirmed down minutes later. At 19:00 Microsoft froze every deployment and configuration change, and twenty minutes later its engineers established that the load came from servers being unavailable, not from more traffic. Kentik’s network measurements show Exchange Online traffic about 85 percent below expected levels from around 17:35, with no comparable drop for Teams, SharePoint or OneDrive.

The night went to attempts that did not hold: servers brought back with an alternative health check fell over again as soon as they received traffic. At 00:30 the team identified connection negotiation, the process that establishes secure connections, as the bottleneck. At 07:20 on Friday it began rolling out an update that reverted the July change, region by region, starting with the regions entering business hours. Europe was healthy again at 11:35, and from 16:10 no customer was actively affected. The mail backlog drained overnight under a delivery timeout extended to 48 hours, so that nothing bounced.

For a European organisation that meant Thursday from 17:20 and all of Friday morning: the window in which the NHS fell back on paper records, Euronext’s listing teams could not receive mail, eighty to ninety Danish municipalities were affected, according to the broadcaster DR, rail operators in Germany, the UK and Canada struggled to send timetables and tickets, and the German Digital Ministry spoke of a very large impact on German business.

Three harmless conditions made one global failure

Microsoft’s preliminary report names three causes, and the instructive thing about them is that none would have been an incident on its own.

The first is a latent defect. The July change was a fix made specifically for the gateway servers; Microsoft says neither the defect nor the fix touches Windows anywhere else. Under certain load it causes lock contention in an internal Windows process: threads queue for the same lock, the processor spends its time waiting and switching between them instead of working, and the server runs out of CPU and memory. Load tests missed it because the labs ran older hardware with fewer cores and a different traffic mix, and contention grows with the number of cores. The defect was invisible where it was tested and dangerous where it was deployed.

The second is the failover. About fifteen hours before the outage, Teams moved its traffic to the primary gateway that also serves Exchange Online, adding load and changing the traffic pattern. Microsoft describes this kind of failover as part of normal operations.

The third is the lost capacity. Nearly 14 percent of the gateway’s servers were out of service after an unrelated issue the day before. The service had headroom for that.

Each condition was within the margin the fleet was built for. All three together were not, and they met in the US business-day peak, around 16:00 UTC. A system can run like this for months, efficient and one ordinary trigger away from collapse, and nothing in its monitoring will say so, because the threshold that matters is not a number anyone has measured. It is set by a defect nobody knows about. Microsoft ruled out an attack and AI, and the mechanism needs neither: lock contention on many-core machines is as old as multiprocessor servers. What is notable is a test gap, production hardware that the lab did not represent.

The safety mechanisms carried the failure

What turned a defect into a global outage was the machinery built to prevent one. Microsoft’s traffic management did what it is designed to do: when a server stops answering, route its users to one that still does. Here, every healthy server had the same defect. Each redirected connection added load to the remaining servers, pushed the next one over its threshold and sent its users onward again. At the low point more than 80 percent of the gateway servers were unresponsive.

Redirection turns a margin into a cliff Line chart, illustrative. Horizontal axis: share of gateway servers out of service, from 0 to 90 percent. Vertical axis: load per remaining server, in percent of the normal peak, from 0 to 300. Two curves rise ever more steeply, because the same traffic is spread over fewer servers: one for Exchange traffic alone, starting at 100 percent, one with the Teams traffic added by the failover, starting at 115 percent. A dashed horizontal line at 125 percent marks where the latent defect wakes up; nobody knew where it was. Point 1: 14 percent of capacity out, Exchange traffic alone, 116 percent, below the line. Point 2: the Teams failover with all servers in, 115 percent, below the line. Point 3: both together at the US peak, 134 percent, above the line; from there every server that fails passes its load to the rest and the curve leaves the chart at about 62 percent out. A dotted line at 80 percent out marks the low point, where the remaining servers would carry about 5.8 times the normal peak, and where a server restarted into the fleet fails again. 0 100 % 200 % 300 % Load per server · % of peak 0 % 20 % 40 % 60 % 80 % Share of gateway servers out of service Exchange traffic alone with Teams traffic (failover) where the latent defect wakes up > 80 % down remaining servers wouldcarry ≈ 5.8 × the peak;a restarted one fails again each failure pushes its load onto the rest ① ② ③ ① ≈ 14 % of capacity out since the day before: within headroom② Teams fails over onto the shared gateway: within headroom③ both at US peak: over the line, and every failure adds load
Fig. 02 Each condition alone stays below the line where the defect wakes up; together they cross it, and from then on every server that fails raises the load on the rest. At the low point, with more than 80 percent of the fleet down, a server brought back alone meets about six times its normal peak and falls over again. Illustrative model, not Microsoft's data. The 1 / (1 − f) shape and the 14 percent are from the preliminary report; the Teams share (+15 percent) and the threshold (125 percent of the normal peak) are assumptions chosen so that each condition alone stays below the line, as the report states. Retries make the real curve steeper.

