Catalog
GPU machine types as on-demand. TPUs are not a GPU hour; we will not pretend.
Provider review / verified 2026-09-05
Google CloudGPU review
GCP sells A3/A2-class GPU VMs and TPU pods inside Google Cloud. If you came for a cheap 4090, you are lost. If you came because your data is already in GCS and you have TPUs in the architecture diagram, keep reading.
Models listed5
From /GPU/hr$0.39
Regions in sample1
Business modelHyperscaler
Our reading
Who this provider is for
I treat GCP GPUs like AWS GPUs: a SKU inside a gravity well. The well is Google’s data/ML stack, Vertex, GCS, TPUs. If you need TPUs, stop comparing H100s and go do that honestly. If you need an H100 and you are not already on GCP, a neocloud will usually hurt less. Quota is the silent stock-out.
Rent it if Teams on GCP/Vertex, or teams who actually want TPUs, not a GPU cosplay.
Skip it if Marketplace shoppers. Anyone without GCP quota who needs a GPU today.
How we compare hoursStrengths
- TPUs exist here. They do not exist on Vast.ai. Stop forcing every model onto an H100 out of habit
- GPU VMs next to GCS and Vertex if that is already home
- Spot/preemptible with the usual caveats, labeled
- Global regions for companies already standardized on Google
Limits
- Quota and availability on the hot A3/A2 shapes can be the whole story
- On-demand GPU hours usually lose to specialist clouds
- TPU vs GPU is a different benchmark religion, do not hide it in one table cell
- Complexity tax same family as AWS
Same standard, every cloud
What we measured on Google Cloud
Four checks we apply to every provider. This is not a score, and it is not a paid ranking.
Price signal
Hyperscaler band. Compare to AWS/Azure and to a neocloud, separately from TPU economics.
Product shape
GCE + accelerator types + Vertex. Platform first.
Still verify
Quota, region, GPU vs TPU. The docs are optimistic. The quota page is honest.
- Model
- Hyperscaler
- Headquarters
- Mountain View, US
- Setup
- GCE GPU VMs, TPUs, Vertex
- On-demand
- Yes, quota permitting
- Spot
- Preemptible, labeled
- Unique
- TPUs
- Why people stay
- GCS / Vertex / IAM
- Compare with
- AWS, Azure, Nebius
Catalog listings
Google Cloud catalog sample
Published on-demand hours first. Spot and waitlists are labeled. Confirm the live rate before you provision.
If you need a TPU, say TPU
Half the GCP GPU content on the internet is people avoiding a TPU decision. If the model is JAX and the budget is a TPU pod, this page’s H100 table is a side quest. I respect TPUs. I will not launder them into a GPU leaderboard.
Quota is stock
Same sermon as AWS. A3 without quota is a brochure. Specialist clouds exist because Google and Amazon both play this game.
How I actually rent it
- 01
GPU or TPU
Pick. Then pick the shape. Then look at quota.
- 02
If GPU: compare a neocloud
Lambda or Nebius, same NVIDIA name, on-demand.
- 03
If you are not on GCP yet, ask why
Moving into Google for one GPU is a lifestyle choice.
FAQ
Google Cloud questions
Is Google Cloud cheaper than AWS for GPUs?+
Sometimes, by SKU and discount. Both lose to neoclouds on raw hours often enough that you should check. Credits distort everything.
Should I use TPUs instead of H100s?+
Only if your stack wants them. It is not a free upgrade. Different software, different queue, different failure mode.
GCP vs Nebius?+
GCP if you are already Google. Nebius if you wanted a GPU cloud without the Google estate.
Who should skip GCP GPUs?+
Anyone whose first GPU should have been a RunPod pod. Climb the ladder later.
Sources
Where these numbers come from
Prices and SKUs are from the GPUBeacon catalog sample. Ranking rules are on the methodology page. Outbound links may be affiliate. They never change the hour we display.
GPUs
Keep comparing
Other providers to compare
Same fields, different product. Compare the GPU hour before you lock a brand.
