AWS · Azure · GCP · NVIDIA T4, L4 and H100 · prices with dates and sources
Which region runs this GPU cheapest?
Every region is ranked by the on-demand hourly list price of the whole GPU instance — one NVIDIA T4, one L4, or an eight-card H100 node — with the exact SKU named beside each number. A region that does not offer the class is listed as unpriced, not quietly dropped: of 60 regions, 54 price a T4, 28 an L4 and 25 an H100.
54 regions · sorted by price
6 regions left out rather than priced at zero: no NC4as_T4_v3 in this region (eastasia, southafricanorth, uaenorth); no T4 capacity in this region (northamerica-northeast2, europe-west6, europe-north1).
T4 priced in 54 of 60 regions — the other 6 are listed under the results with the reason for each. A published price is not a guarantee of capacity either: GPU supply is allocated, and an H100 node in particular usually needs a quota increase or a reservation before a region will launch one for you.
Cheapest regionfor 1 NVIDIA T4 ong4dn.xlarge
US East (N. Virginia)
Ashburn, United States·us-east-1·6 availability zones
- $269/mo less(41%) than South America (São Paulo), the dearest
- 95 region heuristic
Based in North America, 6 AZs, SOC2 and PCI-DSS coverage.
Cheapest on each provider
- AWSUS East (N. Virginia)$0.5260$384/mo
- AzureEast US$0.5260$384/mo
- GCPUS Central (Iowa)$0.5400$394/mo
Every region on one price scale
The dearest region costs 70% more than the cheapest. Select a dot to find it in the table.
Instances: AWS g4dn.xlarge, Azure NC4as_T4_v3, GCP n1-standard-4 + 1x NVIDIA T4. Hourly price of the whole T4 instance, its 1 NVIDIA T4 card included.
54 regions ranked
| Region | Pin | |||
|---|---|---|---|---|
| 1 | AWSUS East (N. Virginia) us-east-1|g4dn.xlarge | $0.5260 | $384 | |
| 2 | AWSUS East (Ohio) us-east-2|g4dn.xlarge | $0.5260 | $384 | |
| 3 | AWSUS West (Oregon) us-west-2|g4dn.xlarge | $0.5260 | $384 | |
| 4 | AzureEast US eastus|NC4as_T4_v3 | $0.5260 | $384 | |
| 5 | AzureEast US 2 eastus2|NC4as_T4_v3 | $0.5260 | $384 | |
| 6 | AzureWest US 2 westus2|NC4as_T4_v3 | $0.5260 | $384 | |
| 7 | AzureWest US 3 westus3|NC4as_T4_v3 | $0.5260 | $384 | |
| 8 | GCPUS Central (Iowa) us-central1|n1-standard-4 + 1x NVIDIA T4 | $0.5400 | $394 | |
| 9 | GCPUS East (South Carolina) us-east1|n1-standard-4 + 1x NVIDIA T4 | $0.5400 | $394 | |
| 10 | GCPUS West (Oregon) us-west1|n1-standard-4 + 1x NVIDIA T4 | $0.5400 | $394 | |
| 11 | AWSEurope (Stockholm) eu-north-1|g4dn.xlarge | $0.5580 | $407 | |
| 12 | AzureSweden Central swedencentral|NC4as_T4_v3 | $0.5580 | $407 | |
| 13 | GCPBelgium europe-west1|n1-standard-4 + 1x NVIDIA T4 | $0.5590 | $408 | |
| 14 | GCPNetherlands europe-west4|n1-standard-4 + 1x NVIDIA T4 | $0.5592 | $408 | |
| 15 | GCPTaiwan asia-east1|n1-standard-4 + 1x NVIDIA T4 | $0.5700 | $416 | |
| 16 | AWSAsia Pacific (Mumbai) ap-south-1|g4dn.xlarge | $0.5790 | $423 | |
| 17 | AzureCentral India centralindia|NC4as_T4_v3 | $0.5790 | $423 | |
| 18 | GCPUS East (N. Virginia) us-east4|n1-standard-4 + 1x NVIDIA T4 | $0.5840 | $426 | |
| 19 | AWSCanada (Central) ca-central-1|g4dn.xlarge | $0.5840 | $426 | |
