CoreWeave sells GPU capacity at scale, and it mostly sells it eight cards at a time. An HGX H100 instance is $49.24 an hour, which works out to $6.16 per GPU. An 8x A100 is $21.60, or $2.70 per GPU. Those per-GPU rates are not unreasonable. The catch is the unit.
That is the single reason most teams go looking for an alternative. Not price, but the size of the smallest thing you can buy and the sales cycle attached to it.
This article covers nine alternatives, what each is actually built for, and where each one stops. Rates were read from each provider's own pricing page on 21 August 2026 unless noted.
| Platform | Minimum purchase | Self-serve | GPU selection | H100 |
|---|---|---|---|---|
| Runpod | 1 GPU | Yes | 24 models | $2.89/hr PCIe Secure |
| CoreWeave | 8 GPUs on HGX, 1 on GH200 | No | 10 configurations | $6.16/hr per GPU |
| Voltage Park | 1 GPU | Yes | H100 and Blackwell | Contact for pricing |
| Lambda | 1 GPU | Yes | 6 published | $3.99/hr SXM |
| Together AI | 1 GPU | Yes | 3 with rates on clusters | $3.99/hr HGX |
| GMI Cloud | Quote | No | 5 published | From $2.00/hr |
| CloudRift | 1 GPU | Yes | 12 models | $1.70/hr reserved |
| TensorDock | 1 GPU | Yes | 45 models | From $2.25/hr SXM5 |
| OVHcloud | 1 GPU | Yes | Several | See provider |
| Vultr | 1 GPU | Yes | Several | See provider |
Competitor rates were read from each provider's own pricing page on 21 August 2026 and change without notice, except Lambda, OVHcloud and Vultr. CoreWeave prices HGX instances in blocks of eight; the per-GPU figure is their own published inference single-GPU price. GMI Cloud and TensorDock publish "from" prices. Runpod rates pull live.
What to look for in a CoreWeave alternative
The criteria that matter shift depending on why you are leaving. Work out which of these is your reason before comparing rate cards.
- Minimum purchase. The most common reason to leave. CoreWeave's HGX instances come in blocks of eight. If you need one card, or four, that is the whole conversation.
- Self-serve access. Whether you can provision from a console or need a sales conversation first. This decides how fast you can start and how much friction sits between an idea and a running job.
- GPU selection. How many models, and how far down the range they go. A platform that starts at an H100 is expensive for a workload that fits on a 24GB card.
- Billing granularity. Per second, per minute, per hour or per month. On short or bursty jobs this changes the bill more than the headline rate does.
- Workload scope. Whether the platform covers development, inference and training, or only one of them. Replatforming later is a cost nobody puts on a pricing page.
- Interconnect. If you are doing genuine multi-node training, InfiniBand availability matters more than per-GPU price.
- Transfer costs. Egress fees can quietly dominate a bill on data-heavy work. CoreWeave charges none, and neither do several platforms here.
1. Runpod
Runpod is the broadest alternative for teams whose problem with CoreWeave is the minimum purchase. You can rent a single GPU from a console in under a minute, and the same account covers three different workload shapes.
Pods handle development and long-running jobs on your own container image or a template. Serverless handles inference that scales to zero between requests, with sub-200ms cold starts via FlashBoot. Clusters handle multi-node training, provisioned self-serve without a contract. All three use the same container images, so moving between them is a configuration change rather than a migration.
Key features
- Twenty-four GPU models, from an RTX A5000 through to a B300, across 31 global regions
- Per-second billing on Pods and Serverless with no minimum, and no charge during provisioning
- Standard Docker images, or Quick Deploy templates and the Hub for open-source models
- No ingress or egress fees
- SOC 2 Type II and independently verified for HIPAA and GDPR, with more than a million developers on the platform
Limitations
- No on-premises option. Runpod is a hosted service only.
- The lineup does not extend to rack-scale NVLink systems like the GB200 NVL72, which CoreWeave does offer.
- Community Cloud runs on third-party hardware, which is a genuine trade rather than a free discount.
Pricing
On Secure Cloud an H100 PCIe is $2.89/hr, an A100 PCIe is $1.59/hr and an RTX A6000 is $0.53/hr. Community Cloud is lower: $1.99/hr, $1.19/hr and $0.33/hr for the same three cards. Entry is $0.16/hr for an RTX A5000. Clusters are $4.31/hr per GPU for an H200 SXM and $1.79/hr for an A100 SXM.
Best for: teams that want to start with one GPU and still have somewhere to go when the workload grows.
2. Voltage Park
One thing first: Voltage Park owns TensorDock, which also appears on this list. They acquired it in March 2025, so treating the two as independent options would be a mistake. Voltage Park has also merged with Lightning AI, so expect a platform mid-integration.
