A neocloud is a cloud provider built specifically to rent GPU compute, rather than a general-purpose cloud that added GPUs later. Neoclouds sell access to AI accelerators and the network fabric between them. They do not sell managed databases, content delivery or the several hundred other services a hyperscaler carries.
The term is a contraction of new cloud. It came into general use through 2025 as a way to separate GPU-first providers from AWS, Azure and Google Cloud, and it now appears in vendor documentation from Cisco and Equinix as well as in analyst research. Before it caught on, the same companies were described as GPU clouds or GPU-as-a-Service providers.
What makes a neocloud different
The distinction is architectural, not just commercial. Three things separate a neocloud from a general-purpose cloud that happens to offer GPU instances.
Minimal virtualization. General-purpose clouds are built on hypervisors that abstract the hardware away from you. That abstraction costs performance, and on a large training job it costs measurably. Neoclouds tend toward bare-metal or lightly virtualized access, so the accelerator behaves the way its specification says it will.
A second network built for GPU-to-GPU traffic. This is the real dividing line. A neocloud runs two networks: a conventional front-end network for management and user access, and a separate high-performance back-end fabric that exists only so accelerators can talk to each other. Distributed training generates enormous east-west traffic between GPUs, and it is sensitive to packet loss in a way that ordinary web traffic is not. That fabric is typically InfiniBand or lossless Ethernet running RDMA or RoCE. If a provider does not have one, it is not really running AI clusters, whatever it says on the pricing page. We cover the trade-offs between those fabrics in our guide to NVLink, InfiniBand and Ethernet.
Hardware visibility. Neoclouds generally expose the topology: which GPUs sit on which node, how they are connected, what the interconnect can actually carry. You need that to tune a distributed job. Hyperscalers usually abstract it away, which is a reasonable decision for a web application and an unhelpful one for a training run.
Neocloud, hyperscaler and traditional cloud
| Traditional cloud | Hyperscaler | Neocloud | |
|---|---|---|---|
| Built around | General-purpose CPU workloads | Breadth of managed services | GPU compute |
| Virtualization | Heavy | Heavy | Minimal or bare metal |
| GPU interconnect | Usually none | Available on specific instance types | Core to the product |
| Service catalog | Broad | Very broad | Narrow by design |
| Provisioning | Minutes | Minutes, quota permitting | Seconds to minutes |
| Typical commitment | On demand or reserved | On demand, reserved or committed spend | On demand, often with no minimum |
| Regional footprint | Varies | Global, dozens of regions | Fewer regions |
The trade is straightforward. A hyperscaler gives you every service you might ever need and charges an abstraction penalty on the one you actually came for. A neocloud gives you the accelerator and the fabric, and expects you to bring the rest.
Who counts as a neocloud
There is no register, and the boundary is genuinely fuzzy at the edges. More useful than a roll-call is knowing that the label covers at least three different businesses, and that the differences between them matter more than the word they share.
Bare-metal and cluster specialists rent large, dedicated GPU clusters on multi-month or multi-year terms. Contract-led, aimed at model builders doing sustained training. Capacity is the product.
On-demand platforms rent by the hour or the second with no commitment, self-serve, aimed at developers who need a GPU now. Runpod, Vast.ai and Nebius sit here, among others.
Serverless inference platforms abstract the machine entirely and bill per request or per second of execution. Modal and Baseten are examples.
Several providers, including us, run more than one of these. So compare the model, not the label. A provider whose entire business is 12-month cluster contracts and one that rents a single GPU for 20 minutes are both called neoclouds, and almost nothing about the two decisions is the same. Our roundup of cloud GPU providers covers the individual providers in detail.
When a neocloud is the right choice
You are training, fine-tuning or serving models. This is what the architecture exists for.
You want the hardware to behave predictably. Minimal virtualization and a real fabric mean the second run resembles the first.
Your demand is uneven. Per-second billing on genuinely elastic capacity suits workloads that burst, which most experimentation does.
