7 Best AI Workloads for Spot GPU Instances in 2026

The latest generation of GPUs is in high demand, and until now, getting your hands on newer chips like NVIDIA’s HGX™ B300 or AMD’s Instinct™ MI350X and MI355X typically meant signing a long-term contract, whether or not your workload needed a dedicated reservation. Spot GPU Droplets®, now in public preview on DigitalOcean, change that. They provide access to idle capacity on these new chips—also known as GPU spot instances—priced below on-demand rates, with no contract required. It’s a direct expression of owning your costs on AI infrastructure: you pay for compute only when you’re using it, at a rate that reflects real-time supply.

When planning usage, it’s important to understand that Spot GPU Droplets are interruptible. DigitalOcean can reclaim one at any time under our Spot capacity terms, aiming to provide at least two hours’ notice by email first. That makes Spot an ideal fit for workloads that are fault-tolerant, checkpointable, and don’t have a live end user depending on that exact instance staying up—from batch training runs to academic research jobs. It also means that DigitalOcean provides a meaningfully longer reclaim window than AWS, Azure, and Google Cloud’s (GCP) own spot and preemptible offerings.

Let’s dive into what you need to know about DigitalOcean’s Spot GPU instances and how to use them for your workloads.

Key takeaways:

• Spot GPU Droplets give contract-free, no-application access to NVIDIA’s HGX B300 and AMD’s Instinct MI350X/MI355X GPUs, priced below on-demand rates.
• DigitalOcean aims to give at least two hours’ notice by email before reclaiming a Spot Droplet—longer than AWS’s, Azure’s, or Google Cloud’s typical interruption windows.
• Pricing is billed per hour, and your rate is locked in for the life of a Droplet once you provision it.
• The best fits for Spot GPU Droplets include fault-tolerant, checkpointable workloads with no live end user waiting on that specific instance. In other words, batch training, batch inference, sweeps, rendering, and research.

What are Spot GPU instances?

GPU Spot instances are a cloud provider’s way of selling unused capacity at a discount. Instead of reserving a specific machine for you indefinitely, the provider hands you access to whatever’s currently idle, at a price that reflects demand. The provider can take it back if a higher-priority workload needs that capacity. It’s the same basic mechanism AWS, Azure, and Google Cloud have offered with general-purpose compute. DigitalOcean is now extending spot pricing to GPU Droplets.

How Spot pricing works on DigitalOcean

Spot GPU Droplets are billed hourly, calculated per second like DigitalOcean’s on-demand GPU Droplets. The Spot rate itself is variable and adjusts based on how much idle GPU capacity is available—but once you successfully provision an instance, your price is locked in for the life of that specific Droplet, even if the market rate moves afterward.

For example, if you provision at a given Spot rate and that rate later rises, you keep paying your original rate until the Droplet is reclaimed or you provision a new one, which would pick up whatever the prevailing Spot rate is at that moment. There’s no proration if a Droplet is reclaimed mid-hour: you’re only billed for the exact time it ran.

At launch, Spot rates are:

• $11.19/GPU/hr for B300
• $7/GPU/hr for MI350X
• $8/GPU/hr for MI355X

All rates are regularly reviewed and subject to change based on capacity and demand. Current Spot Plans pricing can be referenced with our GPU Droplets pricing.

How DigitalOcean’s Spot GPUs differ from AWS, Azure, and GCP

The mechanics are familiar if you’ve used AWS EC2 Spot Instances, Azure Spot VMs, or Google Cloud’s Spot VMs—but the interruption window is not.

Here’s each provider’s typical advance notice before reclaiming a spot instance:

Figures above as of August 2026—see each provider’s own documentation for current terms.

The choice to provide two hours’ notice is designed to give teams enough time to checkpoint a training run, finish a batch job, or otherwise wind down cleanly instead of scrambling. The notice is sent on a best-effort basis to your team’s Owner(s) and the Droplet’s creator.

Contract-free access to B300 and MI350X/MI355X GPUs

Our public preview GPU lineup is exactly the hardware that’s typically been hardest to access without a reservation: NVIDIA’s HGX B300 and AMD’s Instinct MI350X and MI355X.

Previously, getting time on this generation of chips usually meant negotiating a long-term commitment. With Spot GPU Droplets, if capacity is available for the GPU model and data center you want, the Spot plan simply shows up as an option in the standard GPU Droplet creation flow—no contract, application, or opt-in required.

