Compute
GPU compute, on your terms.
Spin up GPU instances for experiments, run bare metal servers when every cycle counts, and reserve dedicated capacity when demand settles. One account, one platform, and a clear path from a single GPU to a full cluster.
GPU Cloud Instance
1–8 GPUs · virtual machine
Development · fine-tuning
On demandBare Metal GPU
8-GPU server · no hypervisor
Training · custom stacks
Full serverReserved GPU Capacity
Committed block · fixed term
Sustained training · inference
ReservedDedicated Compute
Single-tenant servers
Isolated workloads
Isolated
01GPU Cloud Instances
Start fast. Scale when it works.
GPU Cloud Instances are virtual machines with dedicated GPUs attached — the fastest way to go from idea to running code. Use them for prototyping, fine-tuning, evaluation, and inference, and resize as the work changes.
- Single- and multi-GPU instance sizes
- Preconfigured images with drivers and ML frameworks
- Persistent volumes that outlive the instance
- SSH, Jupyter, and Kubernetes access
GPUs per instance
Environment
- Access
- SSH · Jupyter
- Storage
- Persistent volume
- Drivers
- Preinstalled
02Bare Metal GPU
Every cycle, every lane, yours.
Bare Metal GPU gives you whole servers with no hypervisor in the path. You get root access, direct access to GPUs, NVLink, and network adapters, and the freedom to run your own operating system and drivers — ideal for distributed training and performance-sensitive stacks.
- Full eight-GPU servers, single tenant
- No virtualization overhead
- Custom OS, drivers, and kernel settings
- Ready to join high-performance cluster fabrics
Root access
Your OS and stack
Direct I/O
GPUs, NVLink, NICs
Cluster-ready
Join a fabric
03Reserved GPU Capacity
Capacity you can plan around.
When training schedules or inference traffic become predictable, reserve a block of GPUs for a defined term. Reserved capacity gives your roadmap a fixed foundation and makes GPU spend easier to forecast.
- Committed GPUs for the length of your program
- Instances or bare metal under one reservation
- Predictable budgeting for finance teams
- Upgrade path to dedicated compute and clusters
- 01
On demand
Start immediately while requirements take shape.
- 02
Reserved
Commit a block of GPUs for a defined term.
- 03
Dedicated
Move to single-tenant compute or a full cluster.
04Dedicated Compute
Isolation without compromise.
Dedicated Compute reserves physical servers exclusively for one customer. No other tenant shares the hardware, which simplifies security reviews and keeps performance consistent for regulated or sensitive workloads.
- Single-tenant servers, physically isolated
- Network segmentation per customer
- Consistent performance with no noisy neighbors
- A natural step toward Dedicated GPU Clusters
Multi-tenant
Shared hosts
Capacity from a multi-tenant pool
Single-tenant
Dedicated Compute
Every server reserved for one customer
05GPU Resource Management
Know where every GPU-hour goes.
Organize compute into projects, set quotas for teams, and see usage by workload. GPU Resource Management keeps shared capacity fair and makes it easy to tie GPU spend back to the work it supports.
- Projects, teams, and role-based access
- Quotas and limits per team or project
- GPU utilization and usage reporting
- Spend attribution by project and workload
- research-pretrain46% / 50%
- product-inference28% / 30%
- finetune-support11% / 15%
- eval-batch4% / 5%
06Getting started
From sign-up to running workloads.
- 01
Choose your compute
Instances for speed, bare metal for control, reserved or dedicated for predictability.
- 02
Pick an environment
Preconfigured images with drivers and frameworks, or bring your own container.
- 03
Connect your way
SSH, Jupyter, Kubernetes, or Slurm — whichever your team already uses.
- 04
Run and observe
Track GPU utilization, jobs, and spend by project as you go.
07Compare
Choose the right way to compute.
| Option | GPU Cloud Instances | Bare Metal GPU | Reserved GPU Capacity | Dedicated Compute |
|---|---|---|---|---|
| Best for | Development, fine-tuning, inference | Performance-sensitive training | Sustained, predictable demand | Isolation and compliance |
| Tenancy | Shared hosts, isolated VMs | Whole server, single tenant | Instances or bare metal | Single-tenant servers |
| Access | SSH, Jupyter, Kubernetes | Root access, your OS | Same as underlying compute | Root access, your OS |
| Commitment | None — use as needed | Flexible or reserved | Fixed term | Reserved term |
| Scales to | Multi-GPU instances | Multi-node jobs | Larger reserved blocks | Dedicated GPU Clusters |
Need multi-node scale with a dedicated fabric? Explore Dedicated GPU Clusters.
Next step
Get GPUs working for you.
Tell us what you're running and how fast you need to move. We'll recommend the compute option — and the terms — that fit.
