What are the best SaladCloud alternatives for GPU workloads?

SaladCloud is hard to beat on price for interruption-tolerant work on consumer GPUs, so the right alternative depends on what a lost node costs your job. Nodus is the pick for long jobs you don't want to babysit: it lists the NVIDIA RTX 4090 at $0.29/hr and restores your latest checkpoint on a replacement machine when one goes away. RunPod suits fixed-price pods, Vast.ai suits hands-on bargain hunting, and Modal suits short Python functions on datacenter GPUs.

We build Nodus, so read this with that in mind. Every SaladCloud, RunPod, Vast.ai and Modal fact links to that company's own pricing page or docs. The arithmetic is written out so you can redo it with today's prices.

What happens to a long job when a SaladCloud node goes away?

The job stops, and a fresh instance starts elsewhere with an empty disk. SaladCloud says this plainly. Its overview describes a distributed network of consumer GPUs, many of them individually owned PCs, where a node can go offline and SaladCloud automatically starts a replacement on another node.

The replacement does not carry your files. Each instance gets up to 250 GB of disk, kept across restarts on the same node, but a replacement on another node starts empty. SaladCloud tells you to keep checkpoints and results in external storage such as Amazon S3 or Cloudflare R2.

Its long-running tasks guide spells out the pattern for jobs longer than about 30 minutes. Your application regularly uploads its running state to cloud storage, and a new instance downloads it and resumes when it picks up an unfinished job. The pattern is sound. It is also code you write, test and pay storage for.

SaladCloud also publishes how often this happens, from a 2025 reliability test of 100 high-priority instances over 7 days. It counted 182 reallocations in 168 hours, and the interrupted runs averaged 35 hours of uptime. Of the 100 instances, 25 ran the full week without interruption.

What does one lost node cost on a 48-hour job?

It costs every hour since your last saved state, so checkpoints matter more than the hourly rate. We priced one job five ways to show it.

The job: 48 hours on one RTX 4090, interrupted once at hour 30. One interruption is a reasonable planning case. SaladCloud's test works out to 182 / (100 × 168) = 0.011 interruptions per instance-hour, about one every 92 hours, or 0.52 expected in 48 hours at High priority. Medium, Low and Lowest can also be preempted by higher-priority work, and SaladCloud publishes no rate for that, so treat one interruption on Lowest as a best case.

The 0.5 hours is the worst case: the node vanishes just before the next save. Image downloads are free on SaladCloud, and we left out checkpoint upload time and storage fees on both sides because they depend on your model size. Nodus also asks for an urgent checkpoint when capacity gives notice before it is reclaimed, so its loss is often smaller.

Saving state cuts the High-priority bill by $12.98; the tier you pick moves it further.

Two things stand out. Without saved state, one lost node turns a $21.12 High-priority job into $34.32. With your own checkpoints, SaladCloud Lowest is the cheapest line in the table, if the job tolerates preemption and you are willing to own the save-and-resume code.

How do Nodus checkpoints and failover work?

Nodus copies a state directory you choose while the job runs, then restores it into the next attempt. The checkpoints guide shows the contract. Your program writes model, optimizer and step files under /nodus/state and reads them back on start. Nodus decides when to snapshot and where to store it.

With interval: auto, a job gets at least four checkpoints over its expected run time, or you can fix the interval anywhere from 1 minute to 6 hours. This is the setting behind the 30-minute line above:

recovery:
  continuity: Checkpointed
  checkpoint:
    paths: [/nodus/state]
    interval: 30m

Failover follows the attempts and recovery rules. On a reclaim notice, Nodus requests an urgent checkpoint and prepares a replacement machine at the same time. A machine silent for 60 seconds on shared capacity is declared lost and replaced. The replacement starts only once the old machine is provably gone, and Nodus accepts checkpoints only from the current attempt, so a returning machine cannot overwrite its successor's work.

Restores bring back files, not process memory. Your program still has to load its own checkpoint, exactly as on SaladCloud. The difference is that you skip the upload, download and storage plumbing. We walk through the training side in how to resume training after spot preemption.

