Three models under one label
The useful way to sort this market is by what the provider owns, because that decides what it can promise you.
- Capacity builders (CoreWeave, Nebius, Crusoe, and increasingly Lambda) secure power and buildings, buy GPUs by the tens of thousands, and sell most of it on multi-year contracts. On-demand instances are the shop window; the business is the contract book.
- Developer platforms (Together AI, RunPod) put an API or a console in front of capacity they partly own and partly rent. You pay for convenience: serverless endpoints, per-second billing, hosted open models.
- Marketplaces (Vast.ai, and RunPod's Community Cloud tier) list machines operated by third parties. Prices are set by supply and demand, and so is reliability.
The mistake I see most often is comparing an H100 hour across those categories as if it were the same product. It is the same chip. It is not the same network, the same isolation, or the same answer when the node dies at 2 a.m.
The comparison
Prices are on-demand, per GPU-hour, in US dollars, as listed on each provider's own pricing page on 20 October 2026. They exclude tax, storage and egress, and committed pricing is lower everywhere.
| Provider | Model | GPUs offered | H100 on-demand | Who it's for | Status | Notable |
|---|---|---|---|---|---|---|
| CoreWeave | Bare metal / Kubernetes, own data centers | H100, H200, B200, GB200 NVL72, GB300 NVL72 | $49.24 per 8-GPU instance (≈$6.16/GPU) | Labs and hyperscalers buying multi-year clusters; large enterprises | Public (Nasdaq: CRWV, since March 2025) | ~$104B backlog and ~1.5 GW active power at 30 Jun 2026 |
| Nebius | Full-stack AI cloud, own data centers | H100, H200, B200, B300, GB200/GB300 NVL72 | $4.50 | Enterprises, AI labs; Europe-based buyers | Public (Nasdaq: NBIS) | Microsoft (~$17.4B) and Meta (up to ~$27B) contracts |
| Lambda | GPU cloud: on-demand instances and reserved clusters | H100, B200 (on-demand); larger clusters reserved | $3.99 (8x H100 SXM) | ML teams wanting simple self-serve instances; reserved training clusters | Private; IPO targeted 2027 (reported) | Reported $35B Anthropic commitment; Microsoft capacity deal |
| Crusoe | Vertically integrated: builds data centers and power, runs a cloud | H100, H200, B200, B300, GB200, AMD MI300X, MI355X | $4.50 | Large training customers; AMD-curious teams | Private; $30.9B post-money (Sep 2026) | Developer of the Stargate Abilene campus for Oracle and OpenAI |
| Together AI | Inference API plus GPU clusters | H100, H200, B200, GB200 NVL72 | $3.99 | Teams serving open models via API who may later want dedicated GPUs | Private; $8.3B valuation (Jul 2026) | Serverless, provisioned-throughput and dedicated inference |
| RunPod | Developer GPU cloud: pods and serverless | Consumer to data-center: RTX 4090/5090 up to H100, H200, B200 | $3.99 Secure / $2.69 Community | Startups, individual developers, bursty inference | Private; $1B valuation (Jun 2026) | Secure Cloud in vetted data centers; Community Cloud on third-party hosts |
| Vast.ai | Marketplace: hosts list machines, market sets price | 68+ GPU types, RTX 3060 up to B200 | Market-set; no list price | Price-sensitive, fault-tolerant work with checkpointing | Private | On-demand, interruptible (50%+ cheaper) and reserved tiers |
Pricing sources: CoreWeave, Nebius, Lambda, Crusoe, Together AI, RunPod, Vast.ai. For B200 the listed on-demand range was $6.69 (Lambda, 8-GPU instance) to $8.60 (CoreWeave, $68.80 per 8-GPU instance), with RunPod Community Cloud at $5.98.
The capacity builders
CoreWeave
The largest pure-play GPU cloud and the one the category is measured against. CoreWeave listed on Nasdaq in March 2025. For the second quarter of 2026 it reported revenue of about $2.6 billion, a revenue backlog of about $104 billion at 30 June, roughly 1.5 GW of active power and about 3.7 GW contracted, per its Q2 2026 results. It also reported a net loss of about $626 million, with interest expense the reason. Its attempt to buy Core Scientific for about $9 billion ended when Core Scientific shareholders voted it down in October 2025. CoreWeave is the right call when you need a very large, current-generation cluster with engineering support. It is the most expensive on-demand option here, which tells you on-demand is not what it is optimised to sell.
Nebius
Amsterdam-headquartered and Nasdaq-listed, Nebius rebuilt itself as an AI cloud and then signed two contracts that changed its scale: a five-year deal with Microsoft reported at about $17.4 billion, and in March 2026 a Meta agreement for $12 billion of dedicated Vera Rubin capacity plus up to $15 billion of additional compute, with delivery starting early 2027. Its listed on-demand prices undercut CoreWeave on every GPU it publishes. For a European buyer it is also the most obvious non-US option on this list.
Crusoe
Crusoe is as much a data center developer as a cloud: it built the Stargate campus in Abilene, Texas, for Oracle and OpenAI, and it puts power generation next to its sites. In September 2026 it raised a $3.9 billion Series F at a $30.9 billion valuation. Its cloud is the only one here that lists AMD Instinct MI300X on-demand ($3.45 per GPU-hour), which makes it a practical place to test whether your stack runs on AMD before you commit.
