Documentation – OpenMetal Cloud

A placement assessment decides which workloads should stay in public cloud and which would cost less on private infrastructure. This lists the data to gather beforehand, why each piece matters, and what a useful assessment hands back to you.

Workloads built on standard interfaces, like VMs from standard images, containers, open-source databases, and S3-compatible storage, move between providers with little rework. This sorts common workloads by how easily they move and what tends to slow them down.

On an OpenMetal hosted private cloud, we run the hardware, the data center network, and the platform underneath your cloud, while you control everything inside it. This lays out the split layer by layer and how Assisted Management shifts it.

Moving off public cloud works best one workload at a time, and some workloads are better left where they are. This covers the patterns that favor staying, the patterns that favor moving, and how to evaluate a workload that could go either way.

OpenMetal bare metal servers and Hosted Private Cloud VMs can share the same customer-dedicated VLAN, with unmetered private traffic between them. This covers how it works, a step-by-step setup in Central, addressing and security, and common patterns.

DeepSeek-R1’s 671B MoE fits one eight-GPU HGX B300 node at FP8, with room for KV cache. See the fit math and why the node beats a multi-server build.

OpenMetal private clouds serve the Swift object storage API through Ceph’s RADOS Gateway. Most Swift applications work unchanged, but a handful of behaviors differ. This covers each one, why it matters, and how to test before you move.

InfiniBand vs Spectrum-X Ethernet for AI clusters: how NVIDIA’s Quantum-X800 InfiniBand and Spectrum-X Ethernet differ, what ‘InfiniBand-class’ means, and when each fits.

DeepSeek-R1 is a 671B MoE: its FP8 weights span about six discrete H200 cards. See the fit math, why MoE memory follows total params, and how to serve it.

Fixed monthly pricing changes how unit costs work. This walks through calculating usable capacity, cost per VM at your real fill rate, and cost per customer, plus how to handle idle capacity and the step up when you add hardware.

In a hybrid setup, every byte between AWS and your private cloud is billed on one side or the other, depending on direction. This explains where those charges come from, a simple way to estimate them, and design patterns that keep them small.

Adding a node to a Private Cloud Core v4 uses the same automation that built the cloud. This walks through each step from order to usable capacity, what changes for Ceph and security, and what to plan before you grow.

This release delivers Private Cloud Core v4.0.0 (Horizon UI by default), versioned Datadog monitors with the expanded v1.1 set, and a new firewall configuration management API. It also fixes issues with false node-removal alerts, lost Datadog coverage, and IP-whitelisted API keys being rejected.

Every Private Cloud Core v4 node ships with a host-level firewall that denies inbound traffic on its external interface by default. This explains what’s open, what’s trusted, how to grant access in Central or through the API, and how it differs from security groups.

The answer is set by three things: the model’s weight footprint, the KV cache your context and concurrency produce, and how your serving framework spreads both across GPUs. An eight-GPU

OpenMetal has acquired Economize, the agentless cloud cost management platform. Todd and Economize co-founder Anirudh Murali cover why the companies fit, what real cloud spend data shows, and the launch of OpenMetal Labs for complex infrastructure, billing, and migration problems.

The NVIDIA HGX B300 is NVIDIA’s Blackwell Ultra platform in its densest single-node form: eight B300 SXM GPUs wired together on one baseboard by fifth-generation NVLink and NVSwitch into a

New Private Cloud Core v4 clouds use OVN for OpenStack networking by default. The choice between OVN and OVS is made once, before deployment, and decides which load balancer features you get. This covers the tradeoffs and how to choose.

We validated Private Cloud Core v4 with OpenStack’s Tempest suite and published the run guide, configuration, and exclusion list. This walks through running it on your own cloud, reading the results, and understanding why each exclusion exists.

The OpenMetal Central public API covers the full bare metal lifecycle: inventory checks, orders with deployment configurations and saved SSH keys, provisioning status, scaling, power control, and decommissioning, with a sandbox for testing before you touch production.

OpenMetal Central lets you create, assign, and tear down VLANs and IP blocks through the public API. This walks through the object model, a scripted private network between an application cluster and a database cluster, and the rules that prevent accidental deletions.

OpenMetal bare metal can now boot an OS entirely into memory from a kernel and initramfs you host, then configure itself through cloud-init and a node-local metadata service. This covers how it works, image requirements, and where it fits.

As Odoo deployments grow in users and modules, shared cloud hosting and managed platforms hit performance and cost ceilings. This article covers why dedicated bare metal changes the picture, how to size hardware against Odoo’s actual resource model, and when to split the application and database onto separate servers.

OpenMetal bare metal servers support custom OS installs via IPMI, including Windows Server. This article covers the installation path, edition selection and core licensing on dedicated hardware, Hyper-V setup for VM consolidation, and which hardware configurations suit different Windows workloads.

When compliance teams ask for air-gapped infrastructure, they usually mean network isolation: workloads that can’t reach the internet and can’t be reached from it. This article explains the difference between true air-gap and network isolation, how OpenMetal’s private networking architecture enables isolated deployments, and where the honest boundaries are.

