Virtual private clouds
Virtual Private Clouds () are isolated networks used by CoreWeave Kubernetes Service (CKS) clusters to securely communicate with each other and the internet while maintaining privacy and security by defining the network isolation for attached CoreWeave products. VPCs allow for the creation of segregated and secure networks for your resources within the CoreWeave Cloud infrastructure. This isolation ensures that your data and applications remain private and protected. VPCs are essential for organizations using cloud computing while maintaining strict control over network environments.HPC Interconnect
HPC Interconnect provides high-performance, RDMA-enabled fabrics for GPU-to-GPU communication. CoreWeave offers two distinct fabrics: NVIDIA Quantum InfiniBand and NVIDIA Spectrum-X (RoCE). Both are optimized for the high throughput and low latency required for large-scale model training and other demanding parallel workloads. GPUDirect RDMA enables direct access between the main memory of two computers without involving the operating system, cache, or storage, which reduces latency and increases throughput.Direct Connect
Direct Connect lets you connect directly to CoreWeave through private, dedicated links. If you use Equinix or Megaport, you can use your existing connections when establishing direct connections to CoreWeave. You don’t need to add additional dedicated physical ports or establish a network presence in our regions.IP addresses
CoreWeave supports both Public IPv4 addresses and Bring Your Own IP (BYOIP). Public IPv4 addresses ensure that your Services are accessible over the internet, which is necessary for deploying web applications and Services that need to be available globally. BYOIP lets you use your own addresses on CoreWeave Cloud, which is useful for customers who have existing IP address spaces they want to use for workloads on CoreWeave.Ingress Service
The Ingress Service exposes Services running on CKS to the public internet usingLoadBalancer Services and public IPv4 addresses. You can also request a fully qualified domain name (FQDN) or wildcard record to provide a stable, human-readable address for your publicly exposed Services.
Outbound traffic and IP allowlisting
Outbound traffic from your Pods leaves the cluster through a NAT that uses a stable, documented set of public IP ranges per Availability Zone. When an external service needs an allowlist, such as MongoDB Atlas, a partner API, or a firewalled database, allowlist theNAT Egress ranges for the Availability Zones your workloads run in.
Each region page lists its NAT Egress ranges in the Network information table. See Regions and Availability Zones for the full list, or Does outbound traffic have a stable public IP? for a summary.
Bandwidth and data transfer
Per-instance Ethernet bandwidth is bounded by the NIC of the underlying instance type, with no separate per-public-IP throttle. Each instance’s network speed appears in its specification table on the GPU instances and CPU instances pages. For high-bandwidth distributed workloads such as multi-Node training, use the dedicated HPC Interconnect rather than the standard Ethernet path. See What are the bandwidth limits per instance or public IP? for details. Data transfer between CoreWeave Regions is free. See Region features.Private cluster access
By default, self-service CKS clusters expose the Kubernetes API over the internet. Private clusters have no public API endpoint, so you reach the API through a proxy:- Tailscale VPN provides a CoreWeave-managed private access path.
- Traefik ingress, often paired with Direct Connect, provides an in-cluster proxy over dedicated private links.