A neocloud is a cloud provider whose main product is rented GPU compute for AI work. It sells a handful of GPU products where a hyperscaler such as AWS, Azure or Google Cloud sells a broad catalogue of services. The short version of this page: rent single GPUs from a neocloud, use a peer-to-peer marketplace only for work that can be interrupted and holds nothing sensitive, and stay on a hyperscaler when your data, credits or compliance paperwork already live there.
The word was popularised by SemiAnalysis, whose "AI Neocloud Playbook and Anatomy" of 3 October 2024 defined it as "a new breed of cloud compute provider focused on offering GPU compute rental". SemiAnalysis has since said one of its analysts coined the term, but it is the only source for that, so "popularised" is the safe verb. By 1 July 2026 TechCrunch was using "neocloud" in a headline with no explanation.
The definitions in circulation do not match. Uptime Institute's February 2025 list of neoclouds includes Vast AI and established hosts such as Vultr and DigitalOcean. Synergy Research counts OpenAI as one. The SemiAnalysis playbook keeps brokers and marketplaces that own no GPUs in a separate group. That last split matters to anyone renting, so this page uses three classes, not two.
Three kinds of GPU cloud
| Hyperscaler | Neocloud | Peer-to-peer marketplace | |
|---|---|---|---|
| Who | AWS, Azure, Google Cloud, Oracle | CoreWeave, Lambda, Crusoe, Nebius and a long tail of smaller clouds | Vast.ai, and the Community Cloud layer inside Runpod |
| How you buy | A GPU instance type inside a general cloud account, next to storage, databases and networking | A GPU container, virtual machine or bare metal server, often with nothing else attached | A listing from an independent host who sets the price |
| Typical commitment | Pay as you go, or a one-year or three-year commitment for the large discounts | By the second or minute for small buyers; multi-year take-or-pay contracts for large capacity | None for on-demand; optional prepaid reservations |
| What breaks | New Google Cloud trial accounts cannot add GPUs at all; spot instances are reclaimed with two minutes' notice on AWS and 30 seconds on Azure (vendor docs, 21 September 2026); moving data out costs money | First-come capacity that can be gone when you want it; smaller product range; provider financial health | Host reliability and security vary; interruptible instances are stopped when outbid; bandwidth fees differ per host |
How neoclouds differ from hyperscalers
GPU-first product range. Uptime Institute noted in February 2025 that neoclouds "provide only a handful of product lines". Expect GPUs, storage and basic networking. Managed databases, queues and identity services are yours to bring.
Thin virtualisation or none. Latitude.sh rents "fully automated bare metal servers" with GPUs, per its pricing page on 21 September 2026. Vast.ai's documentation says its instances are "Docker containers that give you exclusive GPU access", and Runpod Pods are containers too.
Price. Uptime Institute compared the average on-demand hourly cost of an NVIDIA DGX H100-equivalent instance in February 2025 and found neoclouds 66% cheaper than hyperscalers. The same analysis said most of that gap is margin, not a lower cost base. Those figures are old. AWS cut on-demand prices for its H100 instances by 44% in June 2025, so the gap may have narrowed since.
Contracts. Hyperscaler discounts come from one-year or three-year commitments: AWS Savings Plans, Google committed use discounts and Azure Reserved VM Instances all work this way, per their pricing pages on 21 September 2026. Neoclouds split in two. Small buyers get short terms, such as Runpod's savings plans with "a 3-month or 6-month term upfront". Large buyers sign long ones. CoreWeave's annual report for 2025 describes its committed contracts as typically "take-or-pay", with initial periods that generally range "from one to six years". SemiAnalysis wrote on 2 April 2026 that most of the GPU rental market runs on contracts of six months or more. The hourly market on a comparison site is the small end of this business.
What you give up on a marketplace
A marketplace does not own the GPUs. Vast.ai's documentation says it "uses a marketplace model where hosts set their own prices", and that anyone can become a host by running its software. Three things stop being the provider's promise.
- Security. Vast.ai's own FAQ says "Individual hosts may have less formal security measures", while its Secure Cloud filter limits results to vetted data centres. Its host verification is "fully automated" and based on reliability and configuration, so "verified" describes machine health, not an audit.
- Continuity. On-demand instances on Vast.ai have a maximum duration set by the host, and interruptible instances can be "stopped by higher bids". Runpod's docs rate Community Cloud reliability as "Variable", and Runpod says it is no longer accepting new Community Cloud hosts.
- Predictable extras. Vast.ai's pricing docs say: "Data transfer costs vary by host and include both upload and download traffic." Runpod and Lambda both state they charge nothing for ingress or egress. Our data egress reference lists the fees by provider.
That makes marketplaces a poor fit for customer data and for jobs that cannot restart.
