Vultr Bare Metal

What Vultr Bare Metal actually is

Vultr Bare Metal is the single-tenant line inside the Vultr platform: dedicated physical servers with no virtualization layer, provisioned and billed through the same console, API, and account as a $5 virtual machine. That last part is the pitch, and it is genuinely different from the traditional dedicated-server market. Where legacy dedicated hosts sell you a server after a ticket, a contract, and a day of waiting, Vultr sells bare metal the way it sells cloud compute — on-demand, hourly, automated, and tear-down-able at 3am without emailing anyone.

The product has two shapes. The first is on-demand bare metal servers: whole physical machines — CPU, memory, disks, NICs — assigned to you alone, deployable in minutes with your choice of Linux or Windows and full control of the OS and firmware-level features that matter (custom kernels, hypervisor nesting, hardware passthrough). The second is Vultr Clusters: self-service groups of bare-metal servers, including GPU machines, connected over high-performance networking and created through the console or API with no reservations and no special requests. On the GPU side the catalog is heavy metal — for example an NVIDIA L40S node with 8 x L40S 48 GB GPUs, 2 x Intel Gold 6448H (64 cores / 128 threads), 2 TB RAM, and 4 x 3.84 TB NVMe — aimed at AI training, inference, and rendering at scales where renting eight GPUs by the hour beats signing an enterprise order form.

Everything around the server inherits the platform: the same API and Terraform provider, the same Marketplace images, the same object and block storage (including S3-compatible Vultr Object Storage and storage gateways for bare metal), load balancers, DNS, cloud firewalls, the CDN, Direct Connect, and bring-your-own-IP via BGP announcement in any Vultr region. Layer 3 and layer 4 DDoS mitigation is part of the platform. Automatic backups can be enabled on bare-metal instances (snapshots, notably, are an instance-VM feature and not part of the bare-metal workflow). This platform inheritance matters more than the spec sheet: bare metal on a cloud control plane is what separates this product from renting a box at a colo.

Pricing: what you really pay

Here the story gets more opaque than Vultr's usual transparency, and we want to be exact about what is and is not public. On the public pricing page, bare-metal pricing is described as "starting from" per-GPU-per-hour figures for pre-paid contract terms, with the note to contact sales to discuss pricing; on-demand pricing for bare metal is shown only in the Vultr console or via the API once you are a customer. In other words, there is no public CPU-bare-metal price list to quote. Our database recorded an entry price of $120/month; we could not confirm that on any vendor page in this session, so we have cleared it rather than repeat it. Third-party pricing aggregators still circulate a $120/month entry figure, but we do not treat aggregator numbers as verification.

What we can quote from Vultr directly: the visible GPU bare-metal "starting from" figure of $0.848 per GPU per hour on pre-paid terms, and a storage gateway add-on at $20/month for the highly available version (standalone is free). For context rather than comparison, the compute catalog that shares the platform runs $2.50/month for the smallest VM up to $3,840/month for a 96-vCPU optimized instance, and hourly billing is the default across the platform. Treat on-demand bare metal as hourly-billed, priced per machine class, and confirm the rate in the console before you plan around it.

The structural billing points carry over from the platform and are worth repeating because they apply to physical machines too: automatic backups, if enabled, add 20% of the instance's base price; extra egress is metered; Windows licensing is an add-on; and the terms of service are explicit that there are no refunds and that early termination of services before the billing-cycle end can attract a charge. For a product often rented in bursts — a training run, a migration, a load test — the no-refund policy is the fine print to read first.

Performance & infrastructure

The performance case for bare metal is simple and does not need embellishment: no hypervisor between your workload and the silicon. Full CPU cores with no steal, full memory bandwidth, direct disk access, and NICs you can push without contending with a noisy neighbor. Vultr's FAQ on the product is admirably clear about when that matters — HPC, AI/ML with dedicated GPU access, big-data analytics and databases, large-scale web and gaming platforms, and high-security environments that require physical isolation and complete hardware control. Compliance teams like single tenancy for the same reason: the attack surface and the co-tenancy questions shrink to zero.

The machines are current-generation enterprise hardware — the published L40S configuration shows dual Intel Gold 6448H processors and NVMe storage, and CPU-only classes cover the usual Xeon/EPYC shapes — but as with the pricing, exact per-class specifications for on-demand CPU bare metal are presented in the console rather than on a fixed public spec sheet. That is the one place where this product is less transparent than a traditional dedicated-server catalog, where you can usually compare chassis line by line before signing up.

