Some links here are affiliate links — if you buy through them we may earn a commission at no extra cost to you. It never changes what we recommend. Full disclosure.
Disclosure: This article contains affiliate links. If you sign up through them, we may earn a commission at no extra cost to you. Specifications were checked on DigitalOcean's site in October 2026. GPU pricing changes frequently — always confirm on DigitalOcean's live pricing page.
DigitalOcean now offers three distinct ways to get GPU compute: on-demand GPU Droplets, reserved single-tenant Bare Metal GPUs, and Serverless Inference, where you never see a GPU at all. Picking the wrong one is the most common way AI projects overspend — either by paying for a full eight-GPU server that sits idle, or by renting by the hour for a workload that runs around the clock.
This guide explains what each option is, what the hardware looks like, and how to match them to real workloads.
The Three Options in One Table
| Serverless Inference | GPU Droplets | Bare Metal GPUs | |
|---|---|---|---|
| What you get | An API endpoint for hosted models | A virtual machine with 1 or 8 GPUs | An entire 8-GPU physical server |
| Billing | Per token/request | On demand, billed by usage time | Reserved capacity via sales |
| Setup time | Minutes | Minutes | Arranged with DigitalOcean |
| Tenancy | Shared platform | Virtualized | Single-tenant, no neighbours |
| Best for | Calling models from an app | Experiments, fine-tuning, bursty inference | Sustained training and large-scale inference |
GPU Droplets: On-Demand GPUs in Minutes
GPU Droplets are virtual machines with dedicated GPUs attached, created from the same control panel and API as any other Droplet. DigitalOcean's lineup includes NVIDIA H100, H200 and L40S and AMD Instinct MI300X, available as single-GPU machines or in eight-GPU configurations for multi-GPU work.
Why teams choose them:
- Speed: spin one up, run a job, destroy it. No procurement.
- Granularity: start with a single GPU for prototyping instead of committing to eight.
- Familiar tooling: same API, same networking, same project structure as the rest of your DigitalOcean resources, with pre-installed AI/ML-ready images to skip driver setup.
Watch the bill: like other Droplets, GPU Droplets keep accruing charges while they exist, even when powered off. Destroy them (after snapshotting anything you need) when a job finishes. Third-party price trackers listed on-demand H100-class GPU Droplets roughly in the $3.40–$4.50 per GPU-hour range during 2026; check DigitalOcean's pricing page for the current rate in your region.
Best for:
- Prototyping and benchmarking models
- Fine-tuning runs measured in hours or days
- Batch inference jobs that run on a schedule
- Teams that want to try several GPU types before committing
(ad) Need a GPU for a fine-tune or benchmark this week? Launch a DigitalOcean GPU Droplet — H100, H200, L40S or MI300X, in minutes.
Bare Metal GPUs: The Whole Machine, Reserved
Bare Metal GPUs give you raw, single-tenant access to an entire eight-GPU server. There is no hypervisor between you and the hardware and no other customers on the machine. You can install custom drivers and software, and the configurations are serious:
| Server | GPU memory (total) | CPUs | System RAM | NVMe storage |
|---|---|---|---|---|
| NVIDIA HGX H100 | 640 GB | 2 × Intel Xeon Platinum 8468 | 2,048 GiB | 61.44 TiB |
| NVIDIA HGX H200 | 1,128 GB | 2 × Intel Xeon Platinum 8592+ | 2,048 GiB | 61.44 TiB |
| AMD Instinct MI300X | 1,536 GB | 2 × Intel Xeon Platinum 8568Y | 2,048 GiB | 61.44 TiB |
Bare Metal GPUs are available in DigitalOcean's NYC1 (New York) and AMS3 (Amsterdam) data centers, with capacity reserved by talking to DigitalOcean's team. Pricing includes the substantial local NVMe storage, and DigitalOcean provides high-touch engineering support for these deployments.
Why teams choose them:
- Sustained utilization: if GPUs run most of the day, every day, reserved bare metal is typically more economical than on-demand hours.
- Performance and control: no virtualization overhead, full control over the software stack, and all eight GPUs connected within one chassis for distributed training.
- Privacy and isolation: single-tenant hardware simplifies data-isolation requirements.
- Memory for big models: the H200 and MI300X configurations carry very large GPU memory pools, useful for serving large models with long context windows.
Best for:
- Training or continued pre-training of foundation-scale models
- Production inference serving large models at high, steady volume
- Teams with predictable, long-running GPU demand
(ad) Running GPUs flat out every month? Reserve DigitalOcean Bare Metal GPUs — single-tenant HGX H100, H200 or AMD MI300X servers.
Serverless Inference: When You Don't Need a GPU at All
If your goal is to call a model from an application — chatbots, summarization, classification, agents — you may not need to rent GPUs. DigitalOcean's AI platform offers serverless inference, where you pay per use and DigitalOcean runs the infrastructure. For most apps with spiky or moderate traffic, that is cheaper and far simpler than keeping a GPU warm.
(ad) Explore DigitalOcean's AI Platform and serverless inference — call models by API without managing GPUs.
We compared the economics in detail in GPU rental vs serverless inference: which actually costs less.
A Simple Decision Framework
- Are you calling a model rather than training one? Start with serverless inference.
- Is GPU usage occasional, experimental or measured in hours? Use GPU Droplets and destroy them between runs.
- Will GPUs be busy most hours of most days for months? Talk to DigitalOcean about Bare Metal GPUs.
- Not sure yet? Prototype on a single-GPU Droplet, measure utilization for two to four weeks, then decide.
Tip: Utilization is the number that decides everything. Log GPU usage during your prototype phase (nvidia-smi or your framework's metrics). If average utilization over a month is low, on-demand or serverless wins. If it is consistently high, reserved hardware almost always does.
Don't Forget Storage and Data Movement
Training jobs live or die on how fast they can read data. Three DigitalOcean services matter here:
- Network File Storage — a managed NFS share several GPU machines can mount at once, with a high-performance tier designed for GPU workloads. See our Network File Storage guide.
- Spaces object storage — inexpensive storage for datasets, checkpoints and model artifacts.
- Volumes — block storage attached to a single Droplet.
Common Mistakes
- Leaving GPU Droplets powered off but not destroyed, and paying for them anyway
- Renting eight GPUs when the code only uses one
- Reserving bare metal before measuring real utilization
- Training from slow storage, leaving expensive GPUs waiting on data
- Running a GPU for an app that a per-token inference API would serve more cheaply
Frequently Asked Questions
What is the difference between a GPU Droplet and a Bare Metal GPU?
A GPU Droplet is a virtual machine you create on demand, with one or eight GPUs. A Bare Metal GPU is a reserved, single-tenant physical server with eight GPUs and no virtualization layer.
Which GPUs does DigitalOcean offer?
Its GPU lineup includes NVIDIA H100, H200 and L40S and AMD Instinct MI300X. Bare metal servers come in HGX H100, HGX H200 and MI300X configurations.
Where are DigitalOcean Bare Metal GPUs located?
NYC1 in New York and AMS3 in Amsterdam at the time of writing, with more locations planned.
Can I install my own drivers on Bare Metal GPUs?
Yes — single-tenant bare metal gives you full control of the software stack.
The Bottom Line
Use serverless inference to call models, GPU Droplets to experiment and run bursty jobs, and Bare Metal GPUs once you have proven, sustained demand. DigitalOcean's advantage is that all three live in one account, so you can move between them as a project matures without switching providers.
(ad) Start small and scale up: launch a GPU Droplet today, and talk to DigitalOcean about Bare Metal GPUs when utilization justifies it.