What each tracked provider charges right now
This table shows what each tracked provider charges right now for an H100 and an A100. It mixes neoclouds and marketplaces, so read the provider name before the price.
| Provider | H100 $/GPU-hr | A100 $/GPU-hr |
|---|---|---|
| QuantaCloud | $2.59 | $1.48 |
| Massed Compute | $2.73 | $1.35 |
| Lyceum | $2.79 | $1.59 |
| Ori | $2.90 | none in stock |
| RunPod | $3.19 | $1.59 |
| Lambda Labs | $3.29 | $1.99 |
| VERDA | $3.42 | none in stock |
| Paperspace | $5.95 | $3.09 |
| LeaderGPU | none in stock | $0.68 |
And the cheapest current listing for the newer parts, from any provider:
| GPU | Cheapest $/GPU-hr | Provider | Providers in stock |
|---|---|---|---|
| H100 | $2.59 | QuantaCloud | 8 |
| H200 | $3.43 | QuantaCloud | 4 |
| B200 | $3.75 | Packet.ai | 3 |
AWS, Azure and Google Cloud are not in these tables. This site tracks specialist GPU clouds and marketplaces. For day-by-day movement see the GPU price index.
Who the neoclouds are
The SemiAnalysis playbook named four "AI Neocloud Giants" in October 2024: Crusoe, Nebius, Lambda Labs and CoreWeave, which it called "by far the largest". Below them it placed a long tail of "a couple dozen clouds". One verified fact for each name you will meet most often:
- CoreWeave. Priced its IPO at $40.00 per share and began trading on Nasdaq on 28 March 2025. It reported a revenue backlog of about $104 billion as of 30 June 2026.
- Lambda. Founded in 2012. On 3 November 2025 it announced a multibillion-dollar, multi-year agreement with Microsoft to deploy "tens of thousands" of NVIDIA GPUs.
- Crusoe. Announced the initial close of a $3.9bn Series F at a $30.9bn post-money valuation on 17 September 2026. It claims 1 GW of operational capacity.
- Nebius. Amsterdam-based and Nasdaq-listed. On 16 March 2026 it signed a five-year agreement with Meta worth up to about $27bn, with deliveries from early 2027.
- Fluidstack. Its 20 July 2026 announcement says it raised an $830m Series A at a $7.5bn valuation in January.
- Voltage Park. Merged with Lightning AI in January 2026. The combined company states that it owns and operates more than 36,000 GPUs.
- Runpod. Said on 20 January 2026 that it had passed $120m in annual recurring revenue with more than one million developers.
- Hyperstack. The cloud brand of NexGen Cloud, which raised a $45m Series A on 8 April 2025 and positions itself on European sovereign AI.
Vultr and DigitalOcean show how blurry the edge is: both are established general hosts that Uptime Institute lists as neocloud examples. The full set of clouds we track is on the providers page.
How big the market is
Two analyst firms publish figures, and they measure different things.
- Synergy Research said on 2 April 2026 that neocloud revenue exceeded $25 billion for full-year 2025 and forecast that the market will approach $400 billion by 2031. Synergy includes OpenAI in its neocloud universe, which SemiAnalysis and Uptime do not.
- Gartner predicted that neocloud providers will capture 20% of a $267 billion AI cloud market by 2030, according to Intelligent CIO's report of 26 June 2026. We could not open Gartner's own release, so treat the wording as second-hand. It is a share of AI cloud spending, not of all cloud spending.
Do not add or compare these numbers. One more data point: SemiAnalysis wrote on 6 November 2025 that its highest-rated neoclouds had together booked nearly $400bn in remaining performance obligations since March 2025.
What the ratings tell you
ClusterMAX is a rating system for GPU clouds run by SemiAnalysis. Version 1.0 rated 26 providers on 26 March 2025. Version 2.0 rated 84 on 6 November 2025. It scores ten criteria: security, lifecycle, orchestration, storage, networking, reliability, monitoring, pricing, partnerships and availability. The tiers run Platinum, Gold, Silver and Bronze, the "last tier directly recommended", then Underperforming and Unavailable. Unavailable means "cannot verify yet".
The listing on 21 September 2026 put these tracked providers in the following tiers:
| Tier | Providers we track |
|---|---|
| Platinum | CoreWeave (the only member) |
| Gold | Crusoe |
| Silver | Lambda, Scaleway, Cirrascale, Vultr, Voltage Park |
| Bronze | Hyperstack, Runpod, Verda, DigitalOcean, Hot Aisle, Vast.ai, Latitude.sh, Denvr Dataworks |
| Underperforming | Salad, Massed Compute |
| Unavailable | Ori, listed as Radiant/Ori |
Among hyperscalers, Azure and Oracle are Gold and Google Cloud and AWS are Silver. Our record of the two lowest groups is partial, so a tracked provider missing from this table may be unrated or may sit in one of them.
What a tier tells you: how well a provider runs multi-node training clusters. Read the full review before signing for a cluster.
What a tier does not tell you: much about renting a single GPU for an afternoon. Networking, orchestration and partnerships are three of the ten criteria, and they barely touch a one-card job. A marketplace can rate low as a cluster provider and still suit a fault-tolerant batch on a single card. The rating also ages. Version 2.1 in April 2026 was "not a full re-test of all providers", version 3.0 testing is under way, and Voltage Park is still listed separately from Lightning AI despite the merger.
