A hyperscaler sells GPUs as one line among dozens of products, and a neocloud sells GPU compute and little else. The price shows the difference. On 2026-11-12 the on-demand list price of one L40S, RTX PRO 6000 Blackwell or A100 80GB at AWS, Microsoft Azure or Google Cloud was 1.4 to 3.4 times QuantaCloud's price for the same GPU model (our calculation from each provider's own price list, below). The extra money buys what sits around the GPU: more regions, managed services, compliance programs and one account for everything. It does not buy a different GPU model.
What a hyperscaler is and what a neocloud is#
Neither word has a standards-body definition, so here is each one from a named research source.
| Term | Source | How the source defines it |
|---|---|---|
| Hyperscaler | Synergy Research Group, April 2026 | "the world's major cloud and internet service firms, including the largest operators in SaaS, IaaS, PaaS, search, social networking, e-commerce and gaming" |
| Hyperscaler | Uptime Institute, February 2025 | The "three tech giants" of cloud computing: Amazon Web Services, Microsoft Azure and Google Cloud |
| Neocloud | SemiAnalysis, October 2024 | "a new breed of cloud compute provider focused on offering GPU compute rental" |
| Neocloud | Synergy Research Group, October 2025 | "new or emerging specialized cloud computing platforms that provide high-performance, GPU-centric infrastructure, primarily to support artificial intelligence (AI) workloads" |
Scale is what makes a hyperscaler. Synergy counted 1,360 large data centers run by hyperscale companies at the end of 2025, 48% of the world's data center capacity. In everyday use the word means the three largest public clouds: in the third quarter of 2026 Amazon, Microsoft and Google held 26%, 19% and 16% of worldwide spending on cloud infrastructure services, by Synergy's count, which is 61% between them (our calculation: 26 + 19 + 16).
A neocloud is the opposite shape of company: a short list of products built around GPU compute, sold as virtual machines, bare-metal servers or whole clusters. What is a neocloud covers the definitions and a checklist for evaluating one, and GPU as a service covers how the product is priced.
AWS, Azure and Google Cloud GPU pricing on 2026-11-12#
For the GPUs QuantaCloud also rents, one GPU cost $1.86 to $6.98 an hour on demand at the three hyperscalers on 2026-11-12, depending on the GPU and the cloud. The table lists the smallest instance with one whole GPU that I found on each provider's price list: Linux, on demand, one US region each.
| GPU | Provider | Instance | Region | vCPUs and RAM | List price per hour |
|---|---|---|---|---|---|
| L40S, 48 GB | AWS | g6e.xlarge | US East (N. Virginia) | 4 vCPUs, 32 GiB | $1.861 |
| RTX PRO 6000 Blackwell, 96 GB | AWS | g7e.2xlarge | US East (N. Virginia) | 8 vCPUs, 64 GiB | $3.36312 |
| RTX PRO 6000 Blackwell, 96 GB | Microsoft Azure | Standard_NC144lds_xl_RTXPRO6000BSE_v6 | East US | 144 vCPUs, 264 GiB | $5.50 |
| RTX PRO 6000 Blackwell, 96 GB | Google Cloud | g4-standard-48 | us-central1 (Iowa) | 48 vCPUs, 180 GB | $4.49993 |
| A100, 80 GB | Microsoft Azure | Standard_NC24ads_A100_v4 | East US | 24 vCPUs, 220 GiB | $3.673 |
| A100, 80 GB | Google Cloud | a2-ultragpu-1g | us-central1 (Iowa) | 12 vCPUs, 170 GB | $5.0688 |
| H100, 80 GB | AWS | p5.4xlarge | US East (N. Virginia) | 16 vCPUs, 256 GiB | $6.88 |
| H100 NVL, 94 GB | Microsoft Azure | Standard_NC40ads_H100_v5 | East US | 40 vCPUs, 320 GiB | $6.98 |
QuantaCloud's price for the same GPU models, with each hyperscaler list price as a multiple of it (our calculation):
| GPU | QuantaCloud today | On 2026-11-12 | What the VM came with | Hyperscaler list price against it |
|---|---|---|---|---|
| L40S, 48 GB | $1.09/GPU-hr | $1.09 | 12 vCPUs, 72 GB of RAM, 625 GB of disk | AWS: 1.71 times ($1.861 / $1.09) |
| RTX PRO 6000 Blackwell, 96 GB | $2.16/GPU-hr | $2.39 | 16 vCPUs, 144 GB of RAM, 725 GB of disk | AWS: 1.41 times ($3.36312 / $2.39). Azure: 2.30 times ($5.50 / $2.39). Google Cloud: 1.88 times ($4.49993 / $2.39) |
| A100 80GB, PCIe | $1.48/GPU-hr | $1.48 | 16 vCPUs, 96 GB of RAM, 768 GB of disk | Azure: 2.48 times ($3.673 / $1.48). Google Cloud: 3.42 times ($5.0688 / $1.48) |
| H100, PCIe, 80 GB | $2.59/GPU-hr | Not in the catalog that day | Not listed | No same-day price to compare |
Prices checked 14 Nov 2026, 09:16 UTC
Read the two tables with three cautions. The first is shape: the instances are not the same size. AWS pairs its L40S with 4 vCPUs and 32 GiB of RAM where QuantaCloud's VM had 12 vCPUs and 72 GB, and Azure pairs its A100 with 24 vCPUs and 220 GiB where QuantaCloud's had 16 and 96 GB. If your job is bound by CPU or system memory, compare the whole shape.
