If you've been GPU-shopping in 2026, you've hit the term "neocloud" on a pricing page or a funding headline and probably guessed roughly right — but the word covers companies as different as a publicly-traded 250,000-GPU operator and a marketplace of individual hosts renting out their own rigs. That gap matters when you're deciding who to trust with a training run.
TL;DR: A neocloud is a cloud provider that specializes almost entirely in GPU compute for AI workloads, instead of offering the full stack of services a hyperscaler (AWS, Azure, Google Cloud) does. Neoclouds exist because GPU demand has outpaced hyperscaler supply since 2023, and they typically undercut hyperscaler on-demand GPU pricing — AWS's p5.48xlarge (8x H100) lists at $55.04/hr ($6.88 per GPU) versus $2.50-$3.99/hr per H100 at neoclouds like Hyperstack, Lambda, and RunPod, as of August 2026. What you trade for the lower price is a thinner platform: fewer managed services, less mature identity/networking tooling, and reliability that varies a lot between the leading operators and the marketplace tier.
What a neocloud actually is
A neocloud is a cloud infrastructure provider built specifically around renting out GPUs for AI training and inference, rather than the broad catalog of compute, storage, database, and managed services a hyperscaler sells. The term picked up currency around 2023-2024 as a wave of new (and some older, repositioned) companies raised money specifically to buy GPUs and rent them out, riding the gap between what NVIDIA could ship and what AWS, Azure, and Google Cloud could get their hands on.
The name is doing real work: "neo" as in new, "cloud" as in the delivery model — on-demand compute, no hardware ownership, pay by usage. What makes a neocloud a neocloud rather than just another cloud is the narrowness. A neocloud's entire business is GPUs, the networking between them, and enough software to make a cluster usable. It is not trying to also sell you a relational database, a CDN, or a serverless function runtime.
That focus shows up in three practical differences from a hyperscaler:
- GPU-first hardware allocation. A neocloud's capital goes almost entirely into GPUs, high-bandwidth networking (InfiniBand or similar), and power. A hyperscaler's GPU fleet is one line item in a much larger infrastructure budget that also covers everything else it sells.
- Faster access to new silicon. Neoclouds routinely have new NVIDIA or AMD chips available within weeks of launch, where hyperscaler availability often trails by months as the new hardware gets folded into a much larger provisioning and compliance pipeline.
- Simpler, often cheaper pricing. Fewer SKUs, more transparent per-GPU-hour rates, and — outside of a handful of the largest deals — less of the discount-negotiation layer that dominates hyperscaler enterprise pricing.
Why neoclouds exist
The short version: GPU demand broke the hyperscalers' normal supply model. NVIDIA's H100 and H200 generations sold out faster than hyperscalers could rack, power, and provision them at their usual pace — enterprise GPU lead times through 2026 have been reported as high as 52 weeks in some allocations. That gap created room for companies whose only job is buying GPUs and standing up clusters fast, without the org-wide capacity planning a hyperscaler runs before greenlighting a new SKU.
The capital story matters too. Building a GPU cluster is a multi-billion-dollar bet on power, networking, and NVIDIA allocation, and that scale of capex has made "neocloud" as much a financial category as a technical one — the sector's revenue is projected in the tens of billions in 2026 and growing fast, funded by a mix of venture capital, debt secured against the GPUs themselves, and direct investment from NVIDIA. CoreWeave, the largest public neocloud, reported $5.1 billion in 2025 revenue and a backlog north of $99 billion as of Q1 2026, funded in part by a March 2025 IPO.
