CloudRift rents GPU instances. You pick a card, it boots, you pay per second until you stop it. That is a clean product, with a legible public rate card and no egress, ingress or API-call charges. They also sell a hosted inference API and an orchestration platform for enterprises running their own datacenters.
Runpod covers three different shapes of work on one account: persistent development boxes, inference that scales to zero between requests, and multi-node training clusters. The same container image runs on all three.
So the useful question is not which hourly rate is lower. It is whether renting a machine by the hour is the right shape for what you are building, and what happens to your setup when the answer changes.
What you can actually order from CloudRift
Their rate card lists twelve GPUs with prices next to each, which reads as fully self-serve. It is not, at the tier most teams care about.
The H100, H200 and B200 show reserved rates only. Their own FAQ says the H100 and H200 are not orderable in the console at all: you contact them. If your plan is to sign up this afternoon and have an H100 running, that plan does not survive contact with the product. On Runpod an H100 PCIe deploys from the console in under 30 seconds.
Check the VRAM column before you compare anything. Their H100 row lists 48GB rather than the usual 80GB, and their RTX 4090 and RTX PRO 6000 rows both list 96GB, which is not a standard 4090 configuration. Whatever those figures describe, they are not like-for-like with the same card names elsewhere, and any comparison that ignores them is comparing labels rather than hardware.
The floor is old silicon. CloudRift's cheapest cards are V100s from 2017. Runpod publishes ten current-generation models below the L40S tier, which is where prototyping, embeddings, batch image work and small-model serving actually live.
The capability gap: serving and training
This is the part that does not show up on a rate card at all.
An instance bills whether or not a request arrives. If you are serving a model with uneven traffic, the machine sits warm and metered through every quiet hour. Runpod Serverless scales to zero between requests and starts back up in under 200ms via FlashBoot, so you pay for inference rather than for readiness. That is a structural difference, and at low or bursty traffic it dwarfs any hourly-rate gap in either direction.
CloudRift's answer to this is a hosted inference API rather than serverless GPUs. They run an OpenAI-compatible, vLLM-based endpoint billed per token, which is a good fit if the model you want is one they host. At the time of writing their public catalog lists a single model. If you need to serve your own fine-tune, your own weights or a model they do not carry, that route is closed and you are back to renting an instance that bills continuously.
Multi-node training is not published as part of their rental product. Runpod Instant Clusters provision from the console without a contract or a sales call. If your roadmap includes fine-tuning something that does not fit on one node, that is a platform you would otherwise have to go find separately.
One account, one image, three modes. Prototype on a Pod, deploy the same container to a Serverless endpoint, train on a Cluster. Nothing gets rebuilt when the workload changes shape. Renting instances means assembling the rest yourself.
Where CloudRift is genuinely good
Their pricing is the most legible in this category. On-demand and reserved rates sit side by side for every card, the reserved tiers are selectable in the console rather than negotiated, local storage is included in the hourly rate, and there are no transfer or API charges to model.
They also carry an AMD MI350X, which is rare. If you want current AMD silicon on a public rate card, they have it, and AMD is a first-class path on their platform rather than an afterthought.
Their enterprise product is a different business, and a credible one. CloudRift also sells an orchestration and operations layer for organizations that own GPU hardware and run their own datacenters: VMs, containers, MIG and bare metal from one control plane, with multi-tenancy, quotas, RBAC and audit logging, and the ability to run fully air-gapped. If your problem is that you already bought the GPUs and need to make them usable by internal teams, that is a genuine offering and Runpod does not compete with it.
Pricing
On several mid-range cards CloudRift is cheaper than Runpod Community Cloud. That is true, easy to check, and worth stating plainly.
Their A100 SXM4, L40S and RTX 5090 all sit below the equivalent Runpod Community Cloud rates. The RTX 4090 goes the other way. The H100 comparison is not really a comparison, since theirs is a reserved rate you have to arrange and ours deploys on demand.
The thing to weigh against those numbers is what the hourly rate is buying. A cheaper instance that runs 24 hours to serve six hours of traffic is more expensive than a dearer Serverless worker that scales to zero. Model your actual duty cycle before letting the rate card decide.
Which one should you choose
Choose CloudRift if you want a GPU instance and nothing else, your workload sits on an A100, L40S or RTX 5090, it runs steadily enough to use their reserved tiers, and you value a single legible rate card. On those cards they are cheaper and the pricing is easy to reason about. Choose them too if you own GPU hardware and need an operations layer for it, which is a different product and one Runpod does not offer.
Choose Runpod if you need an H100 on demand rather than by arrangement, if you are serving your own model and want it to scale to zero between requests, if you need multi-node training, if your work fits on a card below their range, or if you would rather not assemble a serving and training stack from separate vendors.
The split: CloudRift is a good, transparently priced GPU rental service with a growing enterprise platform alongside it. Runpod is the AI developer cloud, covering development, serving and training on one account. If you are buying a machine, price both. If you are building a product, the hourly rate is the smallest part of the decision.
Frequently asked questions
What is the main difference between Runpod and CloudRift?
Scope. CloudRift rents GPU instances billed per second, alongside a hosted inference API and an orchestration platform for on-premise fleets. Runpod covers persistent Pods for development, Serverless for running your own container with scale to zero, and Instant Clusters for multi-node training, all on one account with the same container images.
Can I rent an H100 from CloudRift on demand?
Not from the console. Their H100 and H200 are listed at reserved rates only, and their FAQ directs you to contact them for both. Runpod deploys an H100 PCIe from the console at {{gpu:h100-pcie}}/hr on Secure Cloud.
Does CloudRift offer serverless GPUs?
Not in the sense of deploying your own container and having it scale to zero. They offer a hosted, OpenAI-compatible inference API billed per token, which covers the models they host but not your own weights. Runpod Serverless runs your container, scales to zero between requests, and cold starts in under 200ms via FlashBoot.
Can I run multi-node training on CloudRift?
Multi-node clusters are not published as part of their GPU rental product, which lists a maximum of 8 to 16 GPUs per instance. Runpod Instant Clusters provision self-serve from the console at {{cluster:h200-sxm}}/hr for H200 SXM and {{cluster:a100-sxm}}/hr for A100 SXM, with no contract.
Is CloudRift cheaper than Runpod?
On some cards. As of August 2026, their A100 SXM4 at $1.05/hr, L40S at $0.63 and RTX 5090 at $0.65 are below the equivalent Runpod Community Cloud rates. Their RTX 4090 at $0.39 is above Runpod's {{gpu:rtx-4090:community}}/hr. Below the L40S there is little to compare, because their only cheaper cards are V100s from 2017.
How many GPU models does each offer?
CloudRift publishes twelve, from a V100 SXM2 up to an AMD MI350X. Runpod publishes 24, from an RTX A5000 through to a B300, including MI300X, across 31 global regions.
Do either charge egress fees?
Neither does. CloudRift states no egress, ingress or API-call charges. Runpod charges no ingress or egress either.