The second is that these are list prices with no commitment. The February 2025 report from the definitions table notes that hyperscalers "also offer reserved instances and commitment-based enterprise discount schemes", which bring their unit costs closer to neocloud prices, and I have not priced those here. QuantaCloud has no spot instances and no published discount: a committed price comes as a written quote for reserved capacity.
The third is availability. A price on a list is not a GPU you can start. Google Cloud's documentation says its one-, two- and four-GPU H100 machine types must be created as Spot or Flex-start VMs, so it has no on-demand price for a single H100. QuantaCloud's own catalog listed no H100 at all on 2026-11-12.
The one cross-provider average I have found is older. In February 2025 that report put the average on-demand price of an 8-GPU H100 system at $98 an hour at the hyperscalers and $34 at neoclouds, a 66% saving. The hyperscaler side has moved since: on 2026-11-12 the same class of instance listed at $55.04 an hour at AWS (p5.48xlarge), $98.32 at Azure (Standard_ND96isr_H100_v5) and $88.49 at Google Cloud (a3-highgpu-8g), an average of $80.62 (our calculation: ($55.04 + $98.32 + $88.49) / 3).
What the hyperscaler price buys#
The premium pays for breadth, and the same February 2025 report gives the reason it persists: hyperscalers do not cut prices because "they do not have to", since they "have a captive, incumbent user base". It judged that most of the gap in GPU prices comes from "differences in gross margins rather than cost base". Here is what sits on each side, with the hyperscaler figures from their own pages on 2026-11-12.
| What you get | Hyperscalers | QuantaCloud |
|---|---|---|
| Regions | AWS: 39 regions and 124 Availability Zones. Azure: "80+" regions. Google Cloud: 43 regions and 130 zones | US only: us-east-1 in Virginia, and us-midwest-1, us-midwest-2 and us-midwest-4 |
| Catalog | "dozens of products, with millions of individual line items for sale", in the February 2025 report's words | GPU virtual machines on demand, and reserved capacity built to order |
| Services beside the GPU | Google Cloud launches every region with managed Kubernetes, object storage, persistent disks and a managed SQL database, by its own locations page | None of these. You run your own software on a VM, and there is no managed Kubernetes and no managed model endpoint |
| Storage that outlives an instance | Persistent disks and object storage, as products of their own | None: stopping an instance terminates it and deletes its disk |
| One account for everything | "a single set of APIs, management portal and invoice", in the same report's words | Single-user accounts with prepaid credit |
| Compliance programs | AWS lists 143 security standards and compliance certifications, with HIPAA and FedRAMP among them | Not SOC 2 or HIPAA certified |
| Support | AWS sells three paid plans above the Basic Support every customer gets, with published response times | Email and in-console tickets. No SLA |
The right-hand column is the honest list of what a neocloud leaves out, in QuantaCloud's case: no managed Kubernetes, no managed endpoints, no volumes or snapshots, and no region outside the US. If your application needs any of those next to the GPU, the hyperscaler price includes something you would otherwise build yourself.
Quotas and availability#
A hyperscaler makes you ask before you launch: GPU capacity sits behind quotas at all three. AWS counts On-Demand quotas in vCPUs, and its documentation lists a default of 0 for both "Running On-Demand G and VT instances" and "Running On-Demand P instances", so a new account starts with a quota request. Azure keeps two tiers of vCPU quota for each subscription in each region and adds a warning: "Even with sufficient quota, deployments can still fail if Azure doesn't have enough capacity for your requested VM size in the target region or zone". Google Cloud asks you to request quota for each GPU model in each region, plus a global GPU quota that applies across all regions.
QuantaCloud's launch rule is the balance: if a configuration is in the live catalog and your balance covers one hour of it, you can launch it, and most single-GPU VMs are running in about 3 minutes (median). The deploy guide walks through it. The limit is the catalog itself: it holds what is free at that moment, and on 2026-11-12 that was seven GPU models, with no H100 or H200 among them. When a plan depends on a GPU being there on a date, that is a reserved build, not an on-demand launch.
