AWS encodes series and size, Google encodes machine series and configuration, Azure encodes family, vCPUs and sometimes the accelerator, and Oracle uses GPU shape names, according to their naming documentation listed below. For AI, start with AWS P and G, Google A and G, Azure ND and NC, and Oracle BM.GPU. Uptime Institute wrote on 26 February 2025 that neoclouds, which focus on GPU-backed servers and virtual machines, often charge less than the hyperscalers; use the live tables below to compare today's offers.
Decode the machine before comparing its price. Write down the exact GPU, GPU count and memory per GPU. Then check the whole-machine bill. A low per-GPU figure is useful only if you can use the allocation you must rent.
These tables focus on H100, H200, A100, L4 and A10-class comparisons. Memory means nominal capacity per GPU, from the site's verified GPU specification reference, not host RAM or a promised usable VM allocation. Cloud names and GPU counts come from the publishers identified beside each table. Only the AWS table includes network figures; leaving them out of the other tables is not a claim of zero bandwidth.
AWS GPU instances: decode P and G before the size
AWS's naming guide, accessed 28 September 2026, puts the series, generation and optional capabilities before the period. The size follows it. AWS defines P as GPU accelerated and G as graphics intensive. Use the family mapping to identify the GPU; the generation digit alone does not name the NVIDIA product.
| AWS name field | Meaning in AWS's naming guide | Reading example |
|---|---|---|
p or g | GPU accelerated or graphics intensive | Start with the family |
| Generation digit | Generation within the series | p5 is a family identifier |
e | Extra GPU memory | Present in p5e |
n | Network and EBS optimized | Present in p5en |
d | Instance-store volumes | A capability suffix |
.48xlarge | Instance size | Not 48 GPUs |
.metal | Bare-metal size | Deployment form |
AWS's accelerated-computing reference, accessed 28 September 2026, gives the following GPU allocations and P-family network figures. AWS's G5 and G6 product pages, accessed the same day, supply their network figures. GPU memory below uses nominal GB from the verified specification reference, without converting AWS's GiB allocation labels.
| EC2 instance | GPU | GPU count | Nominal memory per GPU | AWS network figure |
|---|---|---|---|---|
p4d.24xlarge | A100 | 8 | 40 GB | 4 × 100 Gigabit |
p4de.24xlarge | A100 | 8 | 80 GB | 4 × 100 Gigabit |
p5.4xlarge | H100 | 1 | 80 GB | 100 Gigabit |
p5.48xlarge | H100 | 8 | 80 GB | 3200 Gigabit |
p5e.48xlarge | H200 | 8 | 141 GB | 3200 Gigabit |
p5en.48xlarge | H200 | 8 | 141 GB | 3200 Gigabit |
g5.xlarge | A10G | 1 | Confirm A10G allocation | Up to 10 Gbps |
g5.48xlarge | A10G | 8 | Confirm A10G allocation | 100 Gbps |
g6.xlarge | L4 | 1 | 24 GB | Up to 10 Gbps |
g6.12xlarge | L4 | 4 | 24 GB | 40 Gbps |
g6.48xlarge | L4 | 8 | 24 GB | 100 Gbps |
For an AWS H100 instance, AWS's reference makes the size distinction concrete: p5.4xlarge has one GPU and p5.48xlarge has eight. Do not choose the larger AWS P5 instance just because its normalized rate looks attractive. First decide whether your job needs the whole allocation.
Keep the suffix when copying EC2 GPU instances into a budget. AWS maps P5e and P5en to H200, while plain P5 maps to H100. For G5, request the exact A10G memory allocation before booking. The live A10 comparison is a family-level price reference, not proof that an A10 offer reproduces an A10G instance.
For a network label marked "up to," do not budget around sustained maximum throughput. AWS's reference, accessed 28 September 2026, identifies credit-based bursting on those sizes. Confirm the baseline for the exact instance. Keep the AWS provider page beside your quote so you can inspect the tracked offer rather than treating a family minimum as the price of every size.
Google Cloud GPU names: read the entire machine type
Google's machine-resource guide, accessed 28 September 2026, distinguishes family, series and machine type, and says a machine type specifies the instance's resource configuration. This matters because similarly shaped numeric suffixes describe different resources.
| Google name field or example | Documented interpretation |
|---|---|
| Family, series, machine type | Three separate terms; a machine type specifies the instance's resource configuration |
a3-highgpu-8g | Google's table maps this type to eight GPUs |
g2-standard-32 | Google's table maps this type to 32 vCPUs and one GPU |
-metal | Machine type without a hypervisor |
Google's accelerator-optimized machine documentation, dated 28 September 2026, gives these mappings. The rows deliberately range from one GPU to eight so you can distinguish the GPU from the amount you rent.
