Start Building →
Bare Metal GPU

The whole machine.
Not a slice of it.

You get a full node, reserved entirely for you. No sharing, no virtualization layer, consistent performance for the length of your term.

node-04 · US-EAST provisioning
GPUsready
Storageready
Networkready
SSH accesssetting up
1 dedicated node · 1-month term
Fixed price
Monthly billing
Full node
No shared tenancy
Root access
Full configuration control
Available silicon

Reserve your node.

B200

New
192GB HBM3eBlackwell
from$5.90/GPU-hr
Deploy
What is bare metal

Nothing between you and the hardware.

Most cloud GPU services rent you a slice of a machine, not the whole thing. Bare metal is different: you get an entire node, with nothing else running on it. No virtualization, no sharing, no neighbors, just your workload talking directly to the hardware. Think of it like renting a desk in a shared office versus leasing the whole building. Same address, very different control.

No noisy neighbors

What running alone on the hardware gets you.

  • Full performance, all the time
  • No scheduler, no contention
  • Nothing else on the silicon

Bare metal

What you get with this tier.

  • Full node, one tenant
  • Consistent, dedicated performance
  • Fixed term, prepaid

Why it matters

Same city, same address, different control.

  • No virtualization overhead
  • Root-level access to configure
  • Built for sustained workloads
How it works

Reserve it. Get access. Run it for the length of your term.

packet.ai
Why Bare Metal

Built for workloads that need the whole machine.

Dedicated power, no sharing

The whole machine's performance is yours, consistently. No other workload is competing for the same resources.

Full control

Root-level access to configure the environment however your workload needs it.

Predictable pricing

A fixed, prepaid price for the length of your term. No hourly meter, no surprise bill at the end of the month.

Built for

Made for the long haul.

AI/ML training and fine-tuning

Teams training or fine-tuning large models that need guaranteed, sustained GPU horsepower for the length of a real training run.

  • Full performance, no contention
  • Predictable cost over the whole term
  • Root access to tune the environment

Training and batch workloads

Teams running training, fine-tuning, or batch workloads on a single node, with no cluster coordination or synchronized bring-up required.

  • Complete node, one price
  • Every GPU comes online together
  • Full node bandwidth, no shared network

Compliance and reliability

Workloads that need isolation, consistent performance, or dedicated hardware to meet compliance or reliability requirements.

  • Single-tenant hardware
  • Consistent, predictable performance
  • Enterprise-friendly invoicing
FAQ

Bare Metal GPU, answered.

For anything not here, reach help@packet.ai.

What is one node?
One node is 8 GPUs.
What does "bare metal" actually mean?
It means you get an entire node, not a slice of one shared with other customers. No virtualization layer, no other workloads on the same hardware.
How is this different from your other GPU options?
Bare Metal is the dedicated, prepaid tier. Instead of paying by the hour, you commit to a fixed term and get a full node reserved just for you. If you want to pay hourly instead, our on-demand options are built for that.
How many nodes do I get?
Bare Metal is provisioned as a single dedicated node per order, at one price, with every GPU in that node coming online together. If you need multiple nodes networked together for distributed training, take a look at GPU Clusters.
How do I pay?
By secure hosted invoice or direct bank wire. There's no credit card or hourly metering, which fits standard enterprise procurement.
How do I access my node once it's provisioned?
Each node comes with its own SSH access and connection details, shown in your dashboard. Credentials are never sent by email.
What happens when my term is ending?
Your dashboard tracks the contract end date, and you'll get automatic email reminders before the term ends so nothing lapses without warning.