Provider risk
The hourly rate does not include the chance that your provider has a bad year. CoreWeave's annual report for 2025, filed with the SEC on 2 March 2026, is the most detailed view in our sources. Figures are as of 31 December 2025.
Customer concentration. The filing says: "We recognized an aggregate of approximately 67% of our revenue from our top customer, Microsoft, for the year ended December 31, 2025." The share was 62% in 2024 and 35% in 2023.
Debt. Total debt principal was $21,615m at the end of 2025, against $8,033m a year earlier. One facility that matures in March 2028 carries a 15% effective interest rate. The filing warns that this debt "could divert our cash flow from operations for debt payments". Nebius said on 8 September 2025 that it would fund a Microsoft build partly through "debt secured against the contract", so CoreWeave is not alone.
Depreciation. CoreWeave depreciates its technology equipment over six years, having changed its estimate "from five to six years" effective 1 January 2023. Depreciation and amortisation was $2,454m in 2025 against revenue of $5,131m, and the company reported a net loss of $1,167m. If GPUs lose their rental value faster than six years, those numbers get worse.
Thin margins and moving prices. If Uptime is right that the neocloud discount is mostly margin, there is little cushion. The SemiAnalysis index of one-year H100 contracts rose almost 40% between October 2025 and March 2026, and the firm reported on 2 April 2026 that on-demand capacity was sold out across GPU types at that time.
We do not know which providers will struggle, and few publish this much detail. For a small renter the defence is cheap: keep prepaid balances low, keep a copy of your data and checkpoints outside the provider, and build your environment from a container image so you can move in an hour.
How to choose
- You need one to eight GPUs for hours or days. Rent on demand from a neocloud. Start from the live table, then check storage and egress terms. A ClusterMAX tier should not decide this.
- Your job can be interrupted and holds no sensitive data. Use a marketplace, on interruptible pricing if your code checkpoints. On Vast.ai, turn on the Secure Cloud filter when you want data-centre hosts only.
- You serve production traffic or handle customer data. Use a neocloud that runs its own data centre capacity, or a hyperscaler if you need its compliance scope and surrounding services. Do not use individual marketplace hosts.
- You need 16 or more GPUs with fast networking for weeks. This is what ClusterMAX measures. Shortlist Silver and above, expect a contract, and compare quotes on our GPU cluster pricing page.
- Your data, credits or committed spend already sit in a hyperscaler. Stay until the credits run out. Google's startup programme page offered up to $350,000 in credits for AI-first startups as of 21 September 2026, and moving terabytes out has a cost of its own.
If you are unsure which one you are, you are the first.
Sources
Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.
All pages accessed 21 September 2026 unless a publication date is shown.
- SemiAnalysis, AI Neocloud Playbook and Anatomy (3 October 2024)
- SemiAnalysis, Most Neoclouds Suck at Security (30 August 2026)
- Uptime Institute, Neoclouds: a cost-effective AI infrastructure alternative (26 February 2025)
- Synergy Research, neocloud revenue and 2030 forecast (13 October 2025)
- Synergy Research, neocloud market forecast to approach $400B by 2031 (2 April 2026)
- Intelligent CIO, Gartner predicts neocloud vendors will capture 20% of the AI cloud market by 2030 (26 June 2026)
- TechCrunch, Neocloud Together AI raises $800M (1 July 2026)
- SemiAnalysis, ClusterMAX 1.0 (26 March 2025)
- SemiAnalysis, ClusterMAX 2.0 (6 November 2025)
- ClusterMAX 2.0 summary, criteria overview, version 2.1 notes and current provider listing
- SemiAnalysis, The Great GPU Shortage: rental capacity (2 April 2026)
- CoreWeave annual report on Form 10-K for 2025 and the SEC copy (filed 2 March 2026)
- CoreWeave, pricing of initial public offering (27 March 2025)
- CoreWeave, second quarter 2026 results (11 August 2026)
- Lambda, agreement with Microsoft (3 November 2025)
- Crusoe, Series F funding (17 September 2026)
- Nebius, agreement with Meta (16 March 2026)
- Nebius, agreement with Microsoft (8 September 2025)
- Fluidstack, Series A announcement (20 July 2026)
- Voltage Park, merger with Lightning AI (21 January 2026)
- Runpod, $120m ARR announcement (20 January 2026)
- NexGen Cloud, Series A funding (8 April 2025)
- AWS, up to 45 percent price reduction for NVIDIA GPU instances (5 June 2025)
- AWS Savings Plans pricing, Google Cloud committed use discounts and Azure Reserved VM Instances
- Google for Startups Cloud Program
- Latitude.sh pricing
- AWS: Spot Instance interruption notices and Azure Spot Virtual Machines (updated 24 June 2026)
- Google Cloud: free trial and free tier features (updated 18 September 2026)
- Runpod docs: Pods overview, choosing a Pod and Pod pricing
- Lambda docs: billing
- Vast.ai docs: instances overview, pricing, general FAQ, security FAQ, rental types, instance types and host verification