Data center power outages are defined as unplanned interruptions to electrical supply that cause partial or complete loss of computing infrastructure. The industry term is “unplanned downtime incident,” and the examples of data center power outages from 2026 alone reveal a consistent pattern: failures rarely stop at a single component. From AWS Northern Virginia to Google’s Ohio region, each incident exposed gaps in UPS maintenance, disaster recovery testing, and architectural design. Studying these data center outage case studies is the most direct path to building genuine operational resilience.
1. Real examples of data center power outages in 2026
The most instructive power failure examples in data centers share a common structure. A primary system fails, a backup system either fails or responds too slowly, and recovery takes far longer than planned. The incidents below are drawn from documented 2026 events and cover the full range of failure types that data center professionals encounter.
2. AWS US-EAST-1 Northern Virginia: cooling failure and thermal cascade
The AWS US-EAST-1 outage in may 2026 began as a thermal event, not a direct electrical fault. A cooling system failure at a Northern Virginia facility triggered a thermal incident that caused hardware failures and a prolonged power outage. The incident started at 5:25 PM PDT on may 7, with cooling systems not stabilizing until more than 20 hours later. That timeline meant EC2 instances and EBS volumes remained degraded or unavailable for an extended window affecting thousands of workloads.
The cascading effect on Coinbase illustrated a deeper problem. A Coinbase postmortem revealed that prolonged downtime resulted from architectural design that tightly coupled critical systems to a single availability zone. Backup systems failed partially, and recovery required manual disaster procedures. The root cause was not just the AWS outage itself. It was the assumption that automated failover would work without ever being tested against a full site power loss.
Disaster recovery procedures often fail in site-wide power outages due to hidden dependencies like control planes and identity management systems that are not fully redundant across regions. These dependencies are invisible until a real outage exposes them.
Pro Tip: Run a full site power loss simulation at least once per year. Tabletop exercises do not reveal the same dependency gaps that a live failover test does.
3. Google Cloud Ohio: UPS battery failure and manual bypass
The march 2026 Google Cloud outage in the Ohio region lasted over six hours. The root cause was a critical UPS battery failure that prevented the system from bridging the gap between utility power loss and generator activation. Engineers manually bypassed the UPS after approximately two hours to restore generator power. That two-hour gap caused widespread service degradation across multiple Google Cloud products.
Key lessons from this data center downtime incident:
- UPS batteries degrade silently. A battery that passes a routine load test can still fail under real fault conditions.
- Generator activation depends on the UPS holding load long enough for the transfer to complete. If the UPS fails first, the generator has nothing to bridge to.
- Manual bypass procedures require trained staff on site. Remote management is not a substitute when physical intervention is needed.
- Geographic scope matters. A single regional UPS failure affected services well beyond Ohio because of cross-region dependencies.
Pro Tip: Schedule UPS battery capacity tests under full load conditions, not just partial load. A battery that holds 50% load may collapse at 90%.
4. Equinix Telecity London: UPS fault and nationwide internet disruption
The june 2026 Equinix Telecity London incident demonstrated how brief a power failure needs to be to cause significant damage. A faulty UPS caused a 22-minute blackout that disrupted UK internet traffic, affecting 10% of customers. Power was restored shortly after the outage, and the majority of traffic rerouted through other paths. The speed of restoration was notable. The scale of impact from just 22 minutes was equally notable.
| Factor | Detail |
|---|---|
| Outage duration | 22 minutes |
| Customers affected | 10% of Telecity customer base |
| Cause | UPS fault at single London facility |
| Traffic response | Majority rerouted through alternate paths |
| Recovery | Power restored shortly after fault |
The Telecity incident reinforces that network resilience depends on traffic rerouting working correctly. Facilities that serve as major internet exchange points carry disproportionate risk. A UPS fault that would be a minor event at a standard enterprise data center becomes a national disruption at a carrier-grade facility.
5. Flexential Portland: the ground fault chain reaction
The Flexential Oregon outage is one of the clearest real-life data center outage examples of how multiple independent failures combine into a single catastrophic event. A ground fault triggered by utility maintenance initiated the sequence. The UPS batteries then failed to hold load. Generator restoration was delayed because the physical access control systems had no battery backup, preventing staff from reaching the generator room.
Outages rarely result from a single failure. The Flexential incident combined a utility event, a UPS failure, an access control design flaw, and an inexperienced overnight crew into one extended outage. Remove any one of those factors and recovery would have been faster.
Pro Tip: Audit every physical access control point in your facility. Any door that blocks access to critical power equipment must have a battery backup or mechanical override.
6. TikTok data center: transfer switch failures and undersized UPS
The TikTok data center power failure exposed a different category of risk. Cascading algorithm and delivery system failures resulted from an undersized UPS, transfer switch malfunctions, and inadequate redundancy. UPS failures caused load loss. Automatic transfer switch (ATS) failures caused slow transfer times and battery depletion before generators came online. The result was a failure chain that the redundancy architecture was designed to prevent but could not stop.
Legacy transfer switch malfunctions remain a common and underestimated cause of power outages. Aging ATS equipment that is not regularly tested produces slow transfers, contact welding, and control logic errors. Critical power professionals consistently identify ATS failures as a frequent contributor to outages that should have been prevented.