You do not need the surrounding platform. If your data already lives somewhere and you only need compute, you are paying a hyperscaler for services you will not use.
When it is not
Your application needs the wider platform. If your architecture depends on managed databases, queues, identity, CDN and analytics from one vendor, a neocloud replaces one component and leaves the rest to solve.
You have strict data residency requirements in a region we do not serve. Neoclouds run fewer regions than hyperscalers. Check the list before committing.
Your existing commitment already covers it. If you have negotiated committed spend with a hyperscaler, the marginal cost of GPU capacity inside that agreement may beat a list price elsewhere.
Your team wants managed services, not infrastructure. Neoclouds hand you lower-level constructs. That is the point, and it is not what every team wants.
Where Runpod fits
Runpod is an AI developer cloud. In the terms above, we run all three models rather than one: Pods for on-demand and interactive work, Serverless for inference that scales to zero between requests, and Clusters for multi-node training.
The specifics that tend to matter when people compare:
- Per-second billing across Pods, Serverless and Clusters, with no minimum
- No ingress or egress fees, which is often the line item that decides a real invoice
- 31 global regions
- Sub-200ms cold starts on Serverless, via FlashBoot
- SOC 2 Type II certified, and HIPAA and GDPR compliant. Reports, BAAs, and DPAs are available for security review through the Runpod Trust Center. Coverage can vary by region and deployment model, so check requirements for your workload. See runpod.io/legal/compliance for current certification details
Current rates start at {{gpu:rtx-4090:community}}/hr on Community Cloud for an RTX 4090, or {{gpu:rtx-4090}}/hr on Secure Cloud. An H100 SXM is {{gpu:h100-sxm}}/hr and an H200 is {{gpu:h200}}/hr on Secure Cloud. Full rates are on the pricing page.
If you are weighing renting against buying hardware outright, that is a different question and we have worked through it separately in buy, build or rent an AI server.
FAQ
What does neocloud mean?
Neocloud means new cloud. It describes a cloud provider built specifically to rent GPU compute for AI workloads, rather than a general-purpose cloud that added GPU instances to an existing platform. The term came into general use through 2025.
What is the difference between a neocloud and a hyperscaler?
A hyperscaler such as AWS, Azure or Google Cloud offers hundreds of managed services across a global footprint, with GPUs as one option among many. A neocloud offers GPU compute and the network fabric between accelerators, and little else. Neoclouds typically use minimal virtualization and expose hardware topology; hyperscalers abstract both away.
Are neoclouds cheaper than hyperscalers?
Often, but the comparison is rarely like for like. Check billing granularity, whether egress is charged, whether the quoted rate includes the vCPU and RAM you need, and whether storage is billed per GB or per GiB. A lower hourly rate with per-hour rounding and egress fees can cost more than a higher rate billed per second with no egress charge.
Which companies are neoclouds?
There is no formal definition or register, so the list depends on who is drawing it. Runpod is one. The group also includes on-demand platforms such as Vast.ai and Nebius, serverless inference platforms such as Modal and Baseten, and a set of contract-led cluster providers aimed at large-scale training. The business models differ enough that the label alone tells you very little.
Is a neocloud the same as GPU-as-a-Service?
Close, and the terms are often used interchangeably. GPU-as-a-Service describes the delivery model. Neocloud describes the kind of company, and usually implies the surrounding architecture as well: the dedicated back-end fabric, the minimal virtualization, the AI-first design. Every neocloud offers GPUaaS; not everything offering GPUaaS is built like a neocloud.
Do neoclouds support multi-node training?
The credible ones do, and it is the clearest test of whether a provider is one. Multi-node training needs a dedicated high-bandwidth fabric between nodes, typically InfiniBand or lossless Ethernet with RDMA. If a provider cannot describe its back-end network, it is renting GPUs rather than running AI clusters.
Get started
The fastest way to evaluate any neocloud is to run your own workload on it. Runpod bills by the second with no minimum and no quota request, so a real test costs very little. See current pricing, or deploy a Pod.
