Definitions for CoreWeave Infrastructure terms used across the platform and documentation
Use this page as a reference for terminology that appears throughout CoreWeave Infrastructure products and documentation, with entries listed alphabetically.
Attribute-Based Access Control (ABAC)
A policy-based access control model where access is determined by evaluating attributes (for example, user, resource, environment).
Audit logging
Tracking and recording system events and user activities for security analysis and compliance.
Availability Zone (AZ)
An AZ is a partition within a Region that hosts one or more data halls.
AZs are physically and operationally independent from each other to prevent failures from propagating across them.
For example, Availability Zone A in the US-EAST-05 Region is named US-EAST-05A.
CoreWeave categorizes AZs as General Access or Dedicated Access.See also: Geo, Super Region, Region, About Regions and Availability Zones
Border Gateway Protocol (BGP)
A standardized exterior gateway protocol that exchanges routing information between different autonomous systems on the internet.
BGP is the protocol used to make core routing decisions on the internet.
Classless Inter-Domain Routing (CIDR)
CIDR is a method for allocating IP addresses and routing IP packets. See also: CIDR notation.
cloud-init
An industry standard method for cloud instance initialization.
CoreWeave uses cloud-init to pass configuration data to Nodes at boot time.
Cluster
A group of interconnected computers working together as a single system.
Control Plane
In Kubernetes, the Control Plane is the set of components that manages the cluster’s desired state, scheduling, and API.
In CoreWeave networking, the Control Plane also refers to the routing-control layer of the EVPN-VXLAN overlay. EVPN advertises routes, while the VXLAN Data Plane forwards encapsulated packets.See also: Data Plane, Ethernet Virtual Private Network (EVPN)
Cordon
A cordoned Node is one the scheduler doesn’t place new work on, while existing Pods keep running. Nodes are typically cordoned for failure investigations, maintenance, upgrades.See also: Node condition, Node lifecycle automation, Node cordoning
CoreWeave AI Object Storage
CoreWeave AI Object Storage is CoreWeave’s purpose-built, S3-compatible object storage for AI training data, checkpoints, and model weights.
In Grafana dashboard names and some telemetry labels, the product appears as CAIOS.See also: LOTA, Multipart upload, Storage
The CoreWeave Node lifecycle controller coordinates Node lifecycle state transitions, including reboots and moves to triage. It’s part of CoreWeave’s Node lifecycle automation. The CoreWeave Node lifecycle controller is separate from the SUNK Node Controller, which synchronizes data between Kubernetes Nodes and slurmd Pods.See also: Node state transitions in CKS, Reboot Nodes, SUNK Node Controller
CPU
A Central Processing Unit (CPU) is the hardware within a computer that carries out the instructions of a computer program by performing basic arithmetic, logical, control, and input/output operations specified by the instructions.
Custom Resource (CR)
A Custom Resource is an instance of a . It’s the actual object created with the Kubernetes API.See also: Custom Resources
Custom Resource Definition (CRD)
CRDs are an extension of the Kubernetes API that lets you define custom resources and controllers that aren’t part of the core Kubernetes API.A CRD is the blueprint for a type of .See also: Custom Resources
Data Plane
The Data Plane forwards application and network traffic.
In Kubernetes clusters, that traffic is the workload path.
In CoreWeave’s EVPN-VXLAN networking overlay, the VXLAN Data Plane forwards encapsulated packets using routing information advertised by the EVPN Control Plane.See also: Control Plane, EVPN-VXLAN integration
Data Processing Unit (DPU)
A programmable infrastructure-on-a-chip that combines an array of ARM-based CPU cores, acceleration engines, and a high-performance network interface.
DPUs function as a “computer-in-front-of-a-computer” and are fully isolated from the host’s CPU.
On CoreWeave Nodes, DPUs provide network, storage, and encryption functions that Nimbus uses for tenant isolation, VPC membership, and Node lifecycle management.See What Is a DPU? at NVIDIA’s Blog.See also: Nimbus
Day 0
The phase in the lifecycle of a CoreWeave Node where it’s initially configured after powering on.
Day 1
The phase in the lifecycle of a CoreWeave Node where it’s intensively validated before delivery to a customer.
Day 2+
The phase in the lifecycle of a CoreWeave Node once it has been delivered to a customer, and is continuously monitored and validated by CoreWeave.
