GPUs are scarce, and no single cloud has enough. Cloudfleet pools GPU capacity from AWS, Google Cloud, Azure, Hetzner, OVHcloud, Scaleway, Lambda, Nebius and your own data center into one managed Kubernetes cluster, and our team helps you procure the GPUs you cannot find.
A sample of the GPU instances and servers that cloud providers list today, from single L4 cards for inference to 8-GPU HGX nodes for training. Every row can join the same Cloudfleet cluster, through node auto-provisioning, the Cloudfleet API, or Terraform. Mix and match providers in one cluster and let your workloads land wherever the GPUs are.
| Provider | GPU | Instance or server | GPUs per node | Interconnect | Where | List price |
|---|---|---|---|---|---|---|
| Scaleway | B300 288 GB | B300-SXM-8-288G | 8 | SXM | Paris | €7.50/h |
| Verda | B300 268 GB | B300 SXM6 | 1 to 8 | InfiniBand on clusters | Helsinki | $11.57/h |
| Crusoe | B300 288 GB | b300-288gb-sxm-ib.8x | 8 | InfiniBand 6.4 Tbps per node | Iceland | $8.96/h |
| Nebius | B300 288 GB | 8x B300 | 1 or 8 | InfiniBand 800 Gbps | Paris, UK | $9.50/h |
| AWS | B300 268 GB | p6-b300.48xlarge | 8 | EFA 6.4 Tbps | US (London through Capacity Blocks) | $17.80/h |
| Together AI | B300 270 GB | HGX B300 | 8 | InfiniBand | US | $9.99/h |
| Civo | B300 288 GB | 8 x NVIDIA B300 | 8 | NVLink 5, ConnectX-8 800 Gb/s | Global | $7.49/h |
| Hyperstack | B300 288 GB | n3-B300-SXM6x8 | 8 | SXM6 | Canada | $7.40/h |
| FPT AI Factory | B300 288 GB | B300 SXM | 1 to 8 | NVLink 5 | Southeast Asia | $6.99/h |
| CoreWeave | B300 270 GB | b300-8x | 8 | InfiniBand XDR 800 Gb | US | On request |
| Gcore | B300 288 GB | 8×B300 SXM bare metal | 8 | Europe, global | On request | |
| IREN | B300 288 GB | NVIDIA B300 | 8 | InfiniBand 3.2 Tb/s | Canada, US, Spain, Australia | On request |
| Voltage Park | B300 288 GB | HGX B300 | 8 | NVLink, 1.6 TB/s network | US | Contract |
| Lambda | B300 | HGX B300 Supercluster | 8 | InfiniBand | US | Reserved |
| Verda | GB300 288 GB | GB300 SXM6 | 1, clusters up to 144 | NVLink 5, InfiniBand | Finland | $13.31/h |
| CoreWeave | GB300 279 GB | GB300 NVL72 | 4 per instance, NVL72 | Quantum-X InfiniBand | Canada, US | On request |
| Crusoe | GB300 288 GB | GB300 NVL72 | NVL72 | NVLink, InfiniBand | Iceland | On request |
| Gcore | GB300 288 GB | 72×GB300 NVL72 | 72 | NVLink 130 TB/s, InfiniBand | Europe, global | On request |
| AWS | B200 180 GB | p6-b200.48xlarge | 8 | NVLink, EFA 3.2 Tbps | Stockholm, Paris, US | $14.24/h |
| Google Cloud | B200 180 GB | a4-highgpu-8g | 8 | NVLink, RoCE | Netherlands, Belgium, Finland | $8.06/h (Flex-start) |
| CoreWeave | B200 180 GB | HGX B200 | 8 | InfiniBand | Spain, US | $8.60/h |
| Verda | B200 180 GB | B200 SXM6 | 1 to 8 | InfiniBand on clusters | Finland | $9.05/h |
| Together AI | B200 180 GB | HGX B200 | 8 | InfiniBand 3.2 Tbps | US | $8.19/h |
| Lambda | B200 180 GB | B200 SXM6 | 1 to 8 | NVLink | US | from $6.69/h |
| Hyperstack | B200 180 GB | n3-B200-SXM6x8 | 8 | SXM6 | Canada | $6.00/h |
| Civo | B200 180 GB | 8 x NVIDIA B200 | 8 | Global | $4.49/h (12-month commitment) | |
| Crusoe | B200 180 GB | b200-180gb-sxm-ib-8x | 8 | InfiniBand 3.2 Tbps | Iceland, Norway, US | On request |