Spot GPU Droplets are fully integrated into DigitalOcean’s existing tooling, too: it’s available through the same Console and API as on-demand GPU Droplets and billed to the same account. There’s no separate Spot marketplace, standalone account, or fragmented invoice to manage.

DigitalOcean’s AMD collaboration has already shown what this hardware can do in production: Character.AI doubled production inference throughput and cut cost per token by 50% running on DigitalOcean’s AMD Instinct MI300X and MI325X platform. Results vary by workload and configuration.

Access is drawn from idle capacity in DigitalOcean’s existing GPU fleet, so what’s available depends on real-time demand in a given region.

During public preview, Spot GPU Droplets are available across four regions:

How to decide when to use Spot GPU instances

Spot GPU Droplets are a good fit if your workload is:

• Fault-tolerant: A job that gets interrupted can resume or restart without breaking anything downstream.
• Checkpointable: Progress is saved somewhere durable (like Spaces), not just on a local disk.
• Free of a live end user: No one is waiting in real time on that specific instance staying up.
• Cost-sensitive rather than uptime-sensitive: You’d rather save on compute than pay a premium for guaranteed availability.
• Batch or asynchronous by nature: Training, inference, sweeps, rendering, preprocessing, or research jobs.

Spot isn’t the right fit for latency-sensitive workloads, anything stateful that doesn’t checkpoint, or anything customer- or production-facing that needs guaranteed uptime. If your workload has a real user (internal or external) depending on that instance staying available, or if losing an in-progress job would be genuinely costly, an on-demand GPU Droplet is the better call.

The best AI workloads for Spot GPU instances

The common thread for the best AI workloads for Spot GPU instances? None of them have a live end user waiting on that exact instance to stay running. If a job gets interrupted, it can checkpoint, resume, or simply restart somewhere else without anyone downstream noticing.

1. Batch model training and fine-tuning with checkpointing

Training runs and fine-tuning jobs that checkpoint regularly are one of the most ideal matches for Spot GPU instances. For example, fine-tuning an open-weight LLM one checkpoint at a time.

If a Droplet is reclaimed, you resume from the last checkpoint, instead of losing the whole run. Spot GPU Droplets are ephemeral at reclamation, so checkpointing to Spaces (or another persistent store) periodically isn’t optional—it’s the mechanism that makes Spot viable for training in the first place.

DigitalOcean customers like Traversal and Workato run training and fine-tuning workloads on GPU Droplets today, the same fleet Spot draws its idle capacity from.

2. Batch and asynchronous inference

Large-scale, non-interactive inference doesn’t need a guaranteed-uptime GPU sitting behind it. This includes tasks such as scoring a dataset overnight, generating embeddings for a document corpus, and running a batch of image or video generation jobs. If a batch gets interrupted partway through, you pick up where you left off.

DigitalOcean’s Batch Inference pairs naturally with Spot GPU Droplets capacity for exactly this kind of job.

3. Hyperparameter sweeps

Sweeps are inherently parallel and disposable by design: you’re running many short, independent trials to find the best configuration, and any individual trial that gets interrupted can simply be relaunched. The cost savings can compound quickly for this use case, since sweeps often mean dozens or hundreds of runs across newest-generation GPUs like the B300 or MI350X.

Follow along as MLWorks creator Mayur Jain walks through how to set up and run a parameter sweep.

4. Rendering and simulation jobs

Rendering pipelines, physics simulations, and similar compute-heavy, checkpoint-friendly workloads suit Spot GPU Droplets for the same reason batch inference does: the job is long-running and resumable. No live user is staring at a loading screen waiting on that specific instance.

5. Experimentation and prototyping

Early-stage work—testing new architecture, prototyping a pipeline, and validating an idea before committing real budget to it—is naturally short-lived and low-stakes if interrupted. Spot GPU Droplets let teams get hands-on with B300 or MI350X/MI355X capacity for this kind of exploration without justifying a long-term reservation for hardware they’re not yet sure they need at scale.

6. Data preprocessing pipelines

Cleaning, transforming, and preparing large datasets ahead of training is typically batch-oriented and checkpoint-friendly. This makes it a reasonable fit for Spot GPU Droplets capacity, especially for preprocessing steps that are themselves GPU-accelerated.