| Google machine type | GPU | GPU count | Nominal memory per GPU |
|---|---|---|---|
a2-highgpu-1g | A100 | 1 | 40 GB |
a2-highgpu-8g | A100 | 8 | 40 GB |
a2-ultragpu-1g | A100 | 1 | 80 GB |
a2-ultragpu-8g | A100 | 8 | 80 GB |
a3-highgpu-1g | H100 SXM | 1 | 80 GB |
a3-highgpu-8g | H100 SXM | 8 | 80 GB |
a3-megagpu-8g | H100 SXM | 8 | 80 GB |
a3-ultragpu-8g | H200 SXM | 8 | 141 GB |
g2-standard-4 | L4 | 1 | 24 GB |
g2-standard-24 | L4 | 2 | 24 GB |
g2-standard-32 | L4 | 1 | 24 GB |
g2-standard-96 | L4 | 8 | 24 GB |
For a GCP GPU quote, copy the middle part of the name too. Google's table assigns H100 to A3 High and Mega, but H200 to A3 Ultra. Treating all A3 machines as interchangeable loses the GPU identity before the price comparison even starts.
For Google Cloud GPU pricing, request the complete machine charge. Google's documentation dated 28 September 2026 says accelerator-optimized billing includes attached GPUs, predefined vCPUs, host memory and bundled Local SSD where applicable. Divide that quote by its GPU count before comparing it with a per-GPU listing. Then check what the competing listing includes. No Google list price is reproduced here.
Azure GPU VM names: the large number means vCPUs
Microsoft's naming convention, dated 1 December 2025, identifies the numeric field as vCPUs, not GPUs. It defines ND as AI training and inference optimized and NC as compute intensive. The accelerator field can name the hardware, but older names still require a lookup.
| Azure name field | Microsoft definition or reading |
|---|---|
ND, NC | AI training/inference optimized; compute intensive |
96 in ND96 | vCPU count |
a | AMD CPU |
d | Local temporary disks |
m | Memory-intensive variant |
r | RDMA/InfiniBand secondary network |
s | Premium SSD compatibility |
H100 | Accelerator identifier |
v5 | VM-series version |
Microsoft's size documentation dated 28 July 2026 supplies these allocations, except the ND H100 page, dated 25 September 2026. Do not infer a numeric network rate from the presence of r; it identifies a feature, not a speed.
| Azure size | GPU | GPU count | Nominal memory per GPU |
|---|---|---|---|
Standard_ND96asr_v4 | A100 | 8 | 40 GB |
Standard_ND96amsr_A100_v4 | A100 | 8 | 80 GB |
Standard_ND96isr_H100_v5 | H100 | 8 | 80 GB |
Standard_ND96isr_H200_v5 | H200 | 8 | 141 GB |
Standard_NC24ads_A100_v4 | A100 PCIe | 1 | 80 GB |
Standard_NC48ads_A100_v4 | A100 PCIe | 2 | 80 GB |
Standard_NC96ads_A100_v4 | A100 PCIe | 4 | 80 GB |
Standard_NC40ads_H100_v5 | H100 NVL | 1 | 94 GB |
Standard_NC80adis_H100_v5 | H100 NVL | 2 | 94 GB |
An H100 label is not enough to match two offers. Microsoft's ND and NC rows above specify different H100 configurations. Use the NVLink, PCIe and SXM guide when checking whether a cheaper listing matches the topology your workload requires.
For Azure GPU pricing, obtain a quote for the exact size, region and billing term. Keep it separate from any account discount until both are visible in your comparison. This article gives no numeric Azure rate and makes no price ranking between Azure and Google.
Oracle BM.GPU names: distinguish the shape from the bill
Oracle's shape reference, accessed 28 September 2026, separates bare-metal and VM GPU shapes. Its explicit mapping makes BM.GPU.H100.8 easy to decode, but do not generalize that example into a rule for every trailing number in OCI.
| Oracle name field | Meaning in Oracle's shape reference |
|---|---|
BM.GPU | Bare-metal GPU shape category |
VM.GPU | Virtual-machine GPU shape category |
H100 in BM.GPU.H100.8 | GPU model |
.8 in BM.GPU.H100.8 | Eight GPUs in this documented shape |
Oracle's same reference gives the allocations below. Use the full shape name when asking for a quote, including the A100 version marker.
| Oracle shape | GPU | GPU count | Nominal memory per GPU |
|---|---|---|---|
BM.GPU.H100.8 | H100 | 8 | 80 GB |
BM.GPU.H200.8 | H200 | 8 | 141 GB |
BM.GPU.A100-v2.8 | A100 | 8 | 80 GB |
VM.GPU.A10.1 | A10 | 1 | 24 GB |
VM.GPU.A10.2 | A10 | 2 | 24 GB |
BM.GPU.A10.4 | A10 | 4 | 24 GB |
Oracle's global price list dated 10 September 2026, observed 28 September 2026 with country USA and currency USD selected, publishes these pay-as-you-go compute rates. The geographic scope is the USA-selected global schedule, not a quote for a named OCI region.
| Oracle shape | Published USD per GPU-hour | Date and geographic scope |
|---|---|---|
BM.GPU.H100.8 | USD 10.00 | 10 September 2026; USA-selected global list |
BM.GPU.H200.8 | USD 10.00 | 10 September 2026; USA-selected global list |
VM.GPU.A10.1, VM.GPU.A10.2, BM.GPU.A10.4 | USD 2.00 | 10 September 2026; USA-selected global list |
Oracle's price list, accessed 28 September 2026, defines the whole-server hourly charge as the listed GPU rate multiplied by the server's GPU count. Compare the published per-GPU rate with the matching live row below, then price the allocation you actually need. These dated list rates do not confirm capacity in your chosen region.