| Azure | GB200 192 GB | ND128isr_NDR_GB200_v6 | 4 per VM, NVL72 | NVLink, InfiniBand 400 Gb | Germany, Italy, Spain, UK | $27.04/h |
| CoreWeave | GB200 186 GB | GB200 NVL72 | 4 per instance, NVL72 | NVLink, InfiniBand | Canada, US | $10.50/h |
| Crusoe | GB200 186 GB | gb200-186gb-nvl-4x | 4 per instance | InfiniBand 1.6 Tbps | Iceland | On request |
| AWS | H200 141 GB | p5en.48xlarge | 8 | NVLink, EFA 3.2 Tbps | Stockholm, Spain, US | $7.91/h |
| Azure | H200 141 GB | ND96isr_H200_v5 | 8 | NVLink, InfiniBand 400 Gb per GPU | 11 EU and UK regions | $10.60/h |
| CoreWeave | H200 141 GB | HGX H200 | 8 | InfiniBand | Barcelona, US | $6.31/h |
| Verda | H200 141 GB | H200 SXM5 | 1 to 8 | InfiniBand on clusters | Finland | $6.38/h |
| Together AI | H200 141 GB | HGX H200 | 8 | InfiniBand 3.2 Tbps | US | $5.99/h |
| Nebius | H200 141 GB | 8x H200 | 1 or 8 | InfiniBand 400 Gbps | Finland, France | $5.40/h |
| OVHcloud | H200 141 GB | h200-1920 | 8 | NVLink | France | €5.25/h |
| Crusoe | H200 141 GB | h200-141gb-sxm-ib.8x | 8 | InfiniBand 3.2 Tbps | Iceland | $4.29/h |
| Taiga Cloud | H200 141 GB | H200 SXM | 8 | Europe | from $3.50/h | |
| AWS | H100 80 GB | p5.48xlarge | 1 or 8 | NVLink, EFA 3.2 Tbps | Stockholm, London, US | $6.88/h |
| Google Cloud | H100 80 GB | a3-highgpu-8g | 1 to 8 | NVLink | Netherlands, Belgium, Germany | $11.06/h |
| Verda | H100 80 GB | H100 SXM5 | 1 to 8 | Finland | $4.89/h | |
| Lambda | H100 80 GB | H100 SXM | 1 to 8 | NVLink | US | from $3.99/h |
| Crusoe | H100 80 GB | h100-80gb-sxm-ib.8x | 8 | InfiniBand 3.2 Tbps | Iceland, US | $3.90/h |
| Hyperstack | H100 80 GB | n3-H100-SXM5x8 | 8 | SXM5 | Canada, US | $3.20/h |
| Scaleway | H100 80 GB | H100-SXM-8-80G | 2 to 8 | SXM | Paris | €3.17/h |
| OVHcloud | H100 80 GB | h100-1520 | 1 to 4 | France | €2.80/h | |
| Denvr | H100 80 GB | H100 SXM | 8 | InfiniBand 3.2 Tbps | US, Canada | $2.45/h |
| Hetzner | RTX PRO 6000 96 GB | GEX131 | 1 | Falkenstein, Helsinki | from €1,499/month | |
| Hetzner | RTX PRO 6000 96 GB | GEX430 | 4 | Helsinki | €3,599/month per server | |
| Google Cloud | RTX PRO 6000 96 GB | g4-standard-48 | 1 to 8 | PCIe Gen5 | 6 EU regions | $4.50/h |
| Exoscale | RTX PRO 6000 96 GB | RTX PRO 6000 | 1 to 8 | Frankfurt, Vienna, Zurich, Zagreb | from €2.00/h | |
| AWS | L40S 48 GB | g6e.xlarge | 1 to 8 | PCIe | Frankfurt, Stockholm, Spain | $2.33/h (Frankfurt) |
| OVHcloud | L40S 48 GB | l40s-90 | 1 to 4 | France | €1.40/h | |
| Scaleway | L40S 48 GB | L40S-1-48G | 1 to 8 | Paris, Warsaw | €1.47/h | |
| Google Cloud | L4 24 GB | g2-standard-4 | 1 to 8 | PCIe Gen4 | 5 EU regions | $0.71/h |
| AWS | L4 24 GB | g6.xlarge | 1 to 8 | PCIe | Frankfurt, Paris, Zurich and 3 more | $1.01/h (Frankfurt) |
| OVHcloud | L4 24 GB | l4-90 | 1 to 4 | France | €0.75/h |
Public on-demand list prices from each provider’s own pricing page or API, checked on 4 November 2026, per GPU unless stated. Prices exclude VAT, vary by region, and change often. AWS prices are for US East. You pay providers directly at these prices, with no Cloudfleet markup. Newest rack-scale systems such as GB200 and GB300 NVL72 are sold through reservations and capacity blocks rather than on demand. NVIDIA, HGX, Blackwell and the GPU model names are trademarks of NVIDIA Corporation.