Dynamic Host Configuration Protocol (DHCP)
A network protocol that automatically assigns IP addresses and other network configuration settings to devices on a network.
Encryption at rest
Protecting stored data using encryption mechanisms.
Encryption in transit
Securing data as it travels across networks using protocols like TLS.
Ethernet Virtual Private Network (EVPN)
EVPN simplifies Control Planes for Virtual Private Network (VPN) services by extending Ethernet (Layer 2) services over a broader network, typically an IP/MPLS network.
EVPN supports multi-tenancy, so different customers’ networks can share the same physical infrastructure while keeping their traffic separate and secure.
EVPN is used in interconnect scenarios and for integrating distributed regional and campus networks.
EVPN brings the advantages of traffic balancing and flexible deployment from IP VPNs into the Ethernet domain.
EVPN Type 5
A Type 5 deals exclusively with IP route advertisement, differentiating it from other types (such as Type 2) that include address advertisement.
EVPN-VXLAN integration
Ethernet VPN-Virtual Extensible LAN (EVPN-VXLAN) combines ’s Control Plane with ’s Data Plane.
Combined, these technologies create virtual Layer 2 networks that span Layer 3 boundaries in large-scale environments.
Fabric (backend network)
A fabric is a distinct high-performance backend network, such as an InfiniBand or RoCE fabric, that connects Nodes for low-latency, high-throughput communication. CoreWeave labels each Node with the fabric it belongs to, so you can keep the Nodes of a single distributed job within one fabric. The fabric label is one of several topology labels CoreWeave exposes for placement.See also: Superpod, Leafgroup, InfiniBand and RoCE labels
Federated identity
An authentication method that lets users log in across multiple systems using a single identity (through OIDC and SAML).
Geo
A Geo is a large geographic area, such as a continent or a large country.
CoreWeave uses Geos to manage global infrastructure, meet regulatory requirements, and plan large-scale redundancy and disaster recovery.
For example, all locations in the United States reside in the US Geo.
A Geo contains one or more .See also: Super Region, Region, Availability Zone (AZ), About Regions and Availability Zones
GPU
A Graphics Processing Unit (GPU) is a parallel processor designed to accelerate vector and matrix operations. GPUs are commonly used in high-performance computing and machine learning applications.
GPUDirect RDMA
GPUDirect RDMA is a technology that enables remote direct memory access () transfers between GPUs and other devices without involving the operating system or CPU.See also: Remote Direct Memory Access (RDMA)
Hard Disk Drive (HDD)
A hard disk drive (HDD) is a non-volatile data storage device that uses spinning magnetic platters.
HDDs are slower than drives, but typically offer higher capacity at lower cost.
On CoreWeave, SUNK shared-storage examples use an HDD volume such as data-hdd mounted at /mnt/hdd for high-capacity bulk storage.See also: Non-Volatile Memory Express (NVMe), Shared storage on SUNK
Helm-installed component
A Helm-installed component is software you install into your own cluster, often from a CoreWeave-provided Helm chart, and that you own and operate. CoreWeave provides the chart and documents the safe configuration, but the running release is yours to manage. This is distinct from a managed component, which CoreWeave installs and reconciles. When deciding who fixes a broken component, a Helm-installed component is your responsibility, while a managed component is CoreWeave’s.See also: Managed component, What CoreWeave manages
HPC Verification
HPC Verification is CoreWeave’s proactive, on-Node testing framework. At the top of each hour, if a Node is idle, meaning no customer Pod holds any of its GPUs, CoreWeave runs a test that uses all of the Node’s GPUs to catch problems that passive monitoring can’t detect, such as silent data corruption, performance regressions, and thermal deficiencies. The test Pods run at a low priority and stop as soon as Kubernetes or Slurm schedules a workload that needs the Node’s GPUs. The tests aren’t billable and can’t be turned off. For test duration and completion checks, see Test schedule.HPC Verification is separate from Node Health Check (NHC): HPC Verification is a fleet-wide CoreWeave test that runs on idle Nodes, while NHC is a tool that SUNK runs on a node after a Slurm job. On SUNK clusters, a passing HPC Verification test is what returns a node whose drain reason carries sunk:verify-undrain to service.See also: Node Health Check (NHC), Day 2+, HPC Verification
Identity and Access Management (IAM)
A framework for managing user identities and access permissions across cloud services.
