Home/Bare Metal
Core product

Bare-metal servers.

Dedicated nodes with zero virtualization overhead. The full silicon — CPU, memory, GPUs, and fabric — belongs to your workload alone. Provisioned in minutes via API, CLI, or console.

What you get

Full control, none of the plumbing.

Single-tenant isolation

No noisy neighbors, no shared caches, no co-tenancy surprises. Your node is yours, physically and logically.

Your OS, your stack

Install anything: custom kernels, InfiniBand drivers, your own container runtime. Full root, full freedom.

Metered power

Per-second billing, spot capacity for burst jobs, and reserved fleets with predictable monthly pricing.

API-first provisioning

Spin nodes up and down from your CI pipeline. A single POST request creates a fully networked node.

Fabric networking

Connect nodes over RDMA fabrics with up to 1.2 TB/s per GPU for tightly coupled distributed training.

Agreed SLAs

Service-level commitments are stated in your agreement, with proactive parts replacement and a live engineer on every ticket.

Node configurations

Pick your chassis.

ConfigAcceleratorGPUs/nodeMemory/GPUBest for
h100-8NVIDIA H100880 GBPrototyping, CI, LoRA tuning
h200-8NVIDIA H2008141 GBProduction fine-tuning, inference
b200-8NVIDIA B2008192 GB70B-class training fabrics
b300-8NVIDIA B3008288 GBFrontier pretraining, long-context inference
Accelerator comparison

Side by side, at a glance.

SpecNVIDIA H100NVIDIA H200NVIDIA B200NVIDIA B300
ArchitectureHopperHopperBlackwellBlackwell Ultra
Memory80 GB HBM3141 GB HBM3e192 GB HBM3e288 GB HBM3e
Memory bandwidth3.35 TB/s4.8 TB/s8 TB/s8 TB/s
FP16 / BF16 (dense)989 TFLOPS989 TFLOPS2.25 PFLOPS3.5 PFLOPS
FP8 (dense)1,979 TFLOPS1,979 TFLOPS4.5 PFLOPS7 PFLOPS
FP4 (dense)9 PFLOPS15 PFLOPS
NVLink900 GB/s900 GB/s1.8 TB/s1.8 TB/s
Power (per GPU)700 W700 W1,000 W1,200 W
RouteH100 →H200 →B200 →B300 →

Ready to push the frontier?

Provision bare-metal GPUs in minutes, scale capacity on demand, and get a real engineer on call — not a ticket queue.