Lambda
Lambda started as the self-serve option for ML engineers and still has the simplest on-demand catalogue on this list. It has since moved toward the capacity builders: TechCrunch reported on 6 October 2026 that Lambda is raising up to $4 billion ahead of an IPO now planned for 2027, that its backlog rose from $15 billion to $50 billion between June and September, and that most of the increase is a $35 billion commitment from Anthropic. That is a strong customer and a concentration risk at the same time.
The developer platforms
Together AI
Together AI is the choice when you want to call an open model through an API today and keep the option of dedicated GPUs later. It sells serverless inference per token, provisioned throughput, dedicated inference endpoints billed per GPU-hour, and GPU clusters with preemptible, on-demand and reserved pricing. It raised an $800 million Series C at an $8.3 billion valuation in July 2026. The same vendor covering prototype and production inference is the main reason teams pick it.
RunPod
RunPod sells pods (a GPU container you control) and serverless endpoints, in two tiers: Secure Cloud in vetted data centers and Community Cloud on third-party hosts. The Community tier is where the low prices on this page come from. RunPod raised $100 million at a $1 billion valuation in June 2026. It is good for bursty inference, fine-tuning runs and anything a small team wants to start in minutes. Confidential data belongs on Secure Cloud, not Community.
Vast.ai
Vast.ai is a marketplace: hosts list machines, and "prices are set by the market, not by Vast", in the company's own words. It offers on-demand, interruptible instances it describes as 50% or more cheaper, and reserved terms of one to six months. Its pricing page did not display static prices when I checked, so there is no H100 figure in the table. Use it for fault-tolerant work that checkpoints often. Do not use it for anything where you have to explain to an auditor who operated the hardware.
How I would choose
- Experiments and fine-tuning runs under a week: RunPod or Lambda on-demand, Vast.ai interruptible if the job checkpoints.
- Serving open models in production: Together AI or a dedicated endpoint, unless the volume justifies reserved GPUs.
- A dedicated training cluster for months: get quotes from CoreWeave, Nebius, Crusoe and Lambda, and negotiate on delivery date and interconnect, not just price.
- Anything regulated: data-center tiers only, with attestations in hand before the data moves.
Whatever you choose, read the provider's financing as carefully as its price sheet. Neocloud expansion is largely debt-funded and concentrated in a handful of very large customers, and the IEA counts hyperscalers and neoclouds together at about $715 billion of planned 2026 capex. If a provider's next tranche of capacity depends on its next financing round, your delivery date does too. Background on where that capacity is being built is on the largest AI data centers.
Related
- AI infrastructure — the cluster hub: what AI data centers are and what limits them
- The largest AI data centers — where neocloud and hyperscaler capacity is being built
- Liquid cooling — why GB200-class capacity is scarce and site-dependent
- Local AI hardware — when owning a box beats renting an H100 hour
Frequently asked questions
What is a neocloud?
A neocloud is a cloud provider built specifically to rent out GPU and accelerator capacity for AI, as opposed to a general-purpose hyperscaler such as AWS, Azure or Google Cloud. The term covers companies that build and run their own data centers (CoreWeave, Nebius, Crusoe), companies that mostly lease space and focus on the developer experience (Lambda, RunPod, Together AI), and marketplaces that aggregate other people's hardware (Vast.ai).
CoreWeave vs Nebius: what is the difference?
Both are publicly listed, build large GPU clusters and sell long-term capacity to hyperscalers and labs. CoreWeave is larger: it reported a revenue backlog of about $104 billion at 30 June 2026 and about 1.5 GW of active power. Nebius, headquartered in Amsterdam, is smaller but growing quickly on the back of contracts with Microsoft (about $17.4 billion) and Meta (up to about $27 billion). On listed on-demand prices, Nebius was cheaper per H100 GPU-hour on 20 October 2026: $4.50 against about $6.16 for CoreWeave.
Which is the cheapest GPU cloud?
On listed on-demand prices for an H100, RunPod Community Cloud was the lowest of the providers on this page at $2.69 per GPU-hour on 20 October 2026, and Vast.ai's marketplace can go lower because hosts set their own prices. Both trade price for isolation and guarantees. Among providers selling dedicated data-center capacity, Lambda and Together AI listed H100 at $3.99. Committed and reserved pricing is lower everywhere and is negotiated, not listed.
Is RunPod or Vast.ai safe for company data?
It depends on the tier. RunPod's Secure Cloud runs in vetted data centers; its Community Cloud and Vast.ai's marketplace include machines operated by third parties. For regulated or confidential data, use the data-center tiers, check the provider's compliance attestations, and do not assume a marketplace host has the controls your auditors expect.
Is Lambda public?
No. As of October 2026 Lambda is private. TechCrunch reported on 6 October 2026 that it is raising up to $4 billion at a $14.5 billion pre-money valuation, possibly its last private round before a planned 2027 IPO, after an earlier plan to list in 2026 slipped.
Should we use a neocloud or a hyperscaler?
Use a hyperscaler when GPUs are one part of a larger workload that already lives there, or when you need its compliance footprint and managed services. Use a neocloud when GPU cost or availability dominates, for example a dedicated training cluster or a large inference fleet. Many teams do both: experiments on a neocloud, production next to their data.
Sources: provider pricing pages (retrieved 20 October 2026); CoreWeave Q2 2026 results via Investing.com; Nebius Form 6-K (16 March 2026); TechCrunch; SiliconANGLE; Data Center Dynamics; IEA. Valuations and funding are as reported; prices change frequently and are not quotes.