NVIDIA H100 PCIe + SXM · 80 GB
Rent an H100 in seconds. No egress fees.
Training & LLM serving. From $2.42/GPU-hr — about 65% below the ~$6.88/hr hyperscaler on-demand rate. Billing starts only after the SSH/Jupyter handshake succeeds.
H100 pricing
| Option | Per GPU-hr | ~Per month (24/7) | Egress |
|---|---|---|---|
| CloudFlux Indie | $2.42 | $1766.60 | $0 (2,000 GB included) |
| CloudFlux Enterprise | $2.35 | $1718.42 | $0 (50,000 GB included) |
| Typical hyperscaler on-demand | $6.88 | $5022.40 | Billed per GB |
Prices last updated 2026-10-11. Pro and Team plans add extra compute discounts; Enterprise 6- and 12-month terms save a further 3–5%. Over the included amount, bandwidth is rate-limited, never charged.
Instance specs
| H100 type | PCIe or SXM5 (select at launch) |
|---|---|
| GPU memory | 80 GB |
| BF16 compute | 989 TFLOPS (dense) |
| Memory bandwidth | 3.35 TB/s |
| GPU-to-GPU (same node) | NVLink 900 GB/s |
| Network | 400 Gbps · 3.2 Tbps IB (Enterprise) |
| GPUs per node | Up to 8 |
| vCPU per GPU | 26 |
| System memory per GPU | 225 GB |
| Local NVMe per GPU | 3.5 TB |
PCIe is available for flexible single-GPU deployments. SXM5 is available from 1–8 GPUs and is the default for multi-GPU training; up to 8 GPUs share one node over NVLink. Larger jobs span nodes over InfiniBand (Enterprise only). CPU, memory, disk and fabric vary by upstream provider; the figures above are the current per-GPU allocation shown by this app.
Live H100 availability
Checking capacity…
What an H100 run costs
Fine-tune
LoRA on an 8B model, 50k samples, 1 epoch
1h 30m · $3.63
Image gen
1,000 SDXL images at 1024×1024
24 min · $0.97
Inference
1M output tokens from a 70B model
2h · $4.84
Estimates from blended public benchmarks, Indie pricing, GPU time only.
Included with every H100
- Pre-cached weights for popular models — zero cold start.
- Idle auto-stop after 15 minutes at 0% GPU (configurable).
- Verified-boot billing with instant refund if the instance fails its health check.
- Zero egress fees and multi-provider failover.
- Budget caps with alerts at 50/80/100% and auto-stop at 100%.
Planning data transfer? Read the AI training bandwidth guide.