For full infrastructure control, use bare-metal servers from a hyperscaler such as AWS .metal instances, Oracle Cloud Infrastructure (OCI) bare metal, or IBM Cloud bare metal when you also need managed services nearby, or choose a dedicated bare-metal provider like Hetzner, OVHcloud, phoenixNAP, or Vultr when you want single-tenant hardware at lower cost. Which option is best depends on which layer you need to control, from the operating system and kernel down to firmware, out-of-band management, and the physical network.

Most comparisons of cloud providers mix root access, single-tenant hardware, IPMI access, and network control into a single bucket labeled “control.” That conflation makes it hard to match a provider to a specific requirement. This guide breaks control into discrete layers, maps each major provider against those layers with verified documentation, and walks through the trade-offs you accept when you take on full operational responsibility.

Disclosure: Control Plane publishes this guide.

TL;DR

Full infrastructure control means different things at different layers.

  • Control has layers. Root SSH access, IPMI access, BGP peering, and firmware-level management are separate capabilities. Define which layers matter before choosing a provider.
  • Hyperscaler bare metal gives you single-tenant hardware inside an ecosystem of managed databases, identity services, and global networking. AWS EC2 .metal instances, OCI bare metal shapes, and IBM Cloud bare metal servers are the strongest options.
  • Dedicated bare-metal providers like Hetzner, OVHcloud, phoenixNAP, and Vultr offer single-tenant servers with transparent pricing and, in some cases, out-of-band management access that most hyperscalers do not expose.
  • Hosted private cloud from providers like OpenMetal gives you an entire OpenStack environment on dedicated hardware, adding API-driven orchestration on top of bare metal.
  • Full control means full operations. You own patching, failover, monitoring, capacity planning, security hardening, and hardware replacement. Plan the team and the runbooks before you commit.
  • Control Plane’s Managed Kubernetes (MK8s) runs Kubernetes and your workloads on servers you control, and Control Plane handles Day 2 operations for you, including autoscaling, built-in observability, and security.

Which cloud providers are best for full infrastructure control?

This table maps what each provider lets you control, based on each provider’s current documentation; “Not documented” means the public docs do not describe the capability.