Identity provider (IdP)
An identity provider (IdP) is an entity that stores and serves user authentication information as an authentication service. IdPs can validate user identity for other services, such as cloud applications.
InfiniBand
A high-performance network architecture that provides high throughput and low latency, commonly used in high-performance computing environments.
Input/output operations per second (IOPS)
IOPS (pronounced eye-ops) is an input/output performance measurement used to characterize computer storage devices.
Internet Protocol version 4 (IPv4)
IPv4 is the fourth version of the Internet Protocol (IP), and one of the core protocols of standards-based internetworking methods in the internet and other packet-switched networks.
Internet Protocol version 6 (IPv6)
IPv6 is the most recent version of the Internet Protocol (IP), the communications protocol that provides an identification and location system for computers on networks and routes traffic across the internet.
Internode Memory Exchange (IMEX)
Internode Memory Exchange (IMEX) is NVIDIA’s software for coordinating GPU memory import and export across OS instances in an NVLink multi-Node deployment. On CoreWeave Kubernetes Service (CKS), customer workloads access IMEX channels through Dynamic Resource Allocation. See IMEX overview.
iPXE (i Preboot eXecution Environment)
The i Preboot eXecution Environment is an open source network boot firmware.
It provides a full implementation with additional features for network booting.
iPXE is commonly used in cloud environments with complex configurations and network installations to boot servers over the network.
From the official FAQ:
Q: What does the “i” in “iPXE” stand for?
A: It doesn’t.
Key Management System (KMS)
A service that manages cryptographic keys used for data encryption and digital signatures.
Kubeconfig
A configuration file containing details like cluster API server addresses, contexts, and user credentials used by kubectl and other Kubernetes tools to authenticate and manage access to Kubernetes clusters. Kubeconfig files let you switch between different clusters and manage multiple environments securely.
Leafgroup
A leafgroup is a group of leaf switches in a backend network topology that serves a common set of Nodes. CoreWeave labels each Node with its leafgroup, so you can keep tightly coupled workloads within one leafgroup to reduce network hops. In the topology hierarchy, a leafgroup belongs to a superpod.See also: Fabric (backend network), Superpod, InfiniBand and RoCE labels
Least privilege
A security principle where users or systems are given only the minimum access required to perform their tasks.
CoreWeave’s Local Object Transport Accelerator (LOTA) is a container that lives on every GPU and CPU Node inside a client’s cluster to accelerate object transfers. Conventional transfer accelerators speed up the data transfer rates of bucket contents over long distances.LOTA uses NVMe SSDs in each Node to create a cluster-wide cache, increasing throughput and reducing latency.
MACsec
MACsec (Media Access Control Security) is an IEEE standard for securing Ethernet networks at the link layer by encrypting and authenticating Ethernet frames.
On CoreWeave, Dedicated Direct Connect supports MACsec between CoreWeave edge routers and your network equipment.See also: IEEE 802.1AE, About Direct Connect
Managed component
A managed component is software that CoreWeave installs into your cluster and keeps reconciled to its intended state. You shouldn’t modify a managed component directly, because CoreWeave automation reverts changes to it. If a managed component misbehaves, that’s a CoreWeave responsibility and a valid reason to open a support ticket. This is distinct from a Helm-installed component, which you install and own.See also: Helm-installed component, What CoreWeave manages
Medium Access Control (MAC) address
A MAC address is a unique identifier assigned to a Network Interface Controller () for use as a network address within a network segment.
Microsegmentation
Dividing networks into smaller zones to enforce granular security controls.
Multi-Factor Authentication (MFA)
A security mechanism requiring multiple forms of verification to access systems, for example, a password and mobile code.
An extension of TLS where both client and server authenticate each other using certificates.
Network Interface Controller (NIC)
A network interface controller (NIC) is a hardware component that connects a computer to a network.
NICs are commonly used to connect computers to Ethernet networks, wireless networks, and other types of networks.
Nimbus
Nimbus is CoreWeave’s control and data-plane software that runs on the NVIDIA BlueField Data Processing Unit (DPU) attached to each Node.
Nimbus provides tenant isolation, VPC membership, public internet routing, and Node lifecycle coordination for CoreWeave Kubernetes Service (CKS) without relying on virtual machines.See also: Data Processing Unit (DPU), Virtual Private Cloud (VPC), CoreWeave Kubernetes Service (CKS)
Node
In CoreWeave Kubernetes Service (CKS) and most Infrastructure docs, a Node is a Kubernetes worker machine that runs Pods.