7. Academic and research workloads

Research computing tends to be exploratory, budget-constrained, and tolerant of the occasional restarted job. It’s also a segment that’s historically had a harder time getting access to newest-generation GPUs at all, contract or not.

Regions and availability during public preview

Spot GPU Droplets are available in public preview across four regions—RIC1, MKC1, ATL1, and MEM1—with GPU-to-region availability detailed in the table earlier in this article. Because Spot draws from idle capacity in DigitalOcean’s existing GPU fleet, real-time availability for a given GPU model and region depends on demand at the moment you go to create a Droplet.

DigitalOcean expects to expand to additional GPU types, regions, and capacity over time based on feedback from this preview, and is evaluating additional safeguards, such as automated snapshotting at reclamation and DigitalOcean Kubernetes (DOKS) node pool support, for future releases. These are current plans, not commitments, and are subject to change.

Spot GPU instances FAQs

Which cloud providers give the longest advance notice before reclaiming a spot GPU instance?

DigitalOcean gives the longest advance notice among major providers. Spot GPU Droplets aim for a best-effort two hours notice by email before reclamation, compared to AWS EC2 Spot Instances’ roughly two-minute notice, Azure Spot VMs’ roughly 30-second Preempt signal, and Google Cloud Spot VMs’ roughly 30-second preemption notice (with an opt-in 120-second window available in preview). Figures are as of August 2026—see each provider’s own documentation for current terms.

What happens when a spot GPU instance gets reclaimed mid-training?

The Spot GPU Droplet is ephemeral at reclamation, so any state that isn’t saved somewhere durable is lost. DigitalOcean sends a best-effort email notice at least two hours ahead to your team’s Owner(s) and the Droplet’s creator. That’s why checkpointing regularly to Spaces (or another persistent store) during training matters.

Should I use Spot GPUs or on-demand GPUs for model training?

It depends on the training job. Spot GPU Droplets are a good fit if the run is fault-tolerant, checkpoints regularly, and doesn’t have a live end user depending on that specific instance staying up. You trade a small risk of interruption for a lower, below-on-demand rate. On-demand is the better call for latency-sensitive, stateful training that doesn’t checkpoint, or any run where losing progress or availability would be genuinely costly.

What AI workloads are best suited for spot GPU instances?

The best-suited workloads share one trait: no live end user waiting on that exact instance staying up. That includes batch model training and fine-tuning with checkpointing, batch and asynchronous inference, hyperparameter sweeps, rendering and simulation jobs, experimentation and prototyping, data preprocessing pipelines, and academic and research workloads. If a job gets interrupted, it can checkpoint, resume, or restart elsewhere without anyone downstream noticing.

Where can I get NVIDIA B300 or AMD MI350X/MI355X GPUs without a long-term contract?

On DigitalOcean, Spot GPU Droplets put NVIDIA’s HGX B300 and AMD’s Instinct MI350X/MI355X in the standard GPU Droplet creation flow, now in public preview. There’s no contract, application, or opt-in required—as long as capacity is available for the GPU model and data center you select.

Where can I get a GPU spot instance right now?

Availability is tied to real-time idle capacity, so it varies by provider, region, and moment. On DigitalOcean, the Spot GPU Droplets option only appears in the Droplet creation flow when capacity actually exists for your chosen GPU model and region. During public preview, that capacity spans four regions: RIC1 (B300, MI350X), MKC1 (B300), ATL1 (MI350X), and MEM1 (MI355X). Check the console at the moment you go to provision, since what’s available can shift with demand.

Run your AI workloads with Spot GPU Droplets

Spot GPU Droplets are one of the fastest ways to get hands-on with next-generation GPU capacity for fault-tolerant, checkpointable workloads—no long-term commitment required.

With Spot GPU Droplets, you get:

• Contract-free access to B300, MI350X, and MI355X GPUs
• Pricing below on-demand rates
• Price locked at creation for the life of the Droplet
• Per-second billing with no proration
• At least two hours’ reclaim notice by email (best effort)

Launch a Spot GPU Droplet from the DigitalOcean console →

NVIDIA, HGX, AMD, and Instinct are trademarks or registered trademarks of their respective owners. Droplet® is a registered trademark of DigitalOcean, Inc.