Summary
- Nebius marked two eu-west2 InfiniBand incidents resolved at 10:52 UTC on 14 October and 16:12 UTC on 15 October.
- A separate connectivity problem between eu-south1 and other regions ran in the incident log from 20:41 UTC on 14 October to 05:50 UTC on 15 October.
- The provider reported corrective actions but has not published a common cause, customer-wide downtime figure or complete technical post-mortem.
Nebius reported renewed network problems affecting European GPU clusters on 14 and 15 October, as well as a separate episode of disrupted connectivity between its eu-south1 cloud region and other regions. The three incidents were marked resolved in its service-status history, but the provider has not published a full technical account of their causes or customer impact.
An InfiniBand issue affecting GPU clusters in eu-west2 was recorded on 14 October. Nebius reported at 09:54 UTC that the cause had been identified and an onsite team was working on it. The incident was marked resolved at 10:52 UTC. The status record does not provide an earlier initial investigation timestamp for that episode, so those updates should not be treated as a measured total outage duration.
A second eu-west2 InfiniBand incident entered investigation at 05:34 UTC on 15 October. Nebius reported identifying the cause and taking corrective action at 09:32 UTC. At 13:13 UTC, the provider said corrective actions had been taken to stabilise affected infrastructure, but recovery and stability checks were continuing. The incident was marked resolved at 16:12 UTC.
Between the two InfiniBand entries, Nebius opened a separate incident affecting connectivity between eu-south1 and other regions at 20:41 UTC on 14 October. It warned that cross-region traffic could be disrupted or degraded. Subsequent updates said network block storage snapshot creation might be affected and that the Console and API could experience elevated latency. Nebius closed the incident at 05:50 UTC on 15 October.
GPU interconnect faults differ from regional networking problems
InfiniBand is a high-throughput, low-latency fabric widely used to interconnect accelerators in distributed computing clusters. Parallel workloads can depend on regular exchanges between GPUs, making network stability important to training and other tightly coupled computation. However, the October incident notices do not quantify the number of affected clusters, identify customer workloads that failed, or establish that every eu-west2 service was unavailable.
Traffic between cloud regions follows a different operational path from communication within a GPU cluster. An application that transfers data between regions, uses remote management interfaces or depends on a service in another location can experience degraded operations even when local compute remains available. Nebius’s own eu-south1 updates specifically identified possible snapshot and management-service effects; they did not state that all instances in the region had stopped working.
The provider used similar language about identifying a cause and preparing or applying fixes in the October entries, but that does not demonstrate a common fault. No disclosed post-incident report identifies a shared failed switch, cabling route, software change, network configuration or other component. The events should therefore be treated as three separate incidents rather than one prolonged outage with an assumed underlying defect.
New incidents follow an earlier cluster disruption
DataCentral reported an earlier eu-west2 InfiniBand incident on 3 October. Nebius investigated that problem from 02:30 UTC and marked it resolved at 08:59 UTC after a fix and monitoring. The events on 14 and 15 October are newly recorded incidents, not a restatement of the 3 October report. The official history does not establish whether the earlier and later incidents were technically related.
Recovery of a distributed workload also differs from restoration of network service. Even after a provider considers an interconnect stable, a training job interrupted earlier may require application-specific recovery or restarting from a saved checkpoint. That is a general characteristic of distributed computation, not a confirmed consequence for particular Nebius customers during these events.
Likewise, a status-page resolution time reflects the provider’s declared operational state. It is not necessarily the precise moment at which every affected application recovered. The record contains no independently audited customer downtime total or service-level compensation assessment.
By 16:12 UTC on 15 October, Nebius had marked the last of the three newly reported incidents resolved. The source establishes repeated disruption across two networking functions within two days, while leaving the affected customer count, physical cause and lasting corrective measures unquantified.