In SUNK and Bare Metal docs, a lowercase node is often a Slurm or physical machine rather than a Kubernetes Node.
When both appear on the same page, SUNK docs capitalize Kubernetes Nodes to keep the distinction clear.See also: Node Pool, Node label, SUNK
Node condition
A Node condition is a status signal that describes some aspect of a Node’s health or readiness. Kubernetes reports standard conditions, and CoreWeave surfaces a subset of customer-visible conditions that explain why a Node is healthy, cordoned, or being serviced. When a Node is cordoned, its condition and reason show you whether the cause is a hardware fault CoreWeave is handling or an action you can take. The Node conditions that CoreWeave sets drive its own Node lifecycle automation: don’t rely on them for your own automation or management, and change them only through documented procedures, such as cordoning a Node yourself or requesting a Node reboot. You can add your own custom Node conditions.See also: Cordon, Node lifecycle automation, Node cordoning
Node Health Check (NHC)
SUNK runs Node Health Check (NHC) as one of the scripts in the Epilog chain after a job, and only when no other job is still running on the node. A failed check drains the node with a reason that begins with NHC:.The text after NHC: names the failed check. Most NHC drains carry sunk:verify-undrain, so they clear on their own once the node passes the next HPC Verification test. NHC is separate from HPC Verification: NHC is a tool that SUNK runs on a node in the Epilog after a job, while HPC Verification is CoreWeave’s hourly test of idle Nodes across the fleet.See also: HPC Verification, Node Health Check (NHC) drains
Node label
A Node label is a key-value pair attached to a Node that describes its hardware, location, or network topology. CoreWeave applies labels automatically, including topology labels for the backend network fabric, superpod, leafgroup, and rack, and labels that identify the instance type and capabilities. You use Node labels in node selectors and affinity rules to control where Pods are scheduled. CoreWeave-applied labels are the supported vocabulary for placement.See also: Fabric (backend network), Superpod, Leafgroup, InfiniBand and RoCE labels
Node lifecycle automation
Node lifecycle automation is the CoreWeave system that manages each Node from Day 0 initialization through Day 2+ production operations. In a CKS cluster, when a Node condition fires, it cordons the Node and, depending on the condition, keeps it cordoned until the condition clears, reboots it, or moves it to triage for investigation. It re-validates a failed Node before the Node returns to production, and replaces Nodes that leave your cluster to keep it at full capacity. Node lifecycle automation is distinct from SUNK’s Node Controller, which synchronizes data between Kubernetes Nodes and slurmd Pods.See also: Cordon, Node condition, Node lifecycle, Node state transitions in CKS, Node Controller
Node Pool
A Node Pool is a logical grouping of Nodes in a CKS cluster with the same Instance Type, Labels, Taints, and Annotations.
Node Pools are useful for managing a group of Nodes as a single entity or assigning workloads to specific Nodes based on their configuration.
Non-Volatile Memory Express (NVMe)
NVMe is a storage protocol that provides high-performance access to non-volatile memory devices. NVMe is designed to take advantage of the low latency and high throughput of modern storage devices, providing improved performance over traditional storage protocols.
OpenID Connect (OIDC)
OpenID Connect (OIDC) is an identity layer laid atop the OAuth 2.0 protocol, which lets users authenticate by verifying their identity through an identity provider (IdP), such as Okta. CoreWeave supports OIDC as an authentication method to CKS clusters.
PCI Express (PCIe)
Peripheral Component Interconnect Express, officially abbreviated as PCIe, is a high-speed serial computer expansion bus standard.
Penetration testing
Simulated cyberattacks used to evaluate the security of cloud environments.
The practice of defining and enforcing security and compliance rules using machine-readable code, for instance, using Open Policy Agent.
Preboot eXecution Environment (PXE)
The PXE specification describes a standardized client-server environment that boots a software assembly, retrieved from a network, on PXE-enabled clients.
On the client side it requires only a PXE-capable network interface controller (), and uses a small set of industry-standard network protocols such as and .
PXE is most often pronounced as “pixie”, and the process is often called “pixie boot”.See also: iPXE
A Region is an area within a Geo that contains multiple Availability Zones (AZs).