Egress#
All three hyperscalers charge for data leaving their network, and QuantaCloud does not. Inbound transfer is free at all four.
| Provider | Free each month | First paid tier, out to the internet | Where and how |
|---|---|---|---|
| AWS | 100 GB | $0.09 per GB, for the next 10 TB | US East (N. Virginia) |
| Microsoft Azure | 100 GB | $0.087 per GB, up to about 10 TB | East US, routed over Microsoft's network |
| Google Cloud | 1 GiB | $0.12 per GiB, up to 1 TiB | us-central1, Premium Tier, to North America |
| QuantaCloud | No allowance needed | No egress or ingress charges | US regions |
Moving 1 TB of checkpoints out in a month costs $81.00 at AWS, $78.30 at Azure and $111.64 at Google Cloud at those rates, if nothing else in the account has used the free allowance (our calculation: 900 GB x $0.09, 900 GB x $0.087, and 930.32 GiB x $0.12, since 1 TB is 931.32 GiB and the first GiB is free). That matters most for the work GPU clouds are used for, where model weights and checkpoints run to hundreds of gigabytes. On QuantaCloud you have to move them anyway, because the disk is deleted when you stop, and QuantaCloud charges nothing for the transfer.
Which one to use#
The rule I follow is to split by where the rest of the stack lives. A job that needs a GPU, a network port and your own software, such as fine-tuning a model or serving one behind an API, belongs wherever the GPU-hour costs least and starts soonest. A system that leans on a hyperscaler's databases, identity, private networking or compliance paperwork stays there, and you pay the GPU premium for not moving it.
The February 2025 report gives an example of running both: an organization trains a model on a GPU cloud and deploys it on a public cloud for inference. The cost to watch in that setup is egress on the hyperscaler side, every time a dataset moves out to the GPU cloud, so price the data movement before the GPUs.
Launch an L40SQuestions about neoclouds and hyperscalers#
What does hyperscaler mean?
A hyperscaler is one of the handful of companies that run data centers at the largest scale, in Synergy's words "the world's major cloud and internet service firms". For cloud buyers it means the three largest public clouds: AWS, Microsoft Azure and Google Cloud.
Which companies are neoclouds?
There are hundreds, and two sources track them by name. Synergy's October 2025 release named five companies as the ones generating substantial revenue and a long tail of 17 more, and described the group as "either relatively new start-up companies, or crypto miners that are transitioning to being providers of high-performance computing services". SemiAnalysis follows 323 providers and rated 77 of them in ClusterMAX 3.0, published on 2026-09-23. I do not rank other providers on this page.
Who are the top neoclouds?
The closest thing to a public ranking is ClusterMAX 3.0, which scores providers across 10 categories of criteria and gave a medal rating to only 19 neoclouds. The hyperscalers sit in the same table: Google Cloud in its Gold tier, Azure in Silver and AWS in Bronze. Note what it measures: in its authors' words, "we are ranking managed clusters", meaning Slurm and Kubernetes clusters that the provider runs for you. QuantaCloud's on-demand product is single GPU VMs, not managed clusters.
Are neoclouds cheaper than hyperscalers?
In every same-day comparison above, yes: on 2026-11-12 the hyperscaler list price was 1.4 to 3.4 times QuantaCloud's for the same GPU model. The February 2025 report found the same direction on average across several neoclouds for 8-GPU H100 systems. Commitments narrow the gap on the hyperscaler side, and egress, storage and quotas change the total, so compare the whole bill for your own job.
Why do hyperscalers charge more for the same GPU?
The February 2025 report's answer is margin, not cost. It found that most of the gap comes from "differences in gross margins rather than cost base", and that customers pay it because staying with one provider is simpler than adding a second.
Is the GPU itself different?
The GPU models are the same, and three details are worth checking. AWS's G7e, Azure's NC RTX PRO 6000 v6 series and Google Cloud's G4 use the RTX PRO 6000 Blackwell Server Edition. NVIDIA sells that GPU in three editions with the same 96 GB, and QuantaCloud's catalog lists only "RTX PRO 6000 Blackwell", so run nvidia-smi -q on the VM to see which edition you have, as the editions guide explains. Azure's A100 is the PCIe version, like the QuantaCloud A100 in the table. And Azure's single-GPU H100 instance uses the H100 NVL with 94 GB, where AWS lists an 80 GB H100.
Is QuantaCloud a neocloud?
Yes. QuantaCloud is a standalone GPU cloud that sells NVIDIA GPU virtual machines by the hour in US regions and builds dedicated servers and GPU clusters to order.
My rule: price the GPU-hours, the egress and the storage for one real month before you choose, because the hourly rate is only one of the three. If the job is mostly GPU time, start with the live GPU prices. If it needs a term, a fixed configuration or hardware the catalog does not list, send a capacity brief and the configuration, lead time and terms come back in writing.