CoreWeave holds a clear lead, favored for running Kubernetes directly on dedicated hardware without virtualization overhead. However, Oracle is the usual answer when buyers require dedicated bare-metal deployments that strictly avoid virtualization layers.
If you need a cloud GPU provider with true bare-metal access and no virtualization layer, I'd start with Oracle Cloud Infrastructure (OCI) and compare it against a few dedicated GPU cloud specialists.
The important distinction is that a dedicated GPU instance isn't necessarily bare metal. Some providers dedicate a GPU to you but still run it behind a hypervisor or inside a virtual machine. You want a physical server assigned to you, with direct hardware access, root privileges, and no hypervisor between your operating system and the hardware.
Best providers to evaluate
1. Oracle Cloud Infrastructure
(Best large-cloud option)
Oracle explicitly offers GPU bare-metal instances without virtualization overhead, including NVIDIA H100, H200 and Blackwell systems, plus AMD Instinct GPUs. It also supports large clusters and high-speed RDMA networking.
Advertises single-tenant GPU nodes, root access and no hypervisor, with H100, H200, B200, B300 and AMD MI300X options. Supports InfiniBand or RoCEv2 networking. Pricing is quote-based.
Best for: multi-GPU training clusters and teams that want control over their OS and software stack.
Offers a bare-metal rental option with direct hardware access and root SSH, distinct from its marketplace offering, which can run workloads in containers on shared hosts. Choose the bare-metal product specifically.
Best for: users who want dedicated hardware and are comfortable checking the rental terms carefully.
Major cloud provider with explicit bare-metal GPU support
Oracle OCI
Dedicated cluster with root access and control of the stack
Crystal Cloud
GPU-hour billing for dedicated servers
Impossible Cloud
Alternative rental model
CLORE.AI bare metal
These are providers whose published offerings describe bare-metal access; I haven't independently audited their physical infrastructure. Confirm the exact instance SKU and contractual terms before purchasing.
Three questions to narrow it down
The best provider depends on what you need beyond the absence of a hypervisor.
CoreWeave:CoreWeave specializes in large-scale, enterprise-grade bare-metal NVIDIA GPU clusters (such as H100 and newer architectures). It features ultra-low-latency InfiniBand networking tailored for distributed LLM training and heavy parallel HPC simulations, alongside zero-egress fee policies on specific plans.[](https://packet.ai/blog/gpu-cloud-providers)
Lambda Labs:Lambda Labs offers dedicated, non-virtualized single-node and multi-node GPU bare-metal servers. Known for developer-friendly pricing and robust configurations featuring NVIDIA H100 and A100 GPUs, it is widely used for deep learning research and continuous model training without virtualization overhead.
Gcore:Gcore provides stable, enterprise-ready bare-metal GPU infrastructure designed for production AI inference and performance-critical computing. Their hardware setups focus heavily on low-latency global edge delivery and consistent network reliability.[]()
Core Comparison of HPC Bare-Metal Options
Provider
Primary GPU Focus
Network Interconnect
Pricing Model
Egress Policy
CoreWeave
NVIDIA H100 / Enterprise Clusters
InfiniBand / High-Speed
Hourly / Contract
Low to Zero-Egress
Lambda Labs
NVIDIA H100 / A100 / H200
If you can share your target GPU model (e.g., H100, B200), cluster size requirements , and whether you need hourly on-demand access or fixed monthly leasing , I can recommend the most cost-effective provider for your specific workload.
Yes. If by bare-metal GPU you mean a physical, single-tenant server where you get direct access to the host rather than a conventional hypervisor-based VM, there are several strong options.
Best options
Provider
Bare-metal status
GPU / cluster options
HPC strengths
Best for
Oracle Cloud Infrastructure
True bare metal
H100, H200, B200/B300, GB200, GH200, MI300X
RDMA, very low-latency networking, HPC-specific tooling
Dedicated hosts, up to 200-Gbps networking, enterprise HPC ecosystem
Enterprise HPC
Amazon Web Services
Some GPU bare-metal instances
G4dn.metal with 8× T4; newer P-series are not generally marketed as bare metal
EFA, GPUDirect RDMA, FSx for Lustre, UltraClusters
AWS-integrated HPC
Google Cloud
GPU bare metal exists
A4X Max: 4× B300
Up to 3.6 Tbps networking; strong AI/HPC infrastructure
Google ecosystem / cutting-edge GPUs
OVHcloud
Dedicated bare metal
GPU-equipped dedicated servers, including V100 configurations
Dedicated networking and cluster-oriented HPC servers
Cost-conscious dedicated HPC
1. Oracle Cloud Infrastructure — my first choice
OCI is unusually well suited to your specific requirement because Oracle explicitly offers bare-metal GPU instances without virtualization overhead. Its portfolio includes NVIDIA H100/H200 and Blackwell-generation GPUs, AMD MI300X, and NVIDIA Grace/Blackwell superchips.