Automation is the differentiator versus the rest of the bare-metal market. Deploy in minutes through the console or API, reinstall the OS yourself, attach Vultr storage via gateways, announce your own IP space over BGP, and script the whole lifecycle with the Terraform provider. Vultr Clusters adds the ability to compose multi-node bare-metal groups — including GPU clusters — without reservations or sales negotiation, which is exactly what machine-learning teams need when a training window opens and hardware must exist today, not in three weeks.

Reliability claims here rest on the same platform SLA as the rest of Vultr: a 100% uptime guarantee covering network and host-node availability, with credits requested through support. Physical servers fail differently from VMs — a dead disk is a hardware event with a human response — and Vultr does not publish bare-metal-specific replacement-time commitments we could find. Support is ticket-based; for physical-hardware issues that is the standard mechanism, but teams with hard recovery-time requirements should confirm response expectations before deploying.

Where it falls short

Pricing opacity is the first complaint. A product whose main competitor is the traditional dedicated-server market — where public price lists are the norm — hides its CPU bare-metal rates behind a login and points contract buyers to sales. We understand why (rates vary by class, region, and commitment), but it costs Vultr its usual transparency advantage, and it makes price comparison before account creation impossible.

Second, bare metal is not available uniformly across the 34-region platform. The footprint is a subset, and the public pages do not enumerate which regions stock which classes; GPU and large-memory machines concentrate in fewer sites than the VM catalog. Buyers targeting a specific metro should verify availability in the console early.

Third, the VM-oriented data-protection model does not fully transfer. Snapshots are not part of the bare-metal workflow; automatic backups can be enabled at a 20% surcharge but cover the instance's file system, not the philosophy of image-level rollback VM users are used to. Physical hardware means hardware failure modes, and the platform's promise of "deploy in minutes" is about provisioning, not about replacing failed components instantly.

Finally: no refunds, ticket-only support, and metered egress — the same platform-level frictions as the compute line, amplified by higher spend. At $120+/month per machine those terms deserve more scrutiny than they do at $5/month.

Who should pick this

Our pick for teams that want dedicated hardware with cloud ergonomics. If you need full machines — for CPU-bound HPC, database clusters with predictable I/O, game backends, compliance-sensitive workloads, or GPU training and inference — and you want to script deployment, billing hourly, without a sales cycle, this is one of the strongest options on the market. The AI story is particularly good: serious GPU nodes (L40S, A100, H100 classes), clusters without reservations, hourly economics, and the same API as the rest of your fleet.

It also fits organizations consolidating onto one provider: the ability to run VMs, bare metal, object storage, managed databases, Kubernetes, and load balancers under one account, one invoice, and one API removes the integration tax of a mixed vendor stack. Bring-your-own-IP via BGP is a further draw for teams migrating existing address space.

We would look at traditional dedicated-server or colocation providers when you need published per-chassis pricing, negotiated terms, hardware SLAs with replacement guarantees, or a specific metro that Vultr's bare-metal subset does not serve. And for pure VM workloads there is no reason to pay for physical isolation you do not need — the optimized compute lines cover that case at lower cost.

Key takeaways

• Vultr Bare Metal sells dedicated physical servers on-demand, hourly, through the same console/API/account as Vultr VMs — with Vultr Clusters for self-service multi-node and GPU groups, no reservations or sales calls required.
• Full hardware access (custom OS/kernels, hardware passthrough), platform add-ons (S3-compatible object storage, storage gateways, load balancers, DNS, CDN, BGP bring-your-own-IP), and L3/L4 DDoS mitigation.
• CPU bare-metal pricing is not public: on-demand rates appear only in the console/API and pre-paid GPU rates are quoted per GPU hour (visible example: from $0.848/GPU/hour); the $120/month entry in our database could not be verified and has been cleared.
• Platform SLA (100% network and host-node uptime, credits via ticket) applies; no bare-metal-specific hardware replacement commitments are published.
• Watch: no refunds in the terms, backup add-on at 20% of instance price, snapshots not part of the bare-metal flow, metered egress, and a bare-metal footprint that is a subset of the 34-region platform.
• In our view the pick for cloud-style automation over physical machines, especially GPU clusters; traditional dedicated hosts still win on published chassis pricing and negotiated hardware terms.

Sources

• https://www.vultr.com/products/bare-metal/
• https://www.vultr.com/pricing/
• https://docs.vultr.com/product-overviews/bare-metal
• https://docs.vultr.com/support/platform/billing/how-much-does-it-cost-to-enable-automatic-backups
• https://www.vultr.com/legal/sla/
• https://www.vultr.com/legal/tos/
• https://www.vultr.com/locations/
• https://stackfreeks.com/vultr-snapshot-vs-backup
• https://platform.tracxn.com/a/d/company/53326a87e4b0398cae238fe8/vultr