How do SaladCloud and its alternatives compare?

The table uses the RTX 4090 where each platform lists one, because that is SaladCloud's flagship class at 24 GB.

Prices checked November 13, 2026: SaladCloud pricing and priority levels, Nodus pricing, RunPod pricing and Pods pricing docs, Vast.ai pricing, pricing overview and instance types, Modal pricing and preemption docs. Modal's L4 rate of $0.000222 per second is multiplied out to an hour. Nodus rates are listed as the hourly supply price on its pricing page.

Nodus costs less than SaladCloud High, Vast.ai's median and both RunPod tiers, and more than SaladCloud Lowest.

Which SaladCloud alternative fits your workload?

Pick by job length and how much state the job carries, then by price.

• Long training, fine-tuning or simulation you don't want to watch: Nodus, because recovery is the platform's job and you can move up to H100-class GPUs when 24 or 32 GB is not enough.
• A fixed-price machine you SSH into and keep for days: RunPod on-demand Pods, which its docs say cannot be displaced by other users. Our Vast.ai alternatives post compares it with marketplace pricing.
• The lowest price on a specific card, with you picking the host: Vast.ai, where on-demand rentals are never interrupted and interruptible ones pause and resume on the same machine.
• Short Python functions and endpoints that scale to zero: Modal, which bills per second and restarts a preempted function on the same input.

Our decision rule: if a lost node would cost you more than an hour of work and you have not built save-and-resume to external storage, don't run it on a preemptible tier anywhere. Either build that code and use SaladCloud Lowest, or let Nodus handle it. If the job is minutes long and retryable, the cheapest tier wins.

Where is SaladCloud the better choice?

SaladCloud wins on stateless, retryable work that fits on a consumer GPU. Its pricing page positions the Lowest tier for retryable batch work and stateless inference behind a queue. That is where its prices are strongest: an RTX 4090 at $0.16/hr, vCPU and RAM included, billed per second only while the container is running.

It also wins on breadth of consumer cards at scale. The same page shows live counts of hundreds of schedulable GPUs in most listed classes. Its long-running tasks guide cites a transcription benchmark of 1 million hours of video on 100 GPUs within a week. If your work is image generation, transcription or embedding jobs pulled from a queue, losing a node costs one retry, and SaladCloud is the cheaper choice.

It is a weaker fit when a request runs past its Container Gateway's 100-second response timeout and you have no queue, per the same guide. It also tops out at 32 GB of VRAM on its listed consumer cards.

How do you move a SaladCloud job to Nodus?

Point the job at a state directory and run it with a checkpoint path and a spend cap. Install the CLI and sign in:

pip install nodus-compute
nodus login

Then make your program write its progress to /nodus/state and load it on start, and launch it:

nodus run --name long-job --gpu L4 --image nodus/pytorch --checkpoint /nodus/state --max-cost 25 -d -- python train.py

Swap L4 for the GPU type you need from the pricing page. If your SaladCloud code already uploads state to S3 or R2, it keeps working; you can delete that code later once the platform checkpoints cover it.

FAQ

Can I connect my SaladCloud account to Nodus?

No. Nodus can connect some cloud accounts for bring-your-own-cloud work, and SaladCloud is not one of them. You move the container and its code, not the account.

Does SaladCloud charge for the time a replacement instance spends starting?

No. SaladCloud bills only while an instance is running, and allocation, image download and container start are free. In its 2025 test, half of instances were online within 10 minutes and 90% within 40 minutes, so a replacement costs time even when it costs no money.

Is SaladCloud High priority safe for long jobs?

High priority is not preempted by other workloads, but SaladCloud's docs say it is still reallocated when the node disconnects. Save state either way.

Does Nodus have a free tier like Modal's monthly credits?

No. Nodus uses prepaid credits, and its public starter credit offer has ended. Modal's Starter plan includes $30 of free credit a month.

Originally published on the Nodus blog.