Neocloud vs hyperscaler: the actual trade-off
The comparison people actually need isn't "neocloud good, hyperscaler bad" — it's which trade-offs you can live with for a given workload.
| Neocloud | Hyperscaler | |
|---|---|---|
| Primary product | GPU compute, narrowly | Full compute/storage/database/networking catalog |
| On-demand H100 pricing (Aug 2026) | ~$2.50-$3.99/GPU-hr (Hyperstack, RunPod) | ~$6.88/GPU-hr (AWS p5.48xlarge, Vantage) |
| New GPU generation availability | Often within weeks of launch | Frequently months behind, gated by internal rollout |
| Managed services beyond compute | Thin — usually just storage and basic networking | Extensive — IAM, VPC peering, managed DBs, compliance certs |
| Enterprise compliance/certifications | Growing, but inconsistent across providers | Mature — SOC 2, HIPAA, FedRAMP, decades of audits |
| Reliability track record | Ranges widely by operator, from datacenter-grade to hobbyist hardware | Consistent SLAs backed by large-scale ops history |
| Support model | Often lean, sometimes excellent (24/7 with sub-6-minute response at some) | Enterprise support tiers, but slower for novel GPU issues |
The honest read: if your workload is GPU compute and not much else — training a model, running batch inference, fine-tuning — a neocloud gets you there cheaper and faster. If your workload needs to sit inside an existing AWS VPC, talk to a managed database over a private link, and pass a specific compliance audit your hyperscaler already has certified, the neocloud's thinner platform becomes real integration work you have to build yourself.
Notable neoclouds and what each is actually known for
There's no single "neocloud tier" — the category spans public infrastructure companies with tens of thousands of GPUs down to marketplaces that resell an individual host's spare hardware. Here's what each of the more visible names actually specializes in, verified against their own sites as of August 2026.
| Provider | What it's known for |
|---|---|
| CoreWeave | Largest public neocloud by revenue ($5.1B in 2025); Kubernetes-native bare-metal platform; customers include OpenAI, Mistral AI, IBM; NASDAQ-listed (CRWV) since March 2025 |
| Lambda | Started as a deep-learning hardware maker, expanded into cloud; per-minute billing, no egress fees; Lambda Stack ships PyTorch/CUDA pre-installed |
| Crusoe | RDMA networking with topology-aware placement for large distributed training runs; NVIDIA (H100/H200/B200/GB200) and AMD (MI300X/MI355X) fleets; zero egress fees |
| Nebius | Amsterdam-headquartered, NASDAQ-listed (NBIS); full-stack AI platform from data prep through production deployment; roots in Yandex's infrastructure engineering |
| Together AI | Combines GPU clusters (on-demand and reserved) with a serverless inference API billed per-token — useful when you don't want to manage a cluster at all |
| RunPod | Both a GPU marketplace and a serverless inference platform; wide GPU catalog from RTX consumer cards up to B300 |
| Vast.ai | Peer-to-peer marketplace — 68+ GPU types across 40+ datacenters, prices set by host competition rather than a catalog; per-second billing |
| Hyperstack | Per-minute prepaid billing across NVIDIA's current lineup (A100 through B300); positions on price transparency |
| Modal | Serverless, container-based GPU execution billed per-second — not a persistent instance model at all, built for bursty inference and batch jobs |
Treat this table as a starting point, not an endorsement — verify current pricing and hardware availability on each provider's own page before committing, since both move often.
What you give up choosing a neocloud
The existing posts on this site win trust by naming what doesn't work, and the same applies here. Three real limitations, stated plainly:
Fewer managed services. A hyperscaler gives you a managed relational database, a message queue, a CDN, and a dozen other building blocks in the same account. Most neoclouds give you a GPU, some storage, and networking — everything else is on you or a third-party tool you bolt on.
Less mature IAM and networking. Fine-grained access control, VPC peering into your existing infrastructure, and audit logging are all things hyperscalers have spent over a decade hardening. Neocloud tooling here is improving but inconsistent between providers, and thinner than what a security team used to AWS IAM will expect on day one.
Variable reliability, especially at the marketplace tier. A publicly-traded operator running its own datacenters is a different reliability bet than an individual host on a peer-to-peer marketplace renting out a rig in a spare room. Both are "neoclouds" by the loose definition, and the term alone tells you nothing about which one you're getting.
How to choose
Match the provider tier to what actually breaks if it goes wrong:
- Production inference serving real users → a top-tier operator with a real support SLA (CoreWeave, Lambda, Crusoe, Nebius) or a managed serverless endpoint (Together AI, Modal, RunPod serverless), not a marketplace host you can't call.