7. The anatomy of a cascading data center failure
Across these data center outage case studies, a clear pattern emerges. No single component failure caused extended downtime. Each incident involved a chain where the primary failure exposed a weakness in the backup layer, which then exposed a weakness in the recovery procedure.
The typical chain looks like this:
- Primary event: Utility fault, cooling failure, or equipment fault
- First backup layer: UPS fails to hold load or ATS transfers too slowly
- Second backup layer: Generator fails to start, or staff cannot access it
- Recovery layer: Disaster recovery procedures fail due to untested dependencies
The Flexential and TikTok incidents both followed this structure. So did the AWS thermal event, where the cooling failure led to hardware damage, which led to partial backup failures, which led to manual recovery procedures that took over 20 hours to complete.
8. Preventing data center power failures: overlooked risks
Preventing data center power failures requires addressing the gaps that standard checklists miss. The most common overlooked risks from 2026 incidents include:
- ATS testing gaps: Transfer switches must be tested under real load conditions, not just inspected visually. Contact welding and control logic errors only appear under load.
- UPS battery capacity under full load: Batteries rated for a given capacity degrade over time. Test at full load, not nameplate capacity.
- Physical access control without battery backup: Any door blocking access to generators, switchgear, or UPS rooms must have independent power.
- Untested disaster recovery plans: Modern distributed systems require thorough testing of failover and disaster recovery plans against real failure modes, not just partial scenarios.
- Single availability zone dependencies: Architectural decisions that couple critical services to one zone eliminate the redundancy that multi-zone designs are supposed to provide.
Pro Tip: Review your emergency power checklist against a total site power loss scenario specifically. Most checklists are written for partial failures, not complete blackouts.
Key takeaways
The most important lesson from 2026 data center power outage case studies is that backup systems fail when they are not tested under real conditions.
| Point | Details |
|---|---|
| Failures chain together | Every major 2026 outage involved multiple interacting faults, not a single root cause. |
| UPS batteries degrade silently | Test UPS capacity under full load; partial load tests miss real failure modes. |
| ATS failures are underestimated | Legacy transfer switches cause slow transfers and contact welding when not regularly tested. |
| Physical access matters | Access control systems without battery backup can block staff from reaching generators. |
| Disaster recovery must be tested live | Untested DR plans fail in site-wide outages due to hidden dependencies across control planes. |
What 2026 outages taught me about backup power assumptions
The pattern I keep seeing across these incidents is not a technology problem. It is a testing problem. Every facility involved had backup systems. Every facility assumed those systems would work. Most of them did not work as expected when the real failure arrived.
The AWS thermal event lasted over 20 hours not because AWS lacks resources, but because the failure mode exposed dependencies that had never been tested at scale. The Google Ohio outage required a manual UPS bypass after two hours because the automated recovery path failed. The Flexential incident was extended by a door that nobody had thought to put on battery backup.
What concerns me most is the architectural coupling issue the Coinbase postmortem revealed. Data center teams spend enormous effort on physical redundancy, then deploy software architectures that eliminate that redundancy by binding critical services to a single zone. The physical infrastructure is sound. The logical architecture undoes it.
The practical advice I give to any data center team is this: run your disaster recovery test against a complete site power loss, not a partial failure. Simulate the scenario where the UPS fails at the same time as the utility. Simulate the scenario where the ATS does not transfer. Those are the scenarios that produced the longest outages in 2026, and they are the scenarios most teams have never actually tested. You can read more about standby generator failure modes to understand where the gaps typically appear.
— Justin
Fuel reliability as a foundation for data center uptime
Real-world outage incidents confirm that generator fuel availability is as critical as the generator itself. A generator that starts but runs dry within hours provides no real protection.
Data Center Fuels, in partnership with Anytime Fuel Pros, provides on-site diesel delivery and DEF supply to data centers nationwide, including emergency resupply during active outage events. The service operates 24/7 and covers both planned fuel maintenance and unplanned emergency response. For facilities running extended generator operations during a power loss scenario, 24/7 fleet fueling keeps generators running through the full recovery window. Contact Anytime Fuel Pros to build a fuel resupply plan before the next outage, not during it.
FAQ
What causes most data center power outages?
Most data center power outages result from a chain of failures rather than a single fault. Common triggers include UPS battery failures, automatic transfer switch malfunctions, cooling system faults, and utility power interruptions.
How long do data center power outages typically last?
Duration varies widely based on failure complexity. The 2026 Google Ohio UPS failure lasted over six hours, while the AWS Northern Virginia thermal event took more than 20 hours to fully stabilize.
What is an ATS and why does it matter in power outages?
An automatic transfer switch (ATS) transfers electrical load from utility power to generator power when utility supply fails. Legacy ATS equipment that is not regularly tested can transfer too slowly or fail entirely, causing battery depletion before generators come online.
How do architectural decisions affect outage recovery time?
Architectural coupling of critical services to a single availability zone eliminates the redundancy that multi-zone designs provide. The Coinbase postmortem showed that tightly coupled systems extended recovery time significantly beyond the underlying AWS outage duration.
What is the most overlooked risk in data center power planning?
Physical access control systems without battery backup are a frequently overlooked risk. The Flexential Portland incident showed that unpowered access controls prevented staff from reaching generators, directly extending the outage.