Per GPU-hour, a Google Cloud on-demand H100 costs about four times the Nodus listed rate, and Google Spot VMs cost about two and a half times as much. Spot narrows the gap, but Compute Engine can reclaim those machines at any time and leaves checkpoint and restart logic to you. Nodus lists an NVIDIA H100 at $2.60/hr on its pricing page, saves your state when interruptible capacity is reclaimed, and can route work through a connected GCP project so you spend Google credits before Nodus capacity.
We build Nodus. Every Google Cloud fact below links to Google's own pricing page or docs, the arithmetic is shown, and one section covers where Google Cloud is the better choice.
How do Google Cloud and Nodus GPU prices compare?
Google Cloud prices accelerator-optimized VMs as a full machine: GPUs, vCPUs, memory and local SSD bundled together. The A3 High machine type gives you eight NVIDIA H100 GPUs, 208 vCPUs, 1,871 GB of memory and 6,000 GB of SSD. Nodus lists the GPU hour on its own, with CPU and memory priced separately on the same page.
| GPU | Google Cloud on-demand | Google Cloud Spot | Nodus listed rate |
|---|---|---|---|
| 8 x H100 SXM (A3 High) | $88.49/hr ($11.06/GPU) | $52.70/hr ($6.59/GPU) | $20.80/hr ($2.60/GPU) |
| 8 x H100 SXM (A3 Mega) | $93.40/hr ($11.68/GPU) | $45.60/hr ($5.70/GPU) | $20.80/hr ($2.60/GPU) |
| 8 x H200 SXM (A3 Ultra) | $84.81/hr ($10.60/GPU) | $50.87/hr ($6.36/GPU) | $28.74/hr ($3.59/GPU) |
| 8 x A100 40 GB (A2 Standard) | $29.39/hr ($3.67/GPU) | $16.75/hr ($2.09/GPU) | $7.23/hr ($0.904/GPU) |
Prices checked October 29, 2026. Google Cloud rates from the accelerator-optimized pricing page for us-central1 (Iowa). Nodus rates from the pricing page.
At list prices, Nodus is about 76% cheaper than Google Cloud on-demand and 61% cheaper than Spot per H100-hour.
What does the same training job cost on each?
We priced the same 8 x H100 run for 10 hours, the worked example we use across our comparison posts.
| Platform and capacity | Hourly rate | 80 GPU-hours |
|---|---|---|
| Google Cloud A3 High, on-demand | $88.49/hr | $884.90 |
| Google Cloud A3 High, Spot | $52.70/hr | $527.00 |
| Nodus | $20.80/hr | $208.00 |
Prices checked October 29, 2026. Google Cloud: accelerator-optimized pricing, us-central1. Nodus: pricing page.
The Nodus run costs less than a quarter of the GCP on-demand price and about 39% of the Spot price for the same GPU-hours. One caveat keeps this fair: Google's price includes 208 vCPUs and 1,871 GB of memory, while the Nodus figure is the GPU line only. Nodus lists CPU and memory separately (4 vCPU with 16 GiB is $0.1054/hr), so a job that needs a large host adds to the Nodus side.
The same 80 H100-hours cost $677 less on Nodus than on Google Cloud on-demand.
Google Cloud's Spot rate looks closer, but the Spot VM docs give a best-effort shutdown period of up to 30 seconds, and the run does not resume by itself. You write the checkpoint code, save to Cloud Storage and relaunch. On Nodus, an interruptible run whose machine is reclaimed gets a replacement with its /nodus/state directory restored, so you save to that path and Nodus handles the restart. That pattern is covered in How do you make spot GPU training survive preemption?.
Can you use both Google Cloud and Nodus?
Yes, and if you hold Google credits it is often the cheapest setup. Nodus can connect a GCP project read-only to show its instances and spend, and, as a separate Beta step, run your single-node Jobs in that project first so they spend your Google Cloud credits before Nodus capacity.
The read-only connection is an OAuth grant limited to read-only scopes (compute.readonly, bigquery.readonly and cloud-platform.read-only). Viewing is free and starts no machines. Running work is an explicit upgrade with a monthly credit limit. Nodus then creates its own network and a service account with roles that can only touch machines it named, and launches in us-central1, us-east1, us-east4 or us-west1.
When the project is full or out of credit, the default is to continue on Nodus capacity at Nodus rates. You can instead make Jobs wait for your approval or wait for your account. Once the owner setup script has finished, the CLI step is one line, copied from the docs:
nodus cloud connect gcp gcp-lab --credit-limit 500 --gcp-project lab-ml-prod
Each attempt on the job page shows whether it ran in your project or on Nodus capacity. Multi-node runs and Sandboxes use Nodus capacity, and the console shows the routing terms before you enable it. The full steps and permissions are in Connect your cloud accounts.
