The important H100 change
Google Cloud now documents a3-highgpu-1g, a3-highgpu-2g and a3-highgpu-4g alongside the original eight-GPU A3 High machine. The smaller shapes must currently use Spot or Dynamic Workload Scheduler Flex-start. They are not regular on-demand VMs.
This distinction matters when comparing providers. The old claim that Google Cloud requires an $88.49 eight-GPU VM for any H100 access is no longer correct. One H100 is listed at $4.79 per hour through Flex-start or at the current Spot rate of $6.64 per hour. Spot rates are variable and can change daily.
How to read Google Cloud GPU prices
A2, A3, A4 and G2 prices cover the complete accelerator-optimized machine, including the attached GPU, predefined vCPUs, memory and bundled local SSD where listed. A T4 attached to an N1 VM uses a different model: the GPU attachment rate is added to the VM price.
The lowest hourly number is not always the most useful option. Spot VMs can be reclaimed and have no SLA. Flex-start has separate scheduling behavior. Regular on-demand VMs suit uninterrupted work, but some newer GPU shapes are unavailable under that model.
Google Cloud GPU prices checked 3 September 2026
| Machine | GPU | Regular | Flex-start | Spot |
|---|---|---|---|---|
g2-standard-4 | 1 x L4, 24 GB | $0.71/h | Not offered | $0.42/h |
a2-highgpu-1g | 1 x A100, 40 GB | $3.67/h | $2.00/h | $2.12/h |
a2-ultragpu-1g | 1 x A100, 80 GB | $5.07/h | $2.40/h | $2.93/h |
a3-highgpu-1g | 1 x H100, 80 GB | Not offered | $4.79/h | $6.64/h |
a3-highgpu-8g | 8 x H100, 640 GB | $88.49/h | $38.32/h | $53.09/h |
a3-ultragpu-8g | 8 x H200, 1,128 GB | $84.81/h | $42.40/h | $49.01/h |
a4-highgpu-8g | 8 x B200, 1,440 GB | Not offered | $64.44/h | $39.63/h |
USD reference rates. Spot prices vary by region and can change daily. Verify your region and account in Google's calculator before starting a job.
Which provisioning model should you use?
- One H100: compare the current
a3-highgpu-1gSpot rate with Flex-start. Regular on-demand is not listed for this shape. - Uninterrupted A100 work: the one-GPU A2 machines have regular on-demand prices and are easier to plan than Spot.
- Checkpointed batch jobs: Spot can reduce cost, but the job must tolerate an unplanned stop and variable capacity.
- Large scheduled jobs: Flex-start gives A3, A4 and A2 workloads a separately priced route when start time is flexible.
- Existing Google Cloud stack: GKE, Vertex AI, Cloud Storage and BigQuery integration can outweigh a lower bare-GPU rate elsewhere.
Primary sources
Google Cloud GPU Pros & Cons
- One-, two-, four- and eight-H100 A3 High shapes
- Regular, Flex-start and Spot models shown side by side
- Integration with GKE, Vertex AI and other Google Cloud services
- L4 through B200 covered in one platform
- Smaller H100 shapes require Spot or Flex-start
- Spot pricing and capacity are variable
- Quota and regional availability can block a planned shape
- Total project cost can include storage and network charges
Verdict
Google Cloud is a practical GPU choice when the workload already depends on its data, orchestration or managed ML services. The new smaller A3 High shapes also remove the former eight-H100 entry barrier, although the one-, two- and four-GPU variants require Spot or Flex-start. For a simple interactive GPU with predictable immediate availability, compare a specialist provider before committing to the extra platform complexity.
Google Cloud GPU FAQ
Can you rent one H100 on Google Cloud?
Yes. The a3-highgpu-1g machine contains one 80 GB H100, 26 vCPUs, 234 GiB of memory and 750 GiB of local SSD. Google currently requires Spot or Dynamic Workload Scheduler Flex-start provisioning for the one-, two- and four-GPU A3 High shapes. Regular on-demand provisioning is listed for the eight-GPU shape.
Does the listed Google Cloud GPU price include the VM?
For accelerator-optimized A2, A3, A4 and G2 machines, the listed machine price includes the attached GPU, predefined vCPUs, memory and bundled local SSD where applicable. T4 GPUs attached to an N1 VM are different: the GPU attachment charge is added to the host VM cost.
Are Google Cloud Spot GPU prices fixed?
No. Google says Spot prices can change as often as once per day, and Spot VMs can be preempted whenever capacity is needed. Use the live price page before starting a job and use Spot only when the workload can checkpoint and resume.
When is Google Cloud a better choice than a specialist GPU cloud?
Google Cloud is most useful when the workload also needs Compute Engine, GKE, Vertex AI, Cloud Storage or BigQuery. A specialist GPU cloud is usually simpler when the requirement is only an interactive GPU pod or one inexpensive accelerator without the rest of the Google Cloud platform.