| 20 | AzureCanada Central canadacentral|NC4as_T4_v3 | $0.5840 | $426 | |
| 21 | AWSEurope (Ireland) eu-west-1|g4dn.xlarge | $0.5870 | $429 | |
| 22 | AzureNorth Europe northeurope|NC4as_T4_v3 | $0.5870 | $429 | |
| 23 | GCPTel Aviv me-west1|n1-standard-4 + 1x NVIDIA T4 | $0.5940 | $434 | |
| 24 | GCPMontréal northamerica-northeast1|n1-standard-4 + 1x NVIDIA T4 | $0.5942 | $434 | |
| 25 | GCPSingapore asia-southeast1|n1-standard-4 + 1x NVIDIA T4 | $0.6044 | $441 | |
| 26 | GCPTokyo asia-northeast1|n1-standard-4 + 1x NVIDIA T4 | $0.6138 | $448 | |
| 27 | AWSEurope (London) eu-west-2|g4dn.xlarge | $0.6150 | $449 | |
| 28 | AWSEurope (Paris) eu-west-3|g4dn.xlarge | $0.6150 | $449 | |
| 29 | AzureUK South uksouth|NC4as_T4_v3 | $0.6150 | $449 | |
| 30 | AzureFrance Central francecentral|NC4as_T4_v3 | $0.6150 | $449 | |
| 31 | AWSUS West (N. California) us-west-1|g4dn.xlarge | $0.6310 | $461 | |
| 32 | GCPMumbai asia-south1|n1-standard-4 + 1x NVIDIA T4 | $0.6382 | $466 | |
| 33 | GCPUS West (Los Angeles) us-west2|n1-standard-4 + 1x NVIDIA T4 | $0.6382 | $466 | |
| 34 | AWSMiddle East (Bahrain) me-south-1|g4dn.xlarge | $0.6450 | $471 | |
| 35 | AWSAsia Pacific (Seoul) ap-northeast-2|g4dn.xlarge | $0.6470 | $472 | |
| 36 | GCPLondon europe-west2|n1-standard-4 + 1x NVIDIA T4 | $0.6548 | $478 | |
| 37 | GCPFrankfurt europe-west3|n1-standard-4 + 1x NVIDIA T4 | $0.6548 | $478 | |
| 38 | AWSEurope (Frankfurt) eu-central-1|g4dn.xlarge | $0.6580 | $480 | |
| 39 | AzureWest Europe westeurope|NC4as_T4_v3 | $0.6580 | $480 | |
| 40 | AzureGermany West Central germanywestcentral|NC4as_T4_v3 | $0.6580 | $480 | |
| 41 | GCPJakarta asia-southeast2|n1-standard-4 + 1x NVIDIA T4 | $0.6588 | $481 | |
| 42 | AWSAsia Pacific (Sydney) ap-southeast-2|g4dn.xlarge | $0.6840 | $499 | |
| 43 | AzureAustralia East australiaeast|NC4as_T4_v3 | $0.6840 | $499 | |
| 44 | AWSAfrica (Cape Town) af-south-1|g4dn.xlarge | $0.6980 | $510 | |
| 45 | GCPHong Kong asia-east2|n1-standard-4 + 1x NVIDIA T4 | $0.7058 | $515 | |
| 46 | GCPSydney australia-southeast1|n1-standard-4 + 1x NVIDIA T4 | $0.7096 | $518 | |
| 47 | AWSAsia Pacific (Tokyo) ap-northeast-1|g4dn.xlarge | $0.7100 | $518 | |
| 48 | AzureJapan East japaneast|NC4as_T4_v3 | $0.7100 | $518 | |
| 49 | AWSAsia Pacific (Singapore) ap-southeast-1|g4dn.xlarge | $0.7360 | $537 | |
| 50 | AzureSoutheast Asia southeastasia|NC4as_T4_v3 | $0.7360 | $537 | |
| 51 | GCPSão Paulo southamerica-east1|n1-standard-4 + 1x NVIDIA T4 | $0.7816 | $571 | |
| 52 | AWSAsia Pacific (Hong Kong) ap-east-1|g4dn.xlarge | $0.8100 | $591 | |
| 53 | AWSSouth America (São Paulo) sa-east-1|g4dn.xlarge | $0.8940 | $653 | |
| 54 | AzureBrazil South brazilsouth|NC4as_T4_v3 | $0.8940 | $653 |
What is being compared
One instance per class, priced whole
Each class maps to one named instance on each provider that sells it. The number is that instance's on-demand list price for one hour, Linux, GPUs included — nothing is divided down to a per-card rate, because the card count differs by class and you cannot rent a fraction of one of these machines. The monthly figure is the same price for 730 hours.