Two further mechanisms should have contained the damage, and neither did. The gateway is built from thousands of capacity units grouped into regions, whose boundaries act as bulkheads; a separate issue, not yet explained, kept them from holding, and the report says the outage would not have become global if they had. And the suite has a second gateway for exactly this case: Teams, SharePoint, OneDrive and the Office apps could fail out of the broken one, which is why Teams chats and meetings mostly worked. Exchange Online, for reasons of scale, has no second path.

Recovery was harder than the failure. A server restarted into a fleet that is 80 percent down meets the entire waiting queue at once and falls over again. Clients added to it: Kentik measured devices opening up to 70 percent more connections to Exchange than expected while responses fell to about a quarter of normal, the signature of retries. Rollback, flighting and feature flags are built for faults that show up soon after a change; this one was four months old. So the engineers built an update that returned the connection-handling code to its last known good state, then took server groups of half a region or more out of service at once and returned them in large batches, so that no small group faced the whole load alone.

None of this is new to people who study large systems. Bronson and colleagues described it in 2021 as a metastable failure: a system running in a vulnerable state is tipped by a trigger into a failure that persists after the trigger is gone, sustained by its own feedback loop, and the root cause is often a feature built to make the system more efficient or more reliable. Google’s SRE book names overload as the most common cause of cascading failure and prescribes the cure Microsoft eventually applied: cut load hard, then bring capacity back gradually. Amazon’s guidance on health checks warns that removing every server when a whole fleet fails its checks makes the outage worse. Microsoft’s first moves that evening, keeping servers from being removed and switching to an alternative health check, came straight from that textbook, after the cascade rather than before it.

Nor was it the first time this year. On 22 January Microsoft 365 was down across North America for almost ten hours, and the preliminary root cause was elevated load on capacity reduced for maintenance. Twice in one year, capacity taken out for maintenance met peak load; the final report has to explain why that margin was too small both times.

Your data stayed in Europe. The failure did not.

Residency draws its line around the data; the failure drew its line around the gateway Schematic in three regions, Americas, Europe and Asia Pacific. In each region, mailboxes and customer data sit at the bottom; above them, every connection passes through the networking gateway, drawn as small capacity units, most of them marked as failed. Dashed bulkheads separate the regions. A bar across all three regions reads: one change to connection handling, rolled out worldwide since July. A solid outline in the adverse colour encloses the bar and the gateway capacity units of all three regions and is labelled: failure domain on 12 November, all regions; the regional bulkheads did not hold. A dashed outline in the accent colour encloses only the European mailboxes and data and is labelled: EU Data Boundary, where the data rests. Outside the failure domain, on the right: an alternate gateway, to which Teams, SharePoint and OneDrive could move but Exchange could not; and Exchange Server on an organisation’s own hardware, which kept working, for example in Germany’s federal administration. Failure domain on 12 November: all regions one change to connection handling, rolled out worldwide since July AmericasEuropeAsia Pacific regional bulkheads (dashed) did not hold mailboxes, customer data mailboxes, customer data mailboxes, customer data EU Data Boundary: where the data rests alternate gateway Teams, SharePoint,OneDrive could move;Exchange could not Exchange Server on own hardware kept working e.g. Germany’s federal administration unresponsive at the low point still serving residency: drawn around the data failure: drawn around the gateway, worldwide
Fig. 03 Two boundaries, drawn differently. The EU Data Boundary encloses where European customers' data rests. The failure of 12 November enclosed the gateway in front of every mailbox, in every region, because one change had been rolled out everywhere and the regional bulkheads did not hold. What kept working sat outside that line: the alternate gateway Teams could move to, and Exchange servers on organisations' own hardware. Schematic, not Microsoft's topology: cell counts and positions are illustrative; 25 of 30 cells are drawn as failed to match the more than 80 percent at the low point.