- Training and fine-tuning runs you can checkpoint and resume → marketplace pricing (Vast.ai, RunPod, Hyperstack) is a legitimate way to cut cost, as long as you're checkpointing often enough that losing a host mid-run costs you minutes, not days.
- Anything that needs to integrate with an existing AWS/GCP/Azure account → budget real engineering time for the integration gap, or stay on the hyperscaler and eat the GPU price premium for that workload specifically.
The portability problem multi-neocloud shopping creates
Once you're comparing neoclouds on price — and the gaps above are real enough that most teams eventually do — you run into the same problem this site's other posts cover in detail: the cheapest GPU right now is rarely on the same provider as the cheapest GPU next month, and every provider switch has traditionally meant rebuilding your Python environment, re-downloading model weights, and losing whatever state you had on the box you just left.
That's the actual reason a live GPU Availability Index across providers is more useful than any single neocloud's pricing page, and it's the problem the marketplace is built to make visible — one place to see what's available and at what price across the neocloud landscape, rather than checking nine tabs.
Frequently asked questions
What is a neocloud in simple terms?
A neocloud is a cloud provider that specializes in renting out GPUs for AI workloads, instead of selling the full range of services a traditional cloud (AWS, Azure, Google Cloud) offers. Examples include CoreWeave, Lambda, Crusoe, and Nebius.
What's the difference between a neocloud and a hyperscaler?
A hyperscaler sells a broad catalog — compute, storage, databases, networking, compliance certifications — built up over more than a decade. A neocloud sells GPU compute specifically, usually at a lower per-GPU price and with faster access to new hardware, but with a thinner platform around it.
Are neoclouds cheaper than AWS or Azure for GPUs?
Generally yes, for on-demand GPU rental specifically. As of August 2026, AWS's p5.48xlarge (8x H100) lists at $55.04/hr, or $6.88 per GPU-hour, while neocloud on-demand H100 rates run roughly $2.50-$3.99 per GPU-hour at providers like Hyperstack and RunPod. The gap narrows for large negotiated enterprise deals on either side.
Is CoreWeave a neocloud?
Yes — CoreWeave is widely considered the largest public neocloud by revenue, reporting $5.1 billion in 2025 revenue and operating dozens of data centers built around NVIDIA GPU fleets, primarily for AI training and inference customers including OpenAI and Mistral AI.
Is a GPU marketplace like Vast.ai the same thing as a neocloud like CoreWeave?
Both get called neoclouds, but they're structurally different. CoreWeave owns or leases its own datacenters and infrastructure. Vast.ai is a peer-to-peer marketplace where independent hosts list their own hardware and prices are set by competition between them — which means reliability and support vary host to host in a way they don't on an operator-run platform.
Sources
- CoreWeave homepage and product pages, checked August 2026. Customer list, cluster goodput, bare-metal Kubernetes positioning.
- CoreWeave financials: $5.1B 2025 revenue, $99.4B backlog as of Q1 2026, via io-fund analysis of CoreWeave/Nebius/NVIDIA financing and public earnings coverage, checked August 2026.
- Lambda GPU Cloud pricing page, checked August 2026. Per-minute billing, GPU lineup, Lambda Stack.
- Crusoe Cloud page, checked August 2026. RDMA networking, GPU lineup (NVIDIA + AMD), zero egress fees.
- Nebius about page and pricing page, checked August 2026.
- Together AI pricing page, checked August 2026. GPU cluster and serverless-inference rates.
- RunPod pricing page, checked August 2026 (page dated July 27, 2026 at last update). On-demand GPU rates.
- Vast.ai pricing page, checked August 2026. Marketplace model, per-second billing.
- Hyperstack GPU pricing page, checked August 2026. Per-minute prepaid billing, GPU rates.
- Modal pricing page, checked August 2026. Per-second serverless GPU billing.
- AWS p5.48xlarge on-demand price ($55.04/hr, 8x H100, $6.88/GPU-hr) via Vantage AWS EC2 instance pricing, checked August 2026.