Provider Bare metal (single-tenant, no provider hypervisor) Custom kernel or your own hypervisor Out-of-band management (IPMI or equivalent) Network control (private networking, BGP, bring your own IP) Regions or locations Managed services alongside
AWS (EC2 .metal instances) Yes. Instances run directly on the host hardware with no AWS hypervisor. Yes. You can run any kernel, KVM, VMware ESXi, or Xen. No. AWS does not expose IPMI or BMC to customers. VPC, placement groups, Elastic IPs, and bring-your-own-IP (BYOIP) via BYOIP for EC2. BGP available through AWS Direct Connect. Dozens of regions across North America, Europe, Asia-Pacific, Middle East, Africa, and South America. Extensive. RDS, S3, IAM, CloudWatch, Lambda, and hundreds more.
Microsoft Azure (BareMetal Infrastructure, Dedicated Hosts) Partial. BareMetal Infrastructure is available for SAP HANA and specialized workloads. Dedicated Hosts provide single-tenant VMs but still run on the Azure hypervisor. BareMetal Infrastructure runs only specific workloads (for example, Nutanix NC2 runs Nutanix AHV); you cannot install a hypervisor of your choice. On Dedicated Hosts, you choose the guest OS and kernel but cannot replace the Azure hypervisor. Not documented for general-purpose use. Azure VNet, ExpressRoute with BGP, and BYOIP support. Regions worldwide. Broad. Azure SQL, Cosmos DB, Entra ID, AKS, and more.
Google Cloud (bare metal instances, Bare Metal Solution) Yes. Bare metal instances (C3, C4, C4A, C4D, X4, Z3, and others) run on a dedicated host server. Bare Metal Solution provides dedicated servers in Google Cloud extensions for Oracle and other workloads. Full OS control; check Google’s docs for hypervisor support. Not documented for customer-facing access. VPC, Cloud Interconnect with BGP, and BYOIP. Regions across North America, Europe, Asia-Pacific, and South America. Extensive. Cloud SQL, BigQuery, GKE, IAM, and more.
Oracle Cloud Infrastructure (bare metal shapes) Yes. OCI bare metal shapes run directly on physical servers with no hypervisor layer. Yes. Full control over OS and hypervisor. Not documented for direct customer IPMI access. VCN, FastConnect with BGP, and BYOIP. OCI also offers sovereign cloud regions. Commercial and sovereign regions across North America, Europe, Asia-Pacific, Middle East, and Latin America. Oracle Autonomous Database, OCI Container Engine, Object Storage, and IAM.
IBM Cloud (bare metal servers) Yes. Single-tenant dedicated servers provisioned hourly or monthly. Yes. Install any OS, custom kernel, or hypervisor including VMware and KVM. Yes on Classic infrastructure (IPMI over the IBM Cloud VPN). Private VLANs, BYOIP not documented, and Direct Link with BGP. Data centers across North America, Europe, Asia-Pacific, and South America. IBM Cloud Databases, Kubernetes Service, Watson AI, and Object Storage.
Hetzner (dedicated root servers, Hetzner Cloud) Yes. Dedicated root servers are single-tenant physical machines. Yes. Full root access allows custom kernels, KVM, Proxmox, or any hypervisor. Yes, on request. KVM console via Robot, free for the first 3 hours. Private Layer 2 networking via vSwitch (VLANs across locations). BYOIP and BGP not documented. Data centers in Falkenstein, Nuremberg, and Helsinki. Hetzner Cloud adds Ashburn, Hillsboro, and Singapore. Limited. Hetzner Cloud offers load balancers, firewalls, and object storage, but no managed databases or IAM.
OVHcloud (dedicated servers, Bare Metal Cloud) Yes. Fully dedicated physical servers. Yes. Install any OS, kernel, or hypervisor. Yes. IPMI access is available on dedicated servers. Private networking (vRack), BYOIP, and BGP support through OVHcloud Connect. Data centers in France, Germany, UK, Canada, US, Singapore, Australia, and others. Managed Kubernetes, managed databases, Object Storage, and Public Cloud services.
phoenixNAP (Bare Metal Cloud) Yes. Single-tenant servers provisioned via API. Yes. Full root access for custom OS and hypervisor installation. Not documented as direct IPMI access; management is through the phoenixNAP portal and API. Private networking, VLAN support, and BYOIP. Self-service BGP peering via the Network API. Data centers in Phoenix, Ashburn, Chicago, Seattle, Amsterdam, and Singapore. Limited managed services. BMC integrates with Rancher and Ansible.
Vultr (Bare Metal) Yes. Dedicated physical servers with no virtualization layer. Yes. Custom ISOs and any OS or hypervisor supported. Not documented for direct IPMI access; management via Vultr dashboard and API. Private networking, BGP support, and BYOIP. Locations across North America, Europe, Asia-Pacific, and South America. Managed databases, Kubernetes (VKE), Object Storage, and load balancers.
OpenMetal (hosted private cloud on OpenStack) Yes. Dedicated hardware running a private OpenStack cloud. Yes. Full control over the OpenStack environment, including hypervisor configuration. Not documented as direct IPMI to individual servers; management is through the OpenMetal and OpenStack interfaces. Private networking within OpenStack, BYOIP, and custom routing. Ashburn, Los Angeles, Amsterdam, and Singapore. OpenStack-native services including block storage, object storage, and identity (Keystone). No hyperscaler-style managed databases.

Hyperscalers give you bare metal hardware but rarely expose out-of-band management like IPMI. IBM Cloud is the exception among hyperscalers, offering IPMI access on its Classic bare metal servers. Among dedicated providers, Hetzner and OVHcloud both document remote console or IPMI access, which matters for firmware updates, crash recovery, and pre-boot diagnostics.

What does full infrastructure control actually mean?

Infrastructure control sits on a spectrum from fully managed platforms where the provider handles nearly everything to running your own data center where you handle every layer.

