Nodus vs Northflank: which is better for GPU training?

If the training job is the whole point, use Nodus; if it is one service inside an app you already deploy on Northflank, stay on Northflank. Nodus lists lower NVIDIA L4, A100, H100 and B200 rates than Northflank Cloud and restores your checkpoint on a new machine when one is lost. Northflank lists the lower 8 x H200 rate and runs GPUs across more kinds of your own cloud accounts.

We build Nodus, so we kept this checkable. Every Northflank fact links to its own pricing page or docs, the arithmetic is shown once, and there is a full section on where Northflank is the better pick.

How do Nodus and Northflank compare at a glance?

They start from different places. Northflank is a full deployment platform: services, jobs, databases, builds and pipelines, where a GPU is a resource you add to a service or job. Nodus is an AI compute platform: you hand it a command or a TrainingJob, and it picks capacity, saves checkpoints and replaces lost machines until the run finishes.

Sources: Northflank pricing, GPUs on Northflank's managed cloud, Northflank jobs, Northflank MCP, Nodus pricing, Nodus attempts and recovery, Nodus cloud accounts. Prices checked November 2, 2026.

Is Nodus cheaper than Northflank per GPU-hour?

On four of the five GPUs both list, yes. Northflank Cloud lists the NVIDIA L4 at $0.80, A100 80GB at $1.76, H100 at $2.74, H200 at $3.14 and B200 at $5.87 per GPU-hour. The Nodus pricing page lists the same GPUs at $0.43, $1.09, $2.60, $3.593 and $5.50.

The gap is widest on the smaller GPUs; on the H200, Northflank's listed rate is lower.

The gap matters most for single-GPU work. A LoRA fine-tune on one A100 80GB pays a 38% lower listed rate on Nodus. On the H100 the difference is about 5%, and on the H200 Northflank is cheaper. Northflank GPU plans also bundle a fixed amount of CPU and system RAM per GPU, so compare the full estimate, not only the GPU line.

What do four real training jobs cost on each?

We priced four common jobs at list rates. The math is GPUs x hours x the listed rate, nothing else.

Nodus wins three of four jobs at list price; the 8 x H200 run is $36.24 cheaper on Northflank.

The decision rule we take from this table: if your run fits on one or two GPUs, the listed rate is the biggest lever and Nodus usually wins clearly. At 8 GPUs the rates are close, so recovery and idle time decide the bill more than the hourly price does.

What happens if the machine fails halfway through training?

The difference is who writes the resume logic. Northflank's GPU guide says services and jobs are stateless and recommends saving checkpoints to persistent volumes. A job's Retry Limit sets how many attempts it gets, and a Time Limit ends it regardless of retries.

On Nodus, a Job writes its state to /nodus/state and Nodus saves and restores that directory. When a machine is reclaimed, Nodus asks for an urgent checkpoint and prepares a replacement at the same time. A machine that goes silent for 60 seconds on shared capacity is replaced. The run loses at most the work since its last committed checkpoint, as the attempts and recovery docs describe. We cover the failure modes in more depth in what happens when GPU capacity runs out.

Here is why that matters on the 8 x H100 job above. If the machine is lost at hour 7 and nothing was checkpointed, you redo 56 GPU-hours: $153.44 at Northflank's listed H100 rate, $145.60 at Nodus's. With a checkpoint every 30 minutes, you redo at most 4 GPU-hours. Northflank can get you there too, but you mount the volume and write the resume path yourself.

Can you use your own cloud account with either?

Yes, and Northflank goes further here. Its bring-your-own-cloud option runs GPUs in your AWS, Google Cloud, Azure, Oracle, CoreWeave or Nebius accounts, your provider bills the GPUs, and Northflank charges $1.00 per GB of VRAM per month. For one 8 x H100 80GB node that fee is 640 GB x $1.00 = $640 a month on top of your cloud bill.

Nodus can run work in a connected AWS account, GCP project or Azure subscription so it spends your cloud credits before Nodus capacity. That mode is in Beta, and the cloud accounts docs list exactly what each connection can do. If your committed spend sits with Oracle or CoreWeave, Northflank covers that today and Nodus does not.

Where is Northflank the better choice?

Pick Northflank when the GPU is one piece of an application. If your inference API, Postgres database, preview environments and CI/CD already live there, adding a GPU job to the same project is less work than adding a second platform.

Northflank is also the better pick in these cases:

• You want 8 x H200 and price is the deciding factor. Its listed $3.14 per GPU-hour is below Nodus.
• You need GPUs in Oracle, CoreWeave or Nebius accounts under one control plane, or timeslicing in your own cluster.
• You need an Enterprise contract with your VPC, on-premises or air-gapped deployment, SAML, SCIM and a custom SLA, all listed on its pricing page.
• You serve models with autoscaling services and want training and serving on the same platform.

Pick Nodus when the job is the product. That means a fine-tune or training run that should finish at a low listed rate, survive a lost machine without your own resume code, and stop at a dollar cap. For other ways to run training, see our Modal alternatives for training and Nodus vs RunPod comparisons.

How do you move a training job from Northflank to Nodus?

If your Northflank job runs a container with a training command, the move is one command. Install the CLI with pip install nodus-compute, run nodus login, then preview the estimate from the directory holding your script:

nodus run --name finetune --gpu H100:8 --image nodus/pytorch \ --max-cost 250 --timeout 12h --dry-run -- torchrun train.py 

Drop --dry-run to start it. To get recovery, write your model, optimizer and step to /nodus/state and load them on start, as the checkpoints guide shows. The flags are copied from the Nodus CLI reference; --max-cost 250 is a cap we picked to sit above the $208.00 estimate, not a quote.

FAQ

Does Northflank offer single H100 GPUs?

Not on Northflank Cloud as of the date above. Its pricing page lists the H100, H200 and B200 at 8 GPUs per container, while the L4, A100 and RTX PRO 6000 come in 1, 2, 4 or 8. Nodus lists the H100 per GPU, so a one-GPU job runs on one GPU.

Does Northflank Cloud use spot GPUs?

Its pricing page describes Northflank Cloud GPUs as on demand, billed by the second. Spot nodes are a bring-your-own-cloud feature, where workloads need a spot tag to land on them. On Nodus, --interruptible takes allow, prefer or never for each run, and recovery handles the reclaims.

Can I try Northflank for free?

The free Sandbox plan includes up to $20 a month of resources on Northflank Cloud, but GPU deployments need at least $50 of account credit. Nodus shows the estimated cost of a run with --dry-run before anything is charged.

Can a coding agent launch training on either platform?

Both have MCP servers. Northflank's sets up with nf agent setup --agent claude, and Nodus installs with nodus mcp install. On Nodus the agent sees the cost estimate before it runs, and a --max-cost cap bounds what it can spend, as covered in how to cap an agent's GPU spending.

Originally published on the Nodus blog.