Compare live AWS GPU pricing with the same GPU elsewhere
Read across the AWS row, then find that GPU in the cheapest-offer table. Both express prices per GPU. The difference is your starting point for comparing suppliers, not a guaranteed saving for a particular instance or job.
| Provider | H100 $/GPU-hr | H200 $/GPU-hr | A100 $/GPU-hr | L4 $/GPU-hr | A10 $/GPU-hr |
|---|---|---|---|---|---|
| AWS | none in stock | none in stock | none in stock | none in stock | none in stock |
| GPU | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|
| H100 | $2.50 | Hyperstack | 9 |
| H200 | $3.43 | QuantaCloud | 6 |
| A100 | $0.68 | LeaderGPU | 11 |
| L4 | $0.49 | RunPod | 2 |
| A10 | $0.37 | LeaderGPU | 2 |
Build a short quote worksheet with the exact machine name, region, GPU variant, allocation, billing term and expected runtime. Keep the per-GPU comparison beside the whole-job estimate. If a listing leaves a field unclear, resolve it before selecting the winner. This also gives your team a repeatable way to revisit the decision when either the workload or the offer changes.
Match memory variant, GPU count, rental term and included host resources before using the difference. For H100, inspect the variants directly:
| Variant | VRAM | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|---|
| H100 PCIe | 80 GB | $2.50 | Hyperstack | 6 |
| H100 SXM5 | 80 GB | $2.90 | Ori | 4 |
| H100 NVL | 94 GB | $3.11 | Massed Compute | 3 |
If you have an AWS contract quote, compare that quote too. A public offer does not tell you the value of your credits or negotiated terms. Use the on-demand, spot and reserved pricing guide to keep different commitments out of the same comparison column.
Check provisioning before choosing a supplier
AWS's Capacity Blocks documentation, accessed 21 September 2026, describes reserving GPU instances for a future date and says cancellations are not allowed. Confirm the start and end of the block against your job plan before committing.
Google's documentation dated 28 September 2026 requires Spot or Flex-start provisioning for a3-highgpu-1g, a3-highgpu-2g and a3-highgpu-4g. The same documentation limits accelerator types to particular regions and zones. A machine name therefore does not establish that your preferred provisioning mode exists in your region.
Microsoft's RTX PRO 6000 BSE v6 overview dated 1 September 2026 reports general availability in West US 2 and Southeast Asia. Treat that as a dated regional announcement, not evidence of spare capacity today. For every cloud, make region, allocation and start time part of the quote request.
Keep the cloud only when its total cost wins
Stay inside your hyperscaler when usable credits, data gravity, compliance requirements or existing contracts outweigh the compute-price difference. Put a value on each reason. Include transfer and storage in the calculation using the data-egress reference. Count the engineering work required to move the job and operate it elsewhere.
Rent the same GPU elsewhere when the workload is portable, the offer meets your requirements and the complete job costs less. The neocloud guide explains that provider category. If your proposed move also changes the accelerator, use the AWS Trainium versus GPU comparison; that is a separate decision from changing who rents you the GPU.
Choose the smallest matching allocation, confirm its terms, and compare total job cost. Stay for a concrete financial or operational advantage. Otherwise, rent the matching GPU from the cheaper qualified offer.
Sources
- AWS instance naming, accessed 28 September 2026.
- AWS accelerated-computing instance reference, accessed 28 September 2026.
- AWS G5, accessed 28 September 2026.
- AWS G6, accessed 28 September 2026.
- Google machine-resource guide, accessed 28 September 2026.
- Google accelerator-optimized machines, 28 September 2026.
- Microsoft VM naming conventions, 1 December 2025.
- Microsoft ND A100 v4, 28 July 2026.
- Microsoft NDm A100 v4, 28 July 2026.
- Microsoft ND H100 v5, 25 September 2026.
- Microsoft ND H200 v5, 28 July 2026.
- Microsoft NC A100 v4, 28 July 2026.
- Microsoft NCads H100 v5, 28 July 2026.
- Oracle compute shapes, accessed 28 September 2026.
- Oracle global price list, USA-selected USD schedule, 10 September 2026; observed 28 September 2026.
- Oracle cloud price list and billing unit, accessed 28 September 2026.
- AWS Capacity Blocks, accessed 21 September 2026.
- Microsoft RTX PRO 6000 BSE v6 overview, 1 September 2026.
- Uptime Institute: Neoclouds as an AI infrastructure alternative, 26 February 2025.