Level What you control What the provider controls Typical use
Managed platform (PaaS) Application code and configuration Runtime, OS, networking, hardware, scaling, patching Web apps, APIs, prototyping
Virtual machines OS, middleware, application stack Hypervisor, hardware, physical network, data center General-purpose workloads, dev/test, internal tools
Dedicated hosts VM placement, OS, socket and core visibility for licensing Hypervisor, firmware, hardware, physical network License-bound software (Windows Server, SQL Server), compliance workloads
Bare metal servers OS, kernel, hypervisor, local storage, software-defined networking Physical hardware maintenance, firmware and BMC on hyperscalers, data center power and cooling, network backbone HPC, databases, real-time systems, custom virtualization
Hosted private cloud Cloud API, tenants, networking, storage; hypervisor upgrades vary (customer on OpenMetal, provider on most managed VMware offerings) Physical hardware, data center facilities Enterprises migrating from on-premises VMware or OpenStack
Colocation Everything above the rack: servers, networking gear, OS, applications Power, cooling, physical security, internet uplinks Organizations with existing hardware and ops teams
Your own data center Everything Nothing Maximum control, maximum operational burden

Each step down the table adds control and adds operational responsibility. The deciding question is how much control your team can operate.

Which provider fits which kind of control?

Each common need maps to a starting point and a practical consideration.

Need Start with Watch out for
Custom kernel modules or eBPF-heavy workloads Any bare metal provider (Hetzner, Vultr, AWS .metal, OCI bare metal) With Secure Boot or kernel lockdown enabled, unsigned modules will not load, so sign them or disable lockdown. On VMs, eBPF runs in the guest kernel, but native XDP and hardware performance counters depend on the virtual NIC driver and what the hypervisor exposes.
Your own hypervisor (VMware, KVM, Proxmox) AWS .metal instances, IBM Cloud bare metal, Hetzner dedicated, OVHcloud dedicated VMware licensing costs are separate. Confirm the provider supports your hypervisor’s network and storage requirements.
Licensing tied to physical cores Dedicated hosts (AWS, Azure) or bare metal servers (IBM Cloud, OCI, Hetzner) Microsoft licenses bought after October 1, 2019 cannot be brought to dedicated hardware at Listed Providers (AWS, Google, Alibaba) without License Mobility, and Windows Server has none; that applies to AWS .metal too. Oracle counts vCPUs on authorized clouds and physical cores elsewhere. Have licensing counsel confirm before you commit.
Predictable performance without noisy neighbors Any bare metal option Bare metal eliminates noisy neighbors but does not eliminate network congestion on shared uplinks or contention on network-attached block storage. Ask about uplink capacity and storage isolation.
Cheapest single-tenant compute Hetzner dedicated root servers, OVHcloud dedicated servers Lower cost often means fewer managed services and fewer regions. Factor in the cost of your own ops team.
Bring your own IP space or BGP Vultr, phoenixNAP, OCI, AWS BYOIP BGP setup complexity varies. Some providers require a support ticket; others offer self-service BGP configuration. BYOIP generally needs at least a /24 (IPv4) or /48 (IPv6) registered with a regional internet registry, plus a valid route origin authorization (ROA).
Strict data residency OCI sovereign cloud regions, OVHcloud (EU-headquartered), Hetzner (EU data centers) Verify that the provider’s terms of service and data processing agreements meet your specific regulatory framework.
Control plus managed databases nearby AWS .metal with RDS, OCI bare metal with Autonomous Database, Vultr bare metal with Managed Databases Managed database traffic between bare metal and the database service may traverse the provider’s internal network. Confirm latency and security posture.

The right provider depends on which layers you need to own and which services you need the provider to run.

What do you give up with full control?

Choosing bare metal or any high-control deployment model means accepting operational responsibilities that managed services handle for you.

