Anvil / Private AI infrastructure

Feature complete

Make the cluster a foundation. Not a project of its own.

Anvil is designed to automate private Kubernetes infrastructure for AI workloads. The focus is repeatable provisioning and a lifecycle that an existing IT team can understand and operate. Deploy Kubernetes in hours without touching a single YAML file.

Inside Anvil

See the infrastructure take shape.

From supported hardware to AI workloads

Start with a defined deployment path.

AI infrastructure combines familiar responsibilities—compute, networks, storage, identity—with model-serving and GPU requirements. Anvil’s role is to bring those responsibilities into a supported operational environment.

Ataraxia will provide fully validated hardware recommendations scoped to your specific workloads. No need to over-deploy, start with what you need today, and add nodes to the cluster as your requirements grow. New GPU worker nodes can be added in under 10 minutes.

Anvil Cloud is designed to operate connected to your cluster, however if you need a fully air-gapped infrastructure, Anvil will automatically relocate itself to your infrastructure and run as Anvil Local. Anvil Local receives upgrades, model downloads, and service packages in the form of encrypted, signed, and sealed bundles from Anvil Cloud. Bundles are created using the Anvil Ladle app (Linux, macOS, Windows) and hand-carried to the Cluster. Corollary updates are hand-carried back to the Cloud to keep Anvil Cloud + Local in sync.

Discuss your deployment requirements →
ProvisioningRecommended deployment · 3 x control-plane nodes + 2 or more AI worker nodes with GPUs
AnvilProvisioning & lifecycle control
HardwareNVIDIAAMDIntel
Platform servicesForgeIn clusterIngotOnsite repositoryGrafanaObserveCiliumNetworkingCephStorage
Control plane3 x nodesSupported configuration
AI workerGPUSupported configuration
AI workerGPUSupported configuration
AI workerGPUAdd as needed
NetworkingHealth signals

Actual sizing depends on the workflows, concurrency, model sizes and facilities.

Product focus

Plan for day two from day one.

Provision

Repeatable installation and configuration of Kubernetes without touching a YAML file.

  • GPU hardware discovery and auto configuration, GPU bonding, vGPU Slicing.
  • Rook/Ceph for storage redundancy.
  • Network support up to 400Gbps.
  • NVIDIA DGX Spark supported as an accelerated GPU worker node.
Features validated

Operate

Cluster health and essential platform services with clear operator visibility.

Observer telemetry monitoring agent, monitors the cluster, corrects minor events, and notifies operators with recommendations for major events.

Features validated

Maintain

Hitless per‑node upgrades, with cluster‑wide upgrade options.

Backups, restores, and rollbacks can be run per‑node or per‑cluster.

Connected clusters can update directly from the cloud, disconnected clusters can be regularly upgraded using hand-carried, encrypted, and signed bundles to maintain isolation.

Features validated

Anvil v1.0.0

Where we started and where we are going.

Initial release priorities

  • Opinionated Kubernetes, configured for Agentic AI
  • No YAML required. Hide the complexity, not the controls.
    • Full access to Kubernetes if it’s needed
  • Air-gapped deployment, and Lifecycle Mgmt.
  • NVIDIA and AMD GPU support (Bonding / Slicing)
  • Hitless upgrades and rollbacks
  • Redundant storage
  • Cloud → Local / Local → Cloud Cluster Orchestration

Roadmap Features

  • Multi-site fleet administration
  • Intel, Cerebras, Tenstorrent, d-Matrix, SambaNova support
  • RoCEv2 (RDMA over Converged Ethernet)
  • VMware and Proxmox deployment auto provisioning
  • AWS and Azure deployment blueprints
  • NVIDIA DGX Spark RDMA Cluster
  • MTU Autodiscovery
  • BMC and Redfish integration

Build with us

Start with one useful workflow.

Bring your infrastructure, your operational constraints, and a problem you need solved. Together, we can define a focused pilot and what success should look like.

Discuss a design partnership