Is Google Cloud Spot stable enough for training?
Spot is stable enough only if your code checkpoints and restarts. Google's Spot VM overview says Compute Engine can preempt Spot VMs at any time, recommends them for fault-tolerant workloads, and gives a best-effort shutdown period of up to 30 seconds. Compute Engine does not manage your checkpoint.
For a 10-hour fine-tune, one lost run without a checkpoint is 10 hours of GPU time and $527 of Spot spend wasted. The safe pattern is to save model, optimizer and step every N minutes and again when you receive the termination signal. On Nodus, you write to /nodus/state and Nodus restores it on a replacement attempt. On GCP, you write to Cloud Storage and write the code that detects and reloads the checkpoint.
Spot discounts also do not combine with commitments. Google's pricing page says Spot prices are final: "Spot prices apply to all Spot VMs (and preemptible VMs), so Spot VMs (and preemptible VMs) cannot receive CUDs or SUDs."
Where is Google Cloud the better choice?
Google Cloud is the better choice in four situations, and we would tell a friend to stay in each of them.
You already have Google credits or commitments
Startup credits, enterprise agreements and committed-use discounts change the math. On Google's pricing page, a one-year resource commitment on A3 High is $61.38/hr and a three-year commitment is $38.86/hr, or $4.86 per GPU-hour. That is still above the Nodus rate but locked in for three years. Credits are also the case where using both wins: connect the project and Nodus spends them first.
Your data lives in Google Cloud Storage
Training next to your data avoids moving it. Sending datasets out of Google Cloud incurs internet egress charges (network pricing), which matter if you retrain often on large data.
You want TPUs
TPUs are only available on Google Cloud, and Nodus does not offer them. If your stack is built on JAX and TPUs, this comparison does not apply.
Compliance requires Google Cloud
Some teams must keep data and compute inside a specific cloud for contractual or regulatory reasons. Then the per-hour price is secondary.
If none of these apply, Nodus is the cheaper choice for GPU training and handles recovery when interruptible capacity is reclaimed. For comparisons with other providers, see Nodus vs Lambda and RunPod alternatives for AI training.
How do you move a training job from Google Cloud to Nodus?
For a simple PyTorch job, install the CLI, add a budget and run the same command:
pip install nodus-compute
nodus login
# 8 x H100, a $40 spending cap, detached
nodus run --gpu H100:8 --max-cost 40 -d -- torchrun train.py
The --gpu H100:8 form and the --max-cost cap come from the nodus run reference; add --interruptible to allow interruptible capacity.
Checkpoint code that saves to a local directory works on Nodus if you point it at /nodus/state. Nodus snapshots that directory periodically and restores it on a replacement attempt. The full pattern is in the checkpoint guide.
If you run Vertex AI custom training today, keep your training script and replace the Cloud Storage checkpoint path with /nodus/state.
FAQ
Does work in my connected GCP project show up on my Google bill?
Yes. Machines Nodus starts in your project are billed by Google to that project, against its credits, and Nodus tracks them against the monthly credit limit you set. That limit is spent at Nodus's list estimate, which can differ from Google's bill, so the docs advise setting it a little below the credit you hold.
Which regions does Nodus use in a connected GCP project?
Nodus launches in us-central1, us-east1, us-east4 and us-west1. If a GPU quota is used up, it skips that GPU model in that region for an hour and runs the work elsewhere.
What happens if my GCP Spot VM is preempted mid-training?
On Google Cloud alone, the VM shuts down after a shutdown period of up to 30 seconds. You restart it or launch a new one and reload your checkpoint from Cloud Storage yourself. On Nodus with interruptible capacity, the platform launches a replacement machine and restores your /nodus/state directory automatically.
Is Google Cloud's DWS Flex-start cheaper than Nodus?
DWS Flex-start lists an A3 High at $38.32/hr ($4.79/GPU-hr), lower than on-demand but still above the Nodus rate of $2.60/GPU-hr. Flex-start works through resize requests that wait for capacity rather than starting at once (Flex-start docs).
How do Nodus and Vertex AI compare for managed training?
Vertex AI keeps training inside Google Cloud, next to your data and IAM. Nodus restores /nodus/state on a replacement attempt for you, and can use your connected GCP project as its first place to run.
Originally published on the Nodus blog.
Top comments (0)