  • Patching and upgrades. You are responsible for OS security patches, kernel updates, and firmware updates. On hyperscaler VMs, the provider patches the hypervisor and host OS. On bare metal, that layer is yours.
  • Failover and high availability. Most dedicated providers do not restart a failed server on another host. You need to design redundancy, configure health checks, and build failover automation.
  • Monitoring and alerting. Dedicated bare metal servers usually do not come with built-in cloud-native monitoring. You deploy your own observability stack or integrate with a third-party service.
  • Capacity planning. Scaling bare metal is not instant. Provisioning a new server can take minutes (API-driven providers like Vultr and phoenixNAP) to hours for stock configurations, and days or weeks for custom builds or sold-out locations. You plan capacity ahead of demand.
  • Security hardening. Firewall rules, intrusion detection, access controls, and audit logging are your responsibility from the OS up.
  • Hardware replacement. When a disk fails or a NIC dies, on dedicated providers you file a ticket, often with the failed disk’s serial number, and wait for the swap. Local data survives only if you run RAID and rebuild it yourself. On hyperscalers the instance is retired and you relaunch it on a new host. Your architecture must tolerate hardware failure gracefully.
  • Out-of-band and network hardening. Keep the BMC and IPMI interface off the public internet, since exposed BMCs are a common attack path. Update BMC and BIOS firmware, and confirm the provider wipes firmware and disks between tenants. Many low-cost servers have a single NIC on a single switch, so ask for bonded uplinks. DDoS protection varies widely by provider, from always-on scrubbing to basic or none, so confirm it before going to production.

The operational cost of bare metal includes the server price and the engineering time to keep everything running, patched, and secure. For teams evaluating what happens when you move off hyperscaler managed services, start by mapping the service equivalents and SLAs you give up.

How Control Plane fits

Control Plane is the unbreakable platform for running production workloads.

It runs serverless, standard, stateful, and cron workloads natively on AWS, GCP, Azure, and Oracle Cloud Infrastructure (OCI), VM workloads on locations joined through Control Plane’s Managed Kubernetes, and workloads on Kubernetes clusters you already run or your own machines, including on premises, active-active across regions and clouds.

Documented infrastructure providers for Control Plane’s Managed Kubernetes (MK8s) are Generic, AWS, GCP, Hetzner, and Triton. To run workloads on bare metal servers or on premises VMs, you use the Generic provider, which connects your own hardware to the Control Plane platform. VM workloads run on locations joined through MK8s on KVM-capable nodes, such as bare metal servers.

On AWS and GCP, MK8s runs worker nodes in your own cloud account or project. You keep infrastructure-level control of those accounts while Control Plane handles Kubernetes lifecycle management, including upgrades and autoscaling.

With MK8s on your own hardware, you still own the servers, the operating system choice, and the physical network; Control Plane runs Kubernetes and your workloads on top. For teams running hybrid Kubernetes across on-premises and public cloud, a Control Plane Global Virtual Cloud (GVC), a named set of locations, runs each workload across cloud regions and on-premises MK8s locations from one workload definition.

Organizations concerned about avoiding cloud vendor lock-in can define workloads in a GVC and run them across any location in that GVC, whether the underlying infrastructure is a hyperscaler region, a Hetzner dedicated server, or a rack in your own data center.

FAQ

Common questions about bare metal cloud providers and infrastructure control.

Which cloud providers are best for full infrastructure control?

For full infrastructure control, the best providers are those offering bare metal servers with single-tenant hardware, root access, and configurable networking. Among hyperscalers, AWS EC2 .metal instances, OCI bare metal shapes, and IBM Cloud bare metal servers provide the deepest control while keeping managed services accessible. Among dedicated providers, Hetzner, OVHcloud, phoenixNAP, and Vultr offer single-tenant servers with transparent pricing and, in some cases, IPMI access that most hyperscalers do not expose. The best choice depends on which layer of control matters most to your workload.

What is bare metal cloud?

Bare metal cloud is a hosting model where you rent an entire physical server with no provider hypervisor between your workload and the hardware. You get root or administrator access, the ability to install any operating system or hypervisor, and dedicated CPU, memory, storage, and network interfaces. Providers like Vultr, Hetzner, OVHcloud, and phoenixNAP provision bare metal servers through APIs, combining the control of dedicated hardware with the on-demand provisioning model of cloud computing.