On the product, Voltage Park is the closest direct answer to CoreWeave if what you want is large-scale capacity without the contract. They sell NVIDIA H100 and, more recently, Blackwell GPUs, from a single card up to 1,016 in a cluster, with 3200 Gbps InfiniBand on the higher tier.
Provisioning is self-serve and takes about fifteen minutes, with no minimum term on the on-demand tiers. There are no ingress, egress or support charges. Long-term reserve covers 32 to more than 8,000 GPUs on six-month-plus contracts. They also publish managed Kubernetes, bare-metal access, virtual machines, storage and observability.
Where it stops: their published hourly rates have been withdrawn. All three tiers now read contact for pricing, though their FAQ still quotes an H100 from $1.99 per hour without a contract. There is also no serverless tier for your own container.
Best for: large H100 or Blackwell clusters with real interconnect, if you are comfortable requesting a quote.
3. Lambda
An established provider aimed at research teams, with both on-demand instances and cluster products. On-demand rates are $3.99/hr for an H100 SXM, $2.79/hr for an A100 SXM 80GB, $1.99/hr for an A100 40GB, $6.69/hr for a B200 SXM6 and $0.79/hr for a Tesla V100.
1-Click Clusters run B200s from $9.86 per GPU-hour at 16 GPUs, dropping to $8.87 at 256 or more. Prices exclude sales tax, VAT and GST, which is worth noting because most of this list quotes tax-inclusive.
Where it stops: no serverless tier, and the published lineup is six models. These figures date from 18 August and have not been re-verified since.
Best for: research teams who want instances and clusters from one established vendor.
4. Together AI
Worth considering if the workload leaving CoreWeave is open-model inference or fine-tuning rather than raw capacity. Together AI sells self-serve GPU clusters alongside a per-token inference API and managed fine-tuning.
Cluster rates are $3.99 per GPU-hour for an HGX H100, $5.99 for an H200 and $8.19 for a B200, with a published reserved ladder bringing the H100 to $3.69 at 7 to 30 days, $3.45 at 31 to 90 and $3.19 at 91 to 180. GB200 NVL72, GB300 NVL72 and HGX B300 are contact-sales. Their separate Dedicated Inference product lists an HGX H100 at $5.49/hr.
Where it stops: they publish nothing below the H100, so small jobs have no economical home. They do offer a Sandbox product for development environments, priced per vCPU and per GiB of RAM, but that is not a GPU workstation with a persistent disk.
Best for: open-model serving and tuning at H100 scale.
5. GMI Cloud
The closest match to CoreWeave's enterprise positioning. GMI Cloud sells dedicated NVIDIA GPUs in its own data centers, with commitment-based pricing and Contact Sales as the primary route in.
Published floors are $2.00 per GPU-hour for an H100, $2.60 for an H200, $4.00 for a B200 in limited availability and $8.00 for a GB200. GB300 is pre-order. They also sell inference-as-a-service and a model library alongside raw GPUs.
Where it stops: these are "from" prices on an enterprise product with no published self-serve tier, so they are not directly comparable to a self-serve rate. If your reason for leaving CoreWeave was the sales cycle, this reproduces it. These figures date from 20 August.
Best for: enterprise buyers who want dedicated capacity and are comparing on commercial terms rather than console access.
6. CloudRift
A transparent rate card with on-demand and reserved tiers selectable in the console, billed per second with no minimum duration and no egress, ingress or API-call charges. Local storage is included in the hourly rate.
On-demand rates include an A100 SXM4 at $1.05/hr, RTX PRO 6000 at $1.29, RTX 5090 at $0.65, L40S at $0.63 and RTX 4090 at $0.39. Reserved tiers run one week at 5% off, one month at 10% and three months at 15%. H100 at $1.70 and H200 at $2.50 are reserved-only. They also run an OpenAI-compatible inference API billed per token, and an orchestration platform for enterprises running their own datacenters.
Read the specifications before comparing. Their H100 row lists 48GB of VRAM rather than the usual 80GB, their RTX 4090 and RTX PRO 6000 rows both list 96GB, and their FAQ notes that H100 and H200 are not orderable directly in the console.
Best for: mid-range cards at a low hourly rate with reserved discounts you can select yourself.
7. TensorDock
Owned by Voltage Park, which also appears on this list, following an acquisition in March 2025. TensorDock's own site does not mention it.
The product is a marketplace rather than a fleet. Independent hosts list their own hardware and set their own prices, which is how it reaches 45 GPU models across more than 100 locations in over 20 countries, and why the same card shows several different rates on one page.
Floors are $2.25 per hour for an H100 SXM5, $1.80 for an A100 SXM4, $0.35 for an RTX 4090 and $0.12 for entry-level consumer cards. KVM virtualization gives root access to a real VM with Windows support, which is unusual in this category. No ingress or egress fees.
