GPU cloud pricing,
priced by the hour.
Reserved when you scale.

On-demand H100 SXM pricing starts at $3.66 per GPU per hour, with hourly billing and no minimum commitment. With a 12-month reserved commitment the rate drops to $2.38 per GPU per hour — about 35% off. H100 PCIe (single-GPU) starts at $2.99/hr on-demand. Multi-GPU 8×H100 SXM nodes with NVLink and InfiniBand are available from $28.50/hr.

On-demand is pay-as-you-go: spin up a GPU, get billed by the hour, shut down when done. No commitment, but you pay the standard hourly rate. Reserved means you commit to 6 or 12 months in exchange for a discount — 22% off for 6-month, 35% off for 12-month. Reserved is right for steady workloads (production inference, ongoing training). On-demand fits bursty work, experiments, and prototyping.

Self-serve single-GPU and multi-GPU instances come up in under 60 seconds for available SKUs. Pre-built CUDA/PyTorch/TensorRT/vLLM images mean no driver setup. For reserved clusters of 32+ H100s with custom networking or bare-metal isolation, expect 24–72 hours from contract to first login.

For throughput-heavy LLM serving (vLLM, TGI, TensorRT-LLM) the L40S usually wins on price-per-token at the 7B–34B range. For large-context or 70B+ models the H100 PCIe is the sweet spot. For edge or always-on small models, the L4 gives the lowest hourly cost. A100 80GB still fits well if you're already on that stack and need predictable behaviour.

A rough rule: parameters × bytes per parameter × 1.25 (for activations and KV cache headroom). A 7B model in BF16 needs roughly 17 GB; in FP8 around 9 GB. A 70B model in BF16 fits on a single 80GB H100; in FP16 you'll need 2× A100 80GB. Use quantization (INT8, INT4, FP8) to shrink the memory footprint — most production inference runs quantized.

No egress fees, no ingress fees. Move datasets and checkpoints freely. Persistent block storage is billed separately at $0.08/GB/month with snapshot support. S3-compatible object storage starts at $0.018/GB/month. There's no platform fee and no minimum monthly spend.

Yes — managed Kubernetes ships with the NVIDIA device plugin, MIG slicing, GPU autoscaler, and Prometheus add-ons. Control plane is $72/cluster/month on Standard, and GPU node pools bill at the same per-hour rate as GPU as a Service — there's no Kubernetes upcharge on the compute itself.

H100 is the right call if your workload benefits from FP8 (modern training, frontier inference) or if you need NVLink 4.0 and 3.35 TB/s memory bandwidth — typically 70B+ training, KV-cache-heavy serving, or latency-sensitive inference. A100 is the cost-efficient default for mature workloads on a stable CUDA stack — 7B–34B fine-tuning, classical deep learning, and production inference that doesn't need FP8.

Standard instances run in per-tenant VPCs with encrypted-at-rest volumes. For regulated workloads we offer single-tenant bare-metal options where the entire host is yours — no shared CPU, no shared memory, no neighbouring tenants. We are SOC 2 Type II and ISO 27001 certified, with DPDP-compliant India data residency available on request.