Microsoft completed its EU Data Boundary in February 2025: the customer data of EU and EFTA customers is stored and processed within the EU and EFTA, with documented exceptions. On 12 November nothing inside it was breached, moved or read. Danish municipalities, the Norwegian parliament, Euronext and German companies lost their mail anyway, from Thursday 17:20 to Friday 12:35 Central European Time.

The reason is that a data boundary and a failure domain are different lines. Residency draws its line around where bytes rest. A failure domain is the set of things that fail together, and on 12 November it was drawn by three facts that have nothing to do with where data rests: every connection to every mailbox passes through the same gateway service; one change to that service had been rolled out to every region; and the bulkheads that should have kept a regional failure regional did not hold. The line was drawn around the planet.

In an earlier piece I described sovereignty in four dimensions: location, jurisdiction, control and portability. The November outage tested none of them. It tested a fifth question that the sovereignty debate rarely asks: what fails with it?

Microsoft’s own European commitments of April 2025 show the gap. Their resilience promise is about geopolitics: a board of European nationals for the European datacentres, a pledge to contest any government order to suspend services, continuity arrangements with European partners, a copy of the code in escrow in Switzerland. All of that addresses a kill switch. None of it addresses a defect shipped to every region.

The political reaction followed the familiar line. Reinier van Lanschot, rapporteur on the Cloud and AI Development Act in the internal market committee, said the outage showed how vulnerable Europe is, and the German Digital Ministry called for redundancies and alternatives. The act, proposed by the Commission on 3 June 2026 and now in committee, would grade cloud services for public buyers on four assurance levels built mainly around establishment, data location, ownership, control and exposure to third-country law. Those are real risks. But a European provider at the highest level, running one global control plane with one deployment pipeline, would have failed exactly as Microsoft did. Switching vendor moves the jurisdiction line; on its own it does not move the failure-domain line. The most precise sentence of the week came from the SPD digital politician Johannes Schätzl: digital sovereignty should be measured by whether “we can remain operational even during outages like this”.

Who did remain operational? Not the organisations with the most sovereign contracts, but the ones whose critical communication ran somewhere else. Germany’s federal administration, running Microsoft’s software on its own hardware rather than in the cloud, reported no major outage, and the Bundestag none although it uses Outlook. Exchange Server installations on customers’ premises were unaffected as long as their traffic stayed on premises; hybrid setups saw mail to and from the cloud pile up. Teams kept working because it could leave the broken gateway. Euronext moved urgent issuer contact to the phone, and traders in Taipei confirmed deals by fax. The common factor was a separate failure domain, sometimes built on purpose and sometimes merely old.

Email is not an application, it is a dependency

The outage also showed how much more than correspondence runs through a mailbox. Brazil’s Supreme Federal Court suspended deadlines because users of its electronic filing system were not receiving their access codes by email. One-time codes, password resets, approval workflows, supplier portals and alerting chains all use mail as a transport, and all of them stopped with it. The AI assistants that read mailboxes, Copilot and the Microsoft 365 connectors of OpenAI and Anthropic among them, inherited the outage, as any agent that acts on mail will.

The incident tooling went down with the incident. On the night of 12 November Microsoft’s impact list included Intune, Defender for Office 365, the admin centre apps and Graph: the tools an administrator would use to manage devices, protect mail and see what is happening. An organisation whose crisis communication, runbooks and contact lists live in Outlook, Teams and SharePoint has put its emergency plan inside its emergency.