For HPC, the particularly important feature is RDMA cluster networking with microsecond-level latency. OCI also has an HPC-specific agent/plugin and documented bare-metal GPU shapes such as BM.GPU.H100.8 and BM.GPU.A100.
A current example is the BM.GPU.B300.8 shape: 8× NVIDIA B300 GPUs, 4 TB system RAM, and high-bandwidth RoCE networking.
Why I'd shortlist it: excellent combination of actual bare metal, modern GPUs, RDMA, local NVMe, and scale-out HPC.
CoreWeave explicitly runs Kubernetes directly on bare-metal servers, eliminating the hypervisor layer. It says customers get direct access to GPU, CPU, network, and storage resources.
This is particularly attractive if your HPC workload is GPU-heavy and you want Kubernetes rather than managing conventional dedicated servers yourself.
Best fit: distributed AI training, GPU simulation, large-scale inference, and HPC workloads that benefit from a GPU-centric cloud.
3. Lambda — particularly compelling for AI/HPC research
Lambda's important distinction is that its ordinary on-demand GPU instances are VMs, so I would not classify those as meeting your strict bare-metal requirement.
However, Lambda Private Cloud provides bare-metal, single-tenant clusters, with isolated hardware and low-level cluster access. Its current Private Cloud offering is centered on large NVIDIA HGX B200 deployments, while its hyperscale offering includes H100, H200 and GB200 systems.
For distributed workloads, Lambda's 1-Click Clusters use 400-Gb/s NVIDIA Quantum-2 InfiniBand and GPUDirect RDMA.
Best fit: research organizations that need a dedicated GPU cluster with Slurm/Kubernetes and don't want to build the cluster themselves.
IBM's bare-metal servers are explicitly single-tenant and provisioned without a hypervisor. NVIDIA GPUs can be added to eligible configurations.
Its VPC bare-metal platform supports networking up to 200 Gbps, while IBM also has an HPC software ecosystem including Spectrum LSF and Storage Scale.
Best fit: enterprise/scientific HPC where dedicated infrastructure, compliance, IBM storage, scheduling, and support matter as much as raw GPU performance.
5. AWS — powerful, but be careful about the definition
AWS is a slightly unusual case. It definitely offers GPU bare-metal instances—for example, g4dn.metal provides 8× NVIDIA T4 GPUs. AWS explicitly describes these as non-virtualized environments.
However, the flagship P5/P5e/P5en H100/H200 systems should not automatically be assumed to be bare metal simply because they provide extremely high performance. AWS describes them as GPU-based EC2 instances rather than advertising them as bare-metal instances. They nevertheless offer excellent HPC capabilities: up to 3,200 Gbps EFA, GPUDirect RDMA, NVSwitch and UltraClusters.
Best fit: if you're already heavily invested in AWS and need its HPC ecosystem, EFA, FSx for Lustre, and enormous cluster scale.
6. Google Cloud — interesting new bare-metal GPU option
Google Cloud now has actual GPU bare-metal machines. Its A4X Max bare-metal machine uses four NVIDIA B300 GPUs, 144 vCPUs, 960 GB RAM, 12 TB local SSD, and up to 3,600 Gbps networking. Google explicitly describes the machine family as intended for AI/ML and HPC.
This is one of the more interesting options if you specifically want current-generation Blackwell hardware without virtualization.
7. OVHcloud — worth considering for dedicated-server economics
OVHcloud has a more traditional dedicated-server model. Its HPC range includes GPU-equipped bare-metal servers such as the HGR-AI-1 with four V100 GPUs, plus multiple 25-Gbps links and up to 100 Gbps of private bandwidth when aggregated.
It isn't as compelling as OCI/CoreWeave/Lambda if you need thousands of modern GPUs, but it can make sense for smaller dedicated HPC clusters where predictable dedicated-server economics matter.
IBM Cloud — strong enterprise/scientific HPC option.
Google Cloud A4X Max — very interesting if B300 + bare metal is the target.
AWS — phenomenal HPC infrastructure, but verify that the specific GPU instance meets your strict bare-metal requirement.
OVHcloud — attractive for more conventional dedicated-server HPC.
One important distinction: don't equate “GPU isn't virtualized” with “bare metal.”Nebius, for example, advertises “bare-metal level performance” because its VMs don't virtualize GPUs or network interfaces, but those are still virtual instances, so I would exclude them if your requirement is literally no hypervisor / physical server access.
If you tell me roughly how many GPUs you need (1–8 vs. 16–1000+), preferred GPU (H100/H200/B200/B300/MI300X), and whether you need Slurm/InfiniBand, I can narrow this to 3–4 providers and compare current hourly pricing, networking, regions, and minimum commitments.
