Voltage Park sells NVIDIA GPU capacity at scale, from a single card up to 1,016 in one cluster, with 3200 Gbps InfiniBand on the higher tier and up to 8,000 GPUs on long-term reserve.
Runpod publishes 24 GPU models at rates you can read on a page, across Pods for development, Serverless for inference that scales to zero, and Instant Clusters for multi-node training.
The comparison used to be about hourly rates on a single card. It is not any more, because Voltage Park no longer publishes them.
What changed at Voltage Park
They have merged with Lightning AI. Their site carries the announcement at the top of every page. Any evaluation of them should account for a platform in the middle of an integration rather than a settled product.
The catalog is no longer H100 only. Blackwell GPUs now sit alongside H100 in their navigation. If you looked at them earlier in the year and ruled them out on hardware age, that reason has weakened.
The published rates are gone. All three tiers now read "Contact for pricing." Their FAQ still quotes $1.99 per hour for a single H100 with no contract, so that figure has not necessarily changed, but it is no longer presented on the pricing table and the InfiniBand rate is not stated anywhere.
That last point is the practical one. A rate you have to request is a different purchasing experience from a rate you can read, and it makes side-by-side comparison impossible without a sales conversation.
What Voltage Park is built for
Large clusters with real interconnect. The InfiniBand tier runs from eight GPUs to 1,016 at 3200 Gbps, and long-term reserve goes from 32 to more than 8,000 GPUs on six-month-plus contracts. For distributed training where GPU-to-GPU bandwidth is the bottleneck, that is the point of the platform.
Self-serve at scale. Getting a hundred GPUs without a procurement cycle is unusual. Voltage Park states self-serve provisioning in under 15 minutes with no minimum terms on the on-demand tiers.
Managed services around the hardware. They publish managed Kubernetes, bare-metal access, virtual machines, storage and observability, plus a Studio layer with hosted models and prebuilt blueprints. This is a broader platform than a bare capacity vendor.
No hidden ingress, egress or support costs, which they state directly.
Where Runpod fits
Published pricing. Every Runpod rate is on the pricing page, for every card, on both Community and Secure tiers. You can build a cost model without talking to anyone.
A range that starts far lower. Voltage Park's hardware is data-center class throughout. A workload that fits comfortably on a 24GB or 48GB card has no economical home there. Runpod publishes ten models at or under {{gpu:l40s}}/hr, starting at {{gpu:rtx-a5000:community}}/hr for an RTX A5000 on Community Cloud.
Serving your own container with scale to zero. Runpod Serverless runs your image and stops billing between requests, with sub-200ms cold starts via FlashBoot. Voltage Park's hosted models cover their catalog rather than your weights.
One account across the project. Development on a Pod, serving on Serverless, training on an Instant Cluster, same container images throughout, per-second billing, 31 global regions, SOC 2 Type II.
Which one should you choose
Choose Voltage Park if you need many GPUs with InfiniBand for distributed training, if you want managed Kubernetes over that hardware, and if you are comfortable requesting a quote rather than reading a rate. At cluster scale they are a serious option and their self-serve provisioning at that size is genuinely unusual.
Choose Runpod if you want to know the price before you talk to anyone, if your work spans more than one card class, if you need to prototype on something cheaper than a data-center GPU, or if you are serving your own model and want it to scale to zero.
The honest split: Voltage Park is built for scale and is expanding fast, but it is mid-merger and has moved its pricing behind a contact form. Runpod is the AI developer cloud, with published rates and a range that covers the small end of the work as well as the large.
Frequently asked questions
How much does Voltage Park cost?
Their pricing page now shows "Contact for pricing" on all three tiers. Their FAQ still states that a single H100 can be rented from $1.99 per hour with no contract, but no InfiniBand or reserve rate is published. Runpod publishes every rate on its pricing page, including {{gpu:h100-pcie}}/hr for an H100 PCIe on Secure Cloud.
What GPUs does Voltage Park offer?
NVIDIA H100 and Blackwell GPUs. Runpod publishes 24 models from an RTX A5000 through to a B300, including H200, B200 and AMD MI300X.
Has Voltage Park been acquired?
They have announced a merger with Lightning AI, described on their site as creating a single place to build and ship AI. Factor an in-progress integration into any evaluation.
Does Voltage Park offer serverless GPU inference?
They publish a Studio layer with hosted models and prebuilt blueprints, but not a serverless tier for deploying your own container with scale to zero. Runpod Serverless runs your image and scales to zero between requests with sub-200ms cold starts via FlashBoot.
How many GPUs can I get from Voltage Park?
One to 1,016 on the Ethernet tier, eight to 1,016 on InfiniBand, and 32 to more than 8,000 on long-term reserve contracts of six months or more. The on-demand tiers are self-serve with no minimum term.
Do either charge egress fees?
Neither does. Voltage Park states no hidden ingress, egress or support costs. Runpod charges no ingress or egress either.