The bill lands with the customer

What did a day without Exchange cost? The Anderson Economic Group, known for cautious estimates, put lost productivity at about 7 billion US dollars. Insurers will see a fraction of that: CyberCube expects insured losses below 100 million dollars, with 350 million as a conservative upper bound, while Morningstar DBRS estimates 300 to 500 million and the broker Acrisure 700 million. Many cyber policies cover only malicious events, and a non-malicious failure at an outside provider only where the policy says so; contingent business-interruption cover typically waits eight to twelve hours and needs proof of a financial loss. CyberCube’s view is that email sits outside the critical transaction path of most businesses: they lost productivity, not revenue.

The contract covers even less. Microsoft promises 99.9 percent monthly availability per service. A full day without Exchange Online, where downtime counts when users cannot send or receive mail in Outlook on the web, leaves about 96.7 percent for a thirty-day month. That entitles the customer to a service credit of 50 percent of the Exchange share of that month’s licence fee: on request, against a future invoice, and as the sole remedy. It is a refund of a fee, not compensation for a loss.

That is not a complaint; it is the allocation of risk. The probability of the failure sits with the provider, and almost all of its consequences with the customer. Whatever a day without the suite costs your organisation, you carry it yourself, so it belongs in your business impact analysis, priced, not in a service level agreement you expect to pay out.

Concentration risk now has a supervisor

For the financial sector this outage has an addressee. On 18 November 2025 the European Supervisory Authorities designated Microsoft Ireland Operations Limited as one of nineteen critical ICT third-party providers under DORA. Its Lead Overseer can examine it and issue recommendations, including on how it deploys updates and on avoiding single points of failure (Articles 33 and 35). The November outage is a test of what that oversight is worth.

The financial entities carry the obligations that decide whether a day like 13 November is an inconvenience or an incident: an ICT business continuity policy, a business impact analysis, yearly tests of continuity and response plans and tested crisis communication (Article 11); exit strategies that take a provider’s failure into account (Article 28(8)); and an assessment of concentration risk (Article 29). The difference between continuity and exit matters. An exit strategy is for months: how you would leave. Continuity is for hours: how you work while the provider is down. On 13 November only continuity counted, and it is the plan most often left untested.

Outside finance, NIS2 asks essential and important entities for business continuity, crisis management and supply-chain security (Article 21(2)(c) and (d)), binding in Germany since 6 December 2025. Public administration at Land level is still a gap, as the Berlin case showed, but the engineering does not depend on who enforces it.

Design for the day the suite is gone

None of this requires leaving Microsoft. It requires deciding which functions must survive Microsoft’s worst day, and giving those, and only those, a second failure domain. That is cheaper than a migration and more effective than a sovereignty label.

What the final report has to answer

Microsoft’s preliminary report is unusually open about what went wrong, and its repair items so far, from the reverted change to new guardrails on traffic distribution, address the symptoms of this incident. The final report, due on 20 November, should answer six questions, which double as the questions to put to any software-as-a-service provider:

  1. What defeated the capacity-unit bulkheads, and is it fixed?
  2. Will Exchange Online get a second gateway path, as Teams has?
  3. How will load tests represent production hardware, core counts and traffic mixes?
  4. How can a change that is months old be reverted in hours rather than a day?
  5. What stops traffic management from piling load onto servers that share a fault, and how much margin is reserved for maintenance after January and November?
  6. Can one region be isolated from another region’s failure — is there an operational counterpart to the EU Data Boundary?

What this means for European organisations

Three conclusions travel beyond this incident.

First, count failure domains, not datacentres. Residency, jurisdiction, control and portability are worth having, and none of them decided anything on 12 November. Add the fifth question to every sourcing decision and every sovereignty assessment: what fails with it, and what do we do in the hours until it comes back?

Second, diversity beats nationality. The organisations that kept working were not the ones with the most European contracts but the ones with a second path for the few functions that could not stop. A European provider with one global pipeline is a single failure domain with a European address.

Third, the next outage is already scheduled; only the date is missing. Microsoft has had two long ones this year, each caused by a margin that turned out to be smaller than its owners thought. The question for your next resilience test is not whether it will happen again. It is what you will be able to do at 17:20 on a Thursday when the mail stops, and through Friday morning until it starts again.