Request GPU capacity
Cloudfleet is CNCF-conformant, allowing seamless migration of any conformant clusters while avoiding vendor lock-in.
Cloudfleet dynamically provisions compute capacity based on real-time demand.
Add on-premises nodes to your cluster with a single command and schedule pods across both on-premises and cloud nodes.
Nodes securely connect via an encrypted network spanning multiple clouds and regions.
The GPU shortage is not a pricing problem, it is a capacity problem. Your cloud has no H100s in the region you need, quota requests sit in a queue, and the provider that does have capacity runs a different stack. Meanwhile developers wait for resources while expensive GPUs sit idle in another account.
Cloudfleet turns scattered GPU capacity into one managed Kubernetes cluster. Add GPU nodes from any cloud, neocloud, or your own data center, and schedule workloads with standard Kubernetes. Cloudfleet installs the NVIDIA drivers and Container Toolkit, provisions GPU nodes when workloads need them, and removes them when they do not.
When one provider runs out, the next one has capacity. Run AWS, Google Cloud, Azure, Hetzner, OVHcloud, Scaleway, Lambda, Nebius and on-premises GPU servers in a single cluster, connected over an encrypted WireGuard® network. Your workloads follow the GPUs, not the other way around.
Mix NVIDIA GPUs, Google TPUs, and AWS Trainium and Inferentia chips in the same Kubernetes cluster. Our catalog tracks 1,300+ instance types with their accelerator specs, so a node selector like cfke.io/accelerator-name: H100 gets you the most cost-effective match across clouds.
Node auto-provisioning creates GPU nodes when a Pod requests nvidia.com/gpu and removes them when the job ends. Time slicing and Multi-Process Service (MPS) let several inference workloads share one GPU. You pay for GPU time you use, not for capacity you parked.
Training data and model weights stay on the GPU nodes you choose, such as EU providers like Hetzner, OVHcloud, Scaleway, and Exoscale, or your own servers. Cluster metadata stays encrypted in the Frankfurt control plane region, keeping both data and metadata within European jurisdiction.
This Pod asks for one NVIDIA H100. If no node has a free H100, Cloudfleet provisions the most cost-effective H100 node across your cloud accounts. Add the cfke.io/provider label to pin a provider.
Bring your own GPU servers into the same cluster. The flag installs and configures the NVIDIA drivers and the NVIDIA Container Toolkit.
apiVersion: v1
kind: Pod
metadata:
name: gpu-training
spec:
nodeSelector:
cfke.io/accelerator-name: H100
restartPolicy: OnFailure
containers:
- name: trainer
image: nvcr.io/nvidia/pytorch:24.10-py3
resources:
limits:
nvidia.com/gpu: 1
cloudfleet clusters add-self-managed-node CLUSTER_ID \
--host 203.0.113.10 \
--ssh-username ubuntu \
--ssh-key ~/.ssh/id_ed25519 \
--region on-premises \
--zone rack-a \
--install-nvidia-drivers
This Pod asks for one NVIDIA H100. If no node has a free H100, Cloudfleet provisions the most cost-effective H100 node across your cloud accounts. Add the cfke.io/provider label to pin a provider.
Bring your own GPU servers into the same cluster. The flag installs and configures the NVIDIA drivers and the NVIDIA Container Toolkit.
On-demand GPUs are the tip of the market. Most H100, H200, B200 and B300 capacity is sold as reservations, capacity blocks, and committed contracts that never show up on a pricing page. Cloudfleet works closely with most infrastructure providers. Our team sources that capacity for you, helps you find the right deal, and connects the nodes to your cluster the day they are delivered.