The trade is consistency. Supply mixes Tier 3 and Tier 4 data centers with, in their own words, converted mining rigs for maximum price-to-performance. A floor price is not a rate card, and hardware varies by listing. Note also that funds are deducted continuously and servers are deleted, not stopped, when the balance reaches zero.
Best for: breadth of hardware and geography, if you can manage variance between hosts.
8. OVHcloud
A European provider offering bare metal servers, dedicated GPU instances and virtual private servers, with a large data centre footprint and strong data residency positioning. It suits teams that want direct hardware access and predictable monthly billing rather than hourly consumption.
Unlike CoreWeave, the GPU offering sits inside a broader general-purpose cloud, so you get storage, networking and managed services alongside it. That breadth is the point, and it is also the trade: this is not a purpose-built AI cloud.
We have not verified their current rates at source, so check their pricing page directly rather than relying on figures quoted elsewhere, including ours.
Best for: European data residency requirements and bare metal with monthly pricing.
9. Vultr
A general-purpose cloud with a GPU offering, aimed at developers and small teams who want quick provisioning through a simple console. Instances deploy in minutes, pricing is pay-as-you-go, and the platform covers compute, bare metal and storage alongside GPUs.
As with OVHcloud, the GPU product sits within a wider cloud rather than being the whole business, which suits teams that want one vendor for several things.
We have not verified their current rates at source. Check their pricing page directly.
Best for: teams already running other workloads on a general-purpose cloud who want GPUs in the same place.
Frequently asked questions
Why do teams look for CoreWeave alternatives?
Most often because of the minimum purchase. CoreWeave's HGX instances are priced in blocks of eight and there is no self-serve signup, so the smallest sensible commitment is large and the buying process runs through sales. Teams that need one or four GPUs, or that want to start today without a contract, tend to look elsewhere. Per GPU the rates are competitive; the unit is the constraint.
Can you rent a single GPU from CoreWeave?
In one case, yes. Their GH200 is a single-GPU instance at $6.50 per hour. They also publish a single-GPU inference price for each configuration, which their pricing page notes applies exclusively to CoreWeave inference platform customers. Everything else in the HGX range comes in blocks of eight, and access still runs through sales rather than a self-serve console.
Is CoreWeave expensive?
Per GPU, no. An HGX H100 instance at $49.24/hr works out to $6.16 per GPU, an 8x A100 at $21.60 is $2.70 per GPU, and spot capacity runs roughly 40 to 60% of on-demand depending on the card. Egress, ingress and transfer are free, their Kubernetes control plane and SUNK are free, and reserved capacity goes up to 60% off. The expense comes from the unit size.
Which CoreWeave alternative is cheapest?
It depends on the card. For H100s, Runpod at $1.99/hr on Community Cloud is among the lowest published rates here. For mid-range cards, CloudRift's A100 SXM4 at $1.05/hr and Runpod's A100 PCIe Community rate at $1.19/hr are competitive. TensorDock's $0.12/hr consumer floor is the lowest entry price, with the variance a marketplace implies.
Which alternatives let me rent a single GPU?
Runpod, Voltage Park, Lambda, Together AI, CloudRift, TensorDock, OVHcloud and Vultr all allow single-GPU rental. GMI Cloud routes through sales, as CoreWeave largely does.
Which offer InfiniBand for multi-node training?
Voltage Park publishes a 3200 Gbps InfiniBand tier. Lambda's 1-Click Clusters and Together AI's GPU Clusters both support multi-node training. Runpod Clusters scale to 64 GPUs and are provisioned self-serve.
Do these platforms charge egress fees?
CoreWeave, Runpod, Voltage Park, TensorDock and CloudRift all state no egress charges. Verify the others against their own pricing pages, since transfer costs can dominate a bill on data-heavy work.
Making the right choice
If you left because of the 8-GPU unit, almost everything here solves it. Runpod, Voltage Park, Lambda, CloudRift and TensorDock all rent single cards from a console.
If you left because of the sales cycle, avoid GMI Cloud, which reproduces it. Voltage Park is the strongest self-serve option at scale, though its rates are now quote-based.
If you are comparing Voltage Park against TensorDock, note that they share an owner.
If you need H100s at volume with real interconnect, Voltage Park's InfiniBand tier is the closest like-for-like.
If your workload does not need a data centre card at all, CoreWeave never had an answer for that. Runpod publishes ten cards at or under $1.09/hr.
If more than one of those is true, and past a certain size they usually all become true together, the case for Runpod is that Pods, Serverless and Clusters sit under one account. You can start with a single RTX A5000 and end on a multi-node cluster without changing vendors.