Regions provide the infrastructure foundation for resiliency, low-latency connectivity, and data residency requirements.
For example, Region US-EAST-05 resides in the US-EAST Super Region and the US Geo.See also: Geo, Super Region, Availability Zone (AZ), About Regions and Availability Zones
Remote Direct Memory Access (RDMA)
RDMA allows data to be transferred directly between the memory of two computers without involving the operating system or CPU.
RDMA provides low latency and high throughput data transfers, making it ideal for high-performance computing environments.See also: GPUDirect RDMA, RDMA over Converged Ethernet (RoCE)
Role-Based Access Control (RBAC)
An authorization method where users are granted permissions based on their role within an organization.
Service mesh
A network layer (such as Istio or Linkerd) that handles secure service-to-service communication with observability and policy enforcement.
Shared responsibility model
A framework outlining security responsibilities split between the cloud provider and the customer.
Slurm
Slurm is a job scheduler commonly used on high-performance computing (HPC) clusters.
On CoreWeave, SUNK runs Slurm on Kubernetes so Slurm jobs and Kubernetes workloads can share the same compute pool.See also: SUNK
SUNK is CoreWeave’s system for running Slurm on Kubernetes.
Slurm nodes run as Kubernetes Pods, so Slurm jobs and Kubernetes workloads can share the same compute pool.
SUNK runs in SUNK Standard or SUNK Self-Service operation modes. SUNK Standard also runs outside CoreWeave as SUNK Anywhere.See also: Slurm, Node Health Check (NHC), CoreWeave Kubernetes Service (CKS)
Super Region
A Super Region is a large section of a Geo that typically contains multiple Regions.
Super Regions group Regions for administrative purposes and provide a high-level view of infrastructure in a geographic area.
For example, all locations in the eastern United States are part of the US-EAST Super Region.See also: Geo, Region, Availability Zone (AZ), About Regions and Availability Zones
Superpod
A superpod is a high-level grouping within a backend fabric that contains multiple leafgroups. CoreWeave labels each Node with its superpod, so you can keep a large distributed job within a single superpod to limit the network distance between its Nodes.See also: Fabric (backend network), Leafgroup, InfiniBand and RoCE labels
System for Cross-domain Identity Management (SCIM)
System for Cross-domain Identity Management (SCIM) is an open standard protocol that automates user and group provisioning across systems.
On CoreWeave, Automated User Provisioning (AUP) uses SCIM to sync users and groups from an identity provider into CoreWeave IAM, and SUNK User Provisioning (SUP) can use SCIM to provision POSIX users in SUNK clusters.See also: Lightweight Directory Access Protocol (LDAP), Identity provider (IdP), Automated User Provisioning, SUNK User Provisioning
Top of Rack (TOR)
A network switch that connects servers in an Availability Zone to the rest of the network.
TOR switches are typically located at the top of a rack of servers and provide network connectivity to the servers within the rack.
Trivial File Transfer Protocol (TFTP)
TFTP is a lightweight file transfer protocol that doesn’t provide authentication or encryption.
TFTP is commonly used for network booting and firmware updates.
Virtual Extensible LAN (VXLAN)
VXLAN addresses the limitations of traditional Virtual Local Area Networks () in large-scale environments.
VXLAN encapsulates Ethernet frames within User Datagram Protocol (UDP) packets, enabling them to traverse across IP networks.
By extending Layer 2 networks over Layer 3 infrastructure, VXLAN allows for greater flexibility and scalability in large-scale, multi-tenant environments.
Virtual Local Area Network (VLAN)
A Virtual Local Area Network (VLAN) is a broadcast domain that’s partitioned and isolated in a computer network at the data link layer (OSI layer 2).
Virtual Private Cloud (VPC)
A VPC is a private network hosted within CoreWeave’s public cloud infrastructure.
A VPC provides a secure environment where resources can be isolated from the public internet and from other VPCs.
On CoreWeave, VPCs define network isolation for attached products such as CKS clusters, including IP address ranges and related networking configuration.See also: About VPCs
Virtual Routing and Forwarding (VRF)
VRF allows multiple instances of a routing table to coexist within the same router at the same time.
Each VRF instance maintains its own routing table, which is separate from the global routing table.
Workload identity
Associating workloads (like containers and VMs) with verifiable identities for secure communication and policy enforcement.