Sources and further reading

  1. Microsoft 365 Status: preliminary post-incident report released under MO1550821 and EX1550543 (opens in a new tab) — Microsoft, 2026
  2. Charles Lamanna: abridged root cause of the Microsoft 365 outage of 12 November (opens in a new tab) — Microsoft, 2026
  3. MO1550821: Users may be unable to connect to multiple Microsoft 365 services (service health record) (opens in a new tab) — Microsoft 365 Service Health, mirrored by Cloud Admin Hub, 2026
  4. Microsoft 365 outage: what went wrong, in plain English (opens in a new tab) — Office Watch, 2026
  5. Exchange Online outage: latent bug in July fix took down gateways (opens in a new tab) — heise online, 2026
  6. Microsoft explains root cause of Exchange Online outage (opens in a new tab) — UC Today, 2026
  7. Microsoft struggling to fix massive cloud outage that's taken down email and Intune (opens in a new tab) — The Register, 2026
  8. Microsoft's Outlook is back and it brought 18 hours of mail with it (Kentik and Cloudflare measurements) (opens in a new tab) — PPC Land, 2026
  9. Microsoft hit by global outages: what happened and why? (opens in a new tab) — Al Jazeera, 2026
  10. Microsoft says affected services fully recovered following worldwide outage (opens in a new tab) — Al Jazeera, 2026
  11. Hours-long Microsoft outage disrupts businesses, governments globally (opens in a new tab) — Reuters, via Insurance Journal, 2026
  12. Microsoft outage disrupts European businesses, public services (opens in a new tab) — Anadolu Agency, 2026
  13. Microsoft Outlook email outage (opens in a new tab) — Euronext Securities, 2026
  14. STF suspende prazos após falha nos serviços da Microsoft (opens in a new tab) — Migalhas, 2026
  15. Digitalministerium: „Sehr große Betroffenheit“ in Wirtschaft durch Microsoft-Panne (opens in a new tab) — Reuters, via onvista, 2026
  16. Microsoft-Störung: Keine größeren Ausfälle in Bundesverwaltung (opens in a new tab) — dts, via IT-Journal, 2026
  17. Microsoft-Störung: Digitalministerium rät zu Redundanzen (opens in a new tab) — ZDFheute, 2026
  18. Microsoft outage revives Europe's tech sovereignty push, Politico says (opens in a new tab) — The Next Web, 2026
  19. Insurance loss estimates for Microsoft outage range from $100M to $700M (opens in a new tab) — Claims Journal, 2026
  20. Microsoft outage could lead to $7B in losses (opens in a new tab) — Security Magazine, 2026
  21. Service Level Agreements (SLA) for Online Services (opens in a new tab) — Microsoft, 2026
  22. Microsoft 365 outage drags on for nearly 10 hours during bad night for North American infra (opens in a new tab) — The Register, 2026
  23. Metastable failures in distributed systems (HotOS ’21) (opens in a new tab) — Nathan Bronson, Abutalib Aghayev, Aleksey Charapko, Timothy Zhu · ACM, 2021
  24. Site Reliability Engineering, chapter 22: Addressing cascading failures (opens in a new tab) — Google, 2016
  25. Implementing health checks (opens in a new tab) — Amazon Builders’ Library, 2019
  26. Microsoft completes landmark EU Data Boundary, offering enhanced data residency and transparency (opens in a new tab) — Microsoft, 2025
  27. Microsoft announces new European digital commitments (opens in a new tab) — Microsoft, 2025
  28. The EU's Cloud and AI Development Act (CADA): towards a sovereignty-focused framework for cloud and AI services (opens in a new tab) — Hogan Lovells, 2026
  29. List of designated critical ICT third-party service providers under Article 31(9) DORA (opens in a new tab) — European Supervisory Authorities (EBA, EIOPA, ESMA), 2025
  30. Regulation (EU) 2022/2554 on digital operational resilience for the financial sector (DORA) (opens in a new tab) — Official Journal of the European Union, 2022
  31. Directive (EU) 2022/2555 on measures for a high common level of cybersecurity across the Union (NIS 2) (opens in a new tab) — Official Journal of the European Union, 2022

Contact

Start with a conversation.

No forms, no funnels. Write me a short note about your situation — I answer personally, usually within two working days.

Mon – Fri, 18:00 – 20:00 CET