AWS Outage of 02/21/2026: Direct Connect Routing Issues after US-EAST-1 – what happened?
On February 21, 2026 (CET), AWS Direct Connect in the US-EAST-1 (N. Virginia) region experienced an outage that was quite noticeable for some customers: previously advertised routes between VPC and On-Prem were temporarily unavailable (or no longer available). The result: connection problems for workloads in the VPC back to your own data center or On-Prem networks.
This is one of those incidents that doesn't look spectacular like 'everything is down' – but can still cause real damage in hybrid architectures because connectivity is the foundation of everything: databases, auth, legacy integrations, VPN fallback, monitoring, deployments.
Source: Symbolic representation of cloud and network infrastructure – fitting for Direct Connect & routing topics.
Brief Summary
- What: Connectivity issues with AWS Direct Connect to the US-EAST-1 region.
- When: 02/21/2026 (CET; night/early morning).
- Symptom: Some customers were missing previously advertised routes between VPC and On-Prem, leading to connection drops/timeouts.
- Scope: Affected were some Direct Connect customers / not 'AWS as a whole'.
- Status (according to updates): Root cause identified; AWS worked on mitigations in parallel, restored routes and observed significant recovery.
- Duration: roughly ~1–2 hours (depending on customer/route; status updates show gradual recovery).
- Takeaway: Hybrid resilience requires DX redundancy and VPN fallback, plus a clean routing strategy.
What happened?
According to the publicly available status updates, a Direct Connect connectivity issue in US-EAST-1 was initially investigated. Shortly thereafter, AWS confirmed that some customers were affected. Specifically: 'missing previously advertised routes' – meaning routes that were previously available were no longer (correctly) announced. AWS identified the cause and worked 'on multiple parallel paths' on the fix, including restoring the affected route advertisements.
Timeline (from status updates)
Note: Time indications are taken from publicly visible status updates; depending on the display, they may be shown in different time zones.
| Time (Status Update) | Update | Practical Significance |
|---|---|---|
| 21.02.2026 00:38 | Investigating (Direct Connect Connectivity Issues) | Initial Confirmation: Issue lies in DX connectivity to US-EAST-1. |
| 21.02.2026 00:59 | Confirmed – missing previously advertised routes | Hybrid traffic (VPC → On-Prem) can fail, even though internal AWS services appear 'green'. |
| 21.02.2026 01:22 | Restoration of non-announced routes in progress, root cause identified | Recovery is route-/control-plane-driven: it can come back at different speeds per connection/VIF. |
| 21.02.2026 01:33 | Early signs of recovery | First customers see stabilization – others may still be affected. |
| 21.02.2026 01:50 | Significant recovery | Most paths/announcements seem to be back; continue monitoring. |
Technical Analysis: Why 'missing routes' are so dangerous
At its core, Direct Connect is routing (BGP) plus physical/virtual paths. When routes 'disappear' or are no longer correctly announced, it often behaves like a classic network error in operation: timeouts, sporadic failures, asymmetric routing, blackholing – and the symptoms vary greatly depending on the application.
What's important is: such events can occur without your EC2 instances or AWS managed services themselves being broken. Your workloads are running, but they temporarily can't reach what they need (e.g., On-Prem DB, AD/LDAP, legacy APIs, MPLS networks).
What you can do (now) specifically – Checklist
- Have VPN fallback active: Site-to-Site VPN as a 'Break-Glass' path that can take over in case of DX issues (automated via routing priorities).
- Direct Connect Redundancy: Two physical connections/locations, separate devices/providers, separate VIFs. Avoid single-cable fate.
- Check Routing Design: Define BGP communities/LocalPref/AS-Path-Prepending cleanly so that failover doesn't 'stick'.
- Blackbox Monitoring: Not just CloudWatch, but synthetic checks over the hybrid path (VPC → On-Prem and back).
- Incident Runbook: 'DX down / Routes missing' as its own runbook category – including quick switching and clear communication components.
If you run hybrid frequently: it's worth consciously testing route tables, propagation, and failover once (chaos-style, but controlled). Because exactly there, minutes – not hours – decide.
Analysis: What has happened in the past
Yesterday's Direct Connect incident is a different class than major 'internet is burning' outages – but it fits a pattern: many outages are not 'compute down', but network/control plane problems that then cause symptoms everywhere secondarily.
Examples (briefly) – for comparison
| Date | Scope | What was roughly the issue? |
|---|---|---|
| 21.02.2026 | AWS Direct Connect (US-EAST-1) | Missing previously advertised routes (VPC ↔ On-Prem) → Hybrid connectivity issues. |
| Okt 2025 | Multiple AWS Services (US-EAST-1) | Major multi-service incident with cascading effects (incl. DNS/Load Balancer dependencies) – widely felt. |
| Dez 2025 | A Single Service (China Regions) | Reported outage of a cost management feature; discussion about internal AI coding tools (Kiro) & process/permission issues. |
These reviews are not meant to create drama – but to show: resilience is almost always architecture. And in hybrid setups, 'network + routing' is the area where a seemingly small incident immediately has a big impact.
Frequently Asked Questions
Was 'AWS down' or just Direct Connect?
The public updates describe a problem with AWS Direct Connect to the US-EAST-1 region. Some customers/connections were affected – not necessarily all AWS services worldwide.
Why are missing routes so noticeable?
Because many hybrid systems have central dependencies on-premises (AD/LDAP, legacy DB, fileshares, central APIs). When BGP announcements are missing, applications continue to run, but they cannot reliably reach their dependencies.
What is the best protection?
Redundant Direct Connect (separate paths) plus VPN fallback, along with synthetic monitoring over the actual hybrid path and a clear runbook for switching.
Does this have anything to do with 'AI tools'?
For the Direct Connect incident, there is no indication of AI tooling in the public updates. The AI connection stems from separate reports about other incidents (e.g., Dec 2025) and is only context here.
Conclusion
The incident of 02/21/2026 is a good example of why you shouldn't just understand outages as 'service down': in hybrid architectures, a disruption in route advertisements is enough, and suddenly a complete business flow is halted. If you take just one thing away from this incident, let it be this: plan the path (routing) just as redundantly as the workloads.