How Temporal Cloud’s per-action billing model works and where it compounds, what the official Temporal documentation says self-hosted deployments require, the infrastructure stack a production cluster needs, and where OpenMetal’s bare metal fits into that.

A practical guide to matching NVIDIA’s H100, H200, B200, and B300 to your model, your latency target, and your budget, starting from the workload instead of the spec sheet. Key

What LangGraph manages and what it explicitly does not, why MemorySaver is not a production checkpointer, what PostgresSaver requires from your infrastructure, the compute profile of long-running agent workflows, and where dedicated bare metal fits.

What Article 12 of the EU AI Act actually requires, why the obligation is an infrastructure problem before it is a software problem, what compliant log storage needs to look like in practice, the current enforcement timeline, and how private cloud in the EU addresses the requirement.

OpenMetal Private Cloud Core v4 is a major platform release that updates the core infrastructure behind new OpenMetal private cloud deployments. The release moves the platform to Ubuntu 24.04, OpenStack

Runway Intelligence is OpenMetal’s executive insight series for late-stage startups and their investors, exploring how cloud economics, infrastructure design, and operational strategy shape valuation, margins, and time to exit.  At

AI-first SaaS companies are reporting 52% gross margins, more than 20 points below the 75-85% range that built the SaaS valuation premium. The cause isn’t just market pressure. It’s an infrastructure decision being made by default.

The four distinct Intel Xeon generations running across OpenMetal’s bare metal and hosted private cloud lineup, what actually changes between them beyond core count, and which generation fits which kind of workload, including where confidential computing eligibility depends on generation and hardware configuration together.

We look at why shared, allocation-based HPC access becomes a real operational problem for life sciences computational research, what changes on dedicated infrastructure that you administer yourself, and how to structure a transition that scales up gradually rather than requiring a large upfront commitment.

We break down how AWS’s Enterprise Discount Program, Azure’s Enterprise Agreements, and GCP’s Workload Agreements actually work, why the discount comes attached to a real financial commitment rather than just usage-based savings, and what that commitment structure costs an enterprise that outgrows or under-uses it.

We walk through a real worked example for sizing a private cloud cluster before committing to a contract, covering how to translate your VM count into node count, why redundancy and replication overhead eat into your raw numbers, and where network bandwidth becomes the limiting factor as a cluster grows.

We break down what Terraform and Ansible each actually do differently, why the overlap between them confuses newcomers, and a practical framework for deciding which one to reach for, or whether you need both, when automating a private cloud deployment.

We look at the real case for using S3-compatible object storage as a default backend, where that approach genuinely breaks down on latency, how serious platforms solve it with a caching layer rather than abandoning object storage entirely, and how to make the same call on a Ceph-based cluster.

We look at the real timing constraint behind real-time bidding auctions, why shared cloud infrastructure quietly eats into that budget through virtualization and cross-zone network hops, and where dedicated bare metal removes that variance directly.

NVIDIA AI Enterprise (NVAIE) is a support and integration product, and knowing that up front settles most of the buying decision. It does not unlock faster GPUs or a private

An account-intelligence system that pre-embeds ten million companies into a resident vector index, and one that dispatches agents to research those same companies live on demand, look like the same

We look at why a company’s production data residency doesn’t automatically cover its CI/CD pipeline, what GitHub Actions actually offers EU-based teams today, and how self-hosting runners on Amsterdam hardware closes that gap directly.

We look at why attack surface management and cyber-risk scanning companies, businesses that scan large swaths of the public internet as their core product, run into trouble on hyperscaler infrastructure, what actually needs to be true about a provider’s acceptable use policy and IP allocation to support this workload, and how that differs from running an internal penetration testing lab.

On latency-bound inference, the Model FLOPs Utilization your optimization stack can actually hold is capped by who else shares the box, not by the kernel that runs on it. Sustained

Runway Intelligence is OpenMetal’s executive insight series for late-stage startups and their investors, exploring how cloud economics, infrastructure design, and operational strategy shape valuation, margins, and time to exit.  A

We walk through the real total cost of building your own dense storage server for off-site backup and archival data versus renting equivalent capacity, covering drive costs in today’s market, the parts of total cost of ownership that don’t show up on a parts list, and how Ceph’s approach to redundancy compares to a single chassis.

We work through the actual cost crossover between GitHub-hosted Actions runners and self-hosted runners on dedicated bare metal, using GitHub’s current 2026 rates, correct a common misconception about a self-hosted runner fee that never took effect, and cover what a self-hosted build pipeline needs beyond just cheaper compute.

We compare three dedicated server providers commonly considered for Proxmox VE hosting, OpenMetal, Hetzner, and OVHcloud, across real hardware specs, current pricing, storage architecture, and support model, so you can match the provider to your actual workload rather than just the sticker price.

We break down what the EU Cyber Resilience Act’s vulnerability reporting obligations actually require starting September 2026, why the tight reporting clock is fundamentally an infrastructure visibility problem, and where dedicated infrastructure and controlled build pipelines make that clock achievable.

Intel TDX confidential VMs run on OpenMetal dedicated bare metal hardware today, available on on XL v5, with OpenStack Nova orchestration slated for after Hibiscus.