Is bare metal faster than virtual machines?

Bare metal eliminates the hypervisor layer, which removes a source of overhead for CPU-bound, memory-bound, and I/O-bound workloads. For applications that are sensitive to latency jitter, such as real-time databases, HPC simulations, and financial trading systems, bare metal typically delivers more predictable performance. The absolute throughput difference varies by workload. Modern hypervisors add minimal overhead for many general-purpose applications, so the performance gain from bare metal is most pronounced for workloads that are sensitive to tail latency or that require direct hardware access.

Can I run my own hypervisor in the cloud?

Yes. AWS EC2 .metal instances, IBM Cloud bare metal servers, OCI bare metal shapes, Hetzner dedicated servers, OVHcloud dedicated servers, and Vultr bare metal all allow you to install and run your own hypervisor, including VMware ESXi, KVM, Proxmox, and Xen. Dedicated hosts on AWS and Azure provide single-tenant hardware, but they still run the provider’s hypervisor, so they are not suitable if you need to replace the hypervisor entirely.

Which cloud provider offers the most bare metal options?

AWS offers .metal sizes across many instance families spanning different processor families (Intel, AMD, Graviton) and accelerator configurations. OCI provides a strong selection of bare metal shapes including GPU instances. Among dedicated providers, OVHcloud offers a broad catalog of dedicated server configurations, spanning entry-level to high-performance storage and compute servers. The right comparison depends on whether you are counting instance families, processor options, or geographic availability.

Is Hetzner good for production?

Yes. Hetzner dedicated root servers and Hetzner Cloud instances run production workloads for many teams. Hetzner offers low pricing and KVM-over-IP remote console access on request for dedicated servers. The trade-offs for production use are limited geographic coverage (dedicated servers in Germany and Finland; Hetzner Cloud also in Ashburn, Hillsboro, and Singapore), fewer managed services compared to hyperscalers, and a support model geared toward technically self-sufficient customers. If your team can handle OS-level operations and you do not need managed databases or a global network backbone from the same provider, Hetzner is a strong option for production workloads.

What is the difference between a dedicated host and bare metal?

A dedicated host gives you a physical server where only your VMs run, but the provider’s hypervisor still sits between your workloads and the hardware. AWS Dedicated Hosts and Azure Dedicated Hosts work this way. Bare metal gives you the physical server with no provider hypervisor. You control the entire software stack from the boot process up. The distinction matters for licensing (some software licenses count every physical core the software can run on), for performance (bare metal eliminates hypervisor overhead), and for compliance (some internal security policies and customer contracts require that no provider software layer sit between the workload and the hardware).

How does Control Plane run workloads on servers I control?

Control Plane’s Managed Kubernetes (MK8s) creates and manages Kubernetes clusters, including upgrades and autoscaling, on infrastructure you choose. You build an MK8s cluster on your bare metal servers or on-premises VMs with the Generic provider. Your servers become locations within a Control Plane Global Virtual Cloud (GVC), and workloads deployed to that GVC run across all its locations from a single workload definition. A workload can span your own bare metal hardware and Control Plane-managed regions on AWS, GCP, Azure, and OCI simultaneously. You keep full control of the hardware, OS, and physical network. Control Plane handles Day 2 operations for you, including autoscaling, built-in observability, and security.

Do I need a dedicated ops team for bare metal?

Not necessarily a large one. Operating bare metal servers in production requires someone to handle OS patching, security hardening, monitoring, failover design, capacity planning, and hardware failure coordination with the provider. At minimum, one engineer or a small platform team owns these responsibilities. Using a platform like Control Plane’s MK8s on top of bare metal takes over Kubernetes and Day 2 operations, including autoscaling, built-in observability, and security, but the infrastructure-layer responsibilities remain yours.

Talk to the Control Plane team or sign up free to run workloads on servers you control with Day 2 operations handled for you.

Originally published on the Control Plane blog.