Five Major System Outages in One Week, Led by Microsoft: Why Is Infrastructure Built for a Previous Era Struggling With the Pace and Demands of the AI Age?

Disruptions affected both classified communications and services used by Apple staff and consumers on 17 November. The NSA confirmed that NSANet, its internal network for transmitting classified information, had suffered an outage. It classified the incident as “Code Yellow,” the second-highest alert in a three-tier system. At Apple, an internal system update that day disrupted the employee portal, internal sign-in systems and remote VPN access, alongside consumer services such as iCloud, the App Store and Apple Pay. Numerous retail stores worldwide also faced interruptions to sales and repair operations.

TMGM Analysis: Financial Market News, Economic Calendar & Market Insights

Transaction processing at JPMorgan had been delayed a day earlier by a software bug in its on-premises hardware systems. More than 3,000 outage reports were logged by Downdetector; customers of Bank of America, Wells Fargo and other institutions reported related disruptions as well. Three days before the JPMorgan incident, a communications failure prompted the FAA to impose a nationwide ground stop. More than 5,100 flights were cancelled, and nearly 10,000 were delayed. The series of failures had started on 12 November with an outage affecting core Microsoft 365 services across at least 70 countries. Outlook, Teams and Copilot stopped working at the same time, and the disruption lasted more than two days.

Across six days, five critical systems suffered failures. Not one incident was identified as an attack or attributed to AI. Nevertheless, the account of these events as separate, non-malicious technical failures reveals a problem more difficult to address than an attack.


Microsoft: Why Did a “Capacity Upgrade” Cause a Two-Day Outage?

A routine rerouting of user traffic set the outage in motion, according to Microsoft’s preliminary customer report, released on 15 November, and Bloomberg’s reconstruction of events. During the evening of 11 November, an automated process sent Teams traffic to another server system that was handling Outlook traffic. Capacity had already been reduced by an unrelated issue. Together, these conditions activated a bug that had been dormant in the company’s data centres since July. As traffic increased, servers ran out of memory and started to crash.

The failure then spread through Microsoft’s automated failover mechanism, which sent more traffic to other servers it judged to be healthy. This compounded the damage. By the morning peak in business traffic on 12 November, more than 80% of Outlook servers had stopped responding. Another 15 hours passed while Microsoft identified the fault, before it started deploying a fix.

A mismatch between laboratory simulations and actual operating conditions emerged from the explanation given by Microsoft Executive Vice President Ramanna. The traffic patterns and hardware configurations used in testing did not mirror those encountered in practice. Tests generally concentrated on older, less powerful systems; on newer hardware, however, the bug responsible for the crashes was “more destructive”.

The technical account, however, is only part of the picture: the human element warrants equal scrutiny. A user identified as a Microsoft employee on Blind, the anonymous workplace forum, claimed that repeated layoffs and voluntary exits had affected Microsoft 365’s core infrastructure teams in recent years. With engineers who had maintained legacy systems for years having departed, those who remained were struggling to identify complex faults promptly. Rollbacks previously completed in around 45 minutes could now require five hours. At one stage during this outage, internal incident-management tools reportedly did not load correctly. Engineers consequently had to collaborate across departments late into the night, examining code and service architectures inherited from colleagues who had left.

Although Microsoft has offered no public response to these claims, they align with both the exceptional length of the disruption and the inability of rollback attempts to fix the problem.


Why Is Infrastructure Renewal Falling Further Behind AI’s Pace?

AI was not identified as the cause of any of the five incidents. The FAA and NSA ruled it out as a factor, and Microsoft expressly denied the involvement of “AI or agents”. There is nevertheless a distinction between “not caused by AI” and “unrelated to the AI age”.

Two trends are moving increasingly far apart. On one side, AI is driving more frequent changes and more intensive workloads. Traffic patterns and hardware configurations are shifting, software updates are gathering pace, and testing environments and incident-response capabilities are finding it difficult to keep up. On the other side are infrastructure renewal cycles. The copper cables and routers used by the FAA operate on timescales spanning decades, while the NSA’s networks develop over generations. Within Microsoft, the ability to maintain legacy systems is weakening quarter after quarter.

By enabling the discovery of large numbers of system vulnerabilities at unprecedented speed, AI puts companies and government agencies under immense pressure to address them immediately. Their response often involves deploying several patches simultaneously, placing substantial strain on systems that were never built for changes at this pace. The causal relationship is therefore not one in which AI produces vulnerabilities. Rather, AI speeds up change, while the systems required to absorb it were designed for a slower pace.

The clearest illustration is Microsoft’s experience. Dormant since July, the bug was activated by routine traffic rerouting and its effects were magnified by automated failover. Recovery then took unusually long because expertise in maintaining legacy systems had been lost. Although none of these individual stages involved AI, all took place amid the pressures of AI-driven growth in cloud services.