The instances are matched on the card, not on the host: a T4 instance is 4 vCPU / 16 GB on AWS, 4 vCPU / 28 GB on Azure, 4 vCPU / 15 GB on GCP. The hourly prices are therefore comparable as GPU machines, not as general-purpose VMs. Where a provider sells nothing comparable at all the table says so rather than leaving a blank — Azure sells no L4 instance.
Reserved, spot, capacity-block and committed-use rates are not in the number, and neither is local SSD, networking or the storage your job reads.
| Class | AWS | Azure | GCP |
|---|---|---|---|
| T4 1 card, 16 GB | g4dn.xlarge1 × NVIDIA T4 GPU memory 16 GB 4 vCPU · 16 GB RAM | Standard_NC4as_T4_v31 × NVIDIA T4 GPU memory 16 GB 4 vCPU · 28 GB RAM | n1-standard-4 + 1x NVIDIA T41 × NVIDIA T4 GPU memory 16 GB 4 vCPU · 15 GB RAM VM and accelerator priced separately; the figure is their sum Google's public pricing page lists this price in 19 of the 22 regions here; the card runs in 23. |
| L4 1 card, 24 GB | g6.xlarge1 × NVIDIA L4 GPU memory 24 GB 4 vCPU · 16 GB RAM | Azure sells no L4 instance | g2-standard-41 × NVIDIA L4 GPU memory 24 GB 4 vCPU · 16 GB RAM Google's public pricing page lists this price in 15 of the 22 regions here; the card runs in 18. |
| H100 8 cards, 640 GB | p5.48xlarge8 × NVIDIA H100 GPU memory 640 GB 192 vCPU · 2048 GB RAM | Standard_ND96isr_H100_v58 × NVIDIA H100 GPU memory 640 GB 96 vCPU · 1900 GB RAM | a3-highgpu-8g8 × NVIDIA H100 GPU memory 640 GB 208 vCPU · 1872 GB RAM Google's public pricing page lists this price for us-central1 only, although the card runs in 13 regions. Elsewhere the price is unknown, not unavailable. |
What the hourly number includes
In the number
- The whole instance's on-demand Linux list price for one hour, its GPUs included.
- Every card in the instance: one for T4 and L4, eight for an H100 node.
- The host vCPU and RAM that come with the instance, whatever the provider pairs with the card.
- Internet egress at the region's first paid tier, once you enter a monthly volume.
- Your assumed service discount when you set the slider; internet egress stays at list price.
Not in the number
- Reserved, spot, capacity-block and committed-use rates, which are where most GPU capacity is actually bought.
- Block or object storage for datasets and checkpoints, local SSD beyond what the instance includes, and networking.
- Managed training and inference services (SageMaker, Azure ML, Vertex AI), which price the same hardware with their own markup.
- Support plans, taxes and negotiated enterprise pricing. Confirm with the provider before you commit.
- Capacity. A published price is not an offer of a machine: GPU supply is allocated per account and per region, and an H100 node in particular usually needs a quota increase or a reservation before it will launch. Price the region here, then check you can get the hardware there.