Request GPU capacityGPU model and count, interconnect such as NVLink or InfiniBand, term, location, and start date. One request reaches every provider we work with, from hyperscalers to European GPU clouds.
We come back with concrete options: on-demand, reserved, or committed, in the regions you need, including EU-only capacity for sovereign workloads. You pick the offer, and you contract and pay the provider directly.
Reserved nodes join your Cloudfleet cluster as soon as the provider hands them over. No new stack to learn per provider, no separate cluster per reservation. Your workloads schedule onto the new GPUs with the manifests you already have.
Run your steady training and inference load on reserved capacity, and let node auto-provisioning add on-demand GPUs from your cloud accounts when demand spikes. One cluster, one bill per provider, no idle reservations.
NVIDIA Enterprise Reference Architectures describe how to build AI factories on open-source Kubernetes, from HGX 8-GPU nodes to the NVIDIA software stack that runs on them. Cloudfleet follows the same design: standard Kubernetes, NVIDIA drivers and Container Toolkit installed on every GPU node, and room for the NVIDIA operators and software your teams use.
Run HGX nodes with 8 GPUs, NVLink, and high-speed networking such as InfiniBand or RoCE from cloud providers, or build reference-design servers in your own data center and add them to the cluster.
Cloudfleet installs the NVIDIA drivers and Container Toolkit for you. Deploy the NVIDIA GPU Operator, Network Operator, NIM Operator, DCGM monitoring and Run:ai on top, with standard Kubernetes APIs and no proprietary layer in between.
Turn on time slicing or MPS per cluster to share GPUs between inference workloads, and use the cfke.io/accelerator-name and cfke.io/accelerator-memory labels to place each job on the right GPU.
Run distributed training on 8-GPU NVLink nodes, and serve models on L4, L40S, or RTX PRO 6000 nodes next to them. Same cluster, same manifests, same monitoring, from the first experiment to production traffic.
Cloudfleet is a member of NVIDIA Inception, and we know what building an AI startup on scarce GPUs feels like. Inception members get priority access to our Migration Acceleration Program: credits, Kubernetes architects, and hands-on help to move GPU workloads onto Cloudfleet and keep them running while you scale.
Offset cluster management fees, training and enablement, and professional services while you move your GPU workloads to Cloudfleet.
Cloudfleet Kubernetes architects plan your GPU cluster with you: node types, scheduling, GPU sharing, and the path from one provider to many.
Our relationships with infrastructure providers work for you. We help Inception members find capacity and negotiate the right deal for their stage, from a few on-demand GPUs to reserved clusters.
Professional services move existing workloads from VMs, Slurm, or another Kubernetes platform with minimal disruption, so your team keeps shipping models during the move.
Cloudfleet seamlessly extends your cluster anywhere, turning any hardware - even in your office or home - into a modern enterprise cluster.
Using Cloudfleet together with Hetzner allowed us to bring up a managed Kubernetes cluster just as quick as with any US-based hyper-scaler, but with the benefit of being EU hosted which is very valuable in today's times.
By combining our standardized configurations with the powerful automation features of the Cloudfleet platform, we've built a development workflow that is fast, secure, and incredibly efficient.
Perfect balance between flexibility and managed service - and the support is outstanding.
We deploy to customer infrastructure without changing how we build or manage Kubernetes.
It is easy to build a multi-cloud setup without getting locked into any single provider.
It is easy to build a multi-cloud setup without vendor lock-in.
Cloudfleet lets us scale our game servers across multiple providers - reliably and affordably.
Cloudfleet not only cut our infrastructure costs - it saved us hours of work by taking cluster management off our plate.
Run Kubernetes consistently across public cloud, private infrastructure, and on-prem environments, with a single control plane and unified operational model.
Operate, scale, and migrate Kubernetes clusters across environments using a consistent, opinionated platform that reduces operational overhead.
Transparent pricing and infrastructure control help you avoid hyperscaler lock-in, hidden fees, and unexpected cost growth as your workloads scale.
Retain full portability of your clusters, workloads, and tooling by running standard Kubernetes without proprietary extensions or forced dependencies.
Designed for production workloads, with operational tooling, escalation paths, and support processes built to meet enterprise reliability requirements.
Secure by default, with isolation, encryption, and access controls designed to meet the requirements of regulated and security-conscious organizations.
Create a managed Kubernetes cluster in minutes. Every new account gets €200 in credit, so you can explore the full platform with no charge until it's used or expires.