[ hardware — blackwell · per second ]
Managed presets on B200/B300/GB300 — billed per second, autoscaled to zero — or full GPU VMs on one platform with Spot discounts.
Wide table — swipe sideways for the $/hr and $/sec columns →
| Preset | VRAM pool | Precision | Best for | Position | $/hr | $/sec | |
|---|---|---|---|---|---|---|---|
| 1× NVIDIA T4 16GB | 1× NVIDIA T4 16GB | FP16 | Cheapest CUDA — experiments, small models | Priced per GPU-hour — see the launch wizard for live rates | $0.44 | $0.0001 | ▾ |
| 1× NVIDIA L4 24GB | 1× NVIDIA L4 24GB | FP8 | Best price/perf up to ~13B + video | Priced per GPU-hour — see the launch wizard for live rates | $0.87 | $0.0002 | ▾ |
| 2× NVIDIA L4 24GB | 2× NVIDIA L4 24GB | FP8 | g2-standard-24 — two-worker inference | Priced per GPU-hour — see the launch wizard for live rates | $1.8 | $0.0005 | ▾ |
| 4× NVIDIA L4 24GB | 4× NVIDIA L4 24GB | FP8 | g2-standard-48 — multi-worker farm on one VM | Priced per GPU-hour — see the launch wizard for live rates | $3.6 | $0.0010 | ▾ |
| 8× NVIDIA L4 24GB | 8× NVIDIA L4 24GB | FP8 | g2-standard-96 — full-node L4 | Priced per GPU-hour — see the launch wizard for live rates | $7.2 | $0.0020 | ▾ |
| 1× NVIDIA RTX PRO 6000 Blackwell (1/8 GPU · 12GB) | 1× NVIDIA RTX PRO 6000 Blackwell (1/8 GPU · 12GB) | FP8 | g4-standard-6 — Blackwell slice for light jobs | Priced per GPU-hour — see the launch wizard for live rates | $0.3 | $0.0001 | ▾ |
| 1× NVIDIA RTX PRO 6000 Blackwell (1/4 GPU · 24GB) | 1× NVIDIA RTX PRO 6000 Blackwell (1/4 GPU · 24GB) | FP8 | g4-standard-12 — quarter GPU | Priced per GPU-hour — see the launch wizard for live rates | $0.6 | $0.0002 | ▾ |
| 1× NVIDIA RTX PRO 6000 Blackwell (1/2 GPU · 48GB) | 1× NVIDIA RTX PRO 6000 Blackwell (1/2 GPU · 48GB) | FP8 | g4-standard-24 — half GPU | Priced per GPU-hour — see the launch wizard for live rates | $1.2 | $0.0003 | ▾ |
| 1× NVIDIA RTX PRO 6000 Blackwell 96GB | 1× NVIDIA RTX PRO 6000 Blackwell 96GB | FP8 | g4-standard-48 — try any app from $2.40/hr (park/wake receipted D27) | Priced per GPU-hour — see the launch wizard for live rates | $2.4 | $0.0007 | ▾ |
| 2× NVIDIA RTX PRO 6000 Blackwell 96GB | 2× NVIDIA RTX PRO 6000 Blackwell 96GB | FP8 | g4-standard-96 — two Blackwell GPUs | Priced per GPU-hour — see the launch wizard for live rates | $4.8 | $0.0013 | ▾ |
| 4× NVIDIA RTX PRO 6000 Blackwell 96GB | 4× NVIDIA RTX PRO 6000 Blackwell 96GB | FP8 | g4-standard-192 — four GPUs | Priced per GPU-hour — see the launch wizard for live rates | $9.6 | $0.0027 | ▾ |
| 8× NVIDIA RTX PRO 6000 Blackwell 96GB | 8× NVIDIA RTX PRO 6000 Blackwell 96GB | FP8 | g4-standard-384 — full-node Blackwell | Priced per GPU-hour — see the launch wizard for live rates | $19.2 | $0.0053 | ▾ |
| 8× NVIDIA B300 288GBpopular | 288GB ×8 = 2.3TB | FP8 | Frontier dense models, 1M-ctx serving | vs H100 8×: ~2× tokens/$ on MoE | $96 | $0.0267 | ▾ |
| 8× NVIDIA B200 180GBbest fitpopular | 180GB ×8 = 1.4TB | NVFP4 | Large MoE throughput kings | vs A100 8×: ~3× throughput | $80 | $0.0222 | ▾ |
Spec sheet
NVFP4 precision NVFP4 — best tokens/$ for large MoE models; 4-bit weights, frontier quality. Best tokens/$ for large MoE models. | |||||||
| 4× NVIDIA B300 288GBpopular | 288GB ×4 = 1.1TB | NVFP4 | Smallest frontier-capable footprint | vs H200 4×: lower $/hr, more VRAM | $48 | $0.0133 | ▾ |
| 4× NVIDIA B200 180GBpopular | 180GB ×4 = 720GB | NVFP4 | ≤70B dense models, evals | vs L40S 4×: ~2.5× inference speed | $40 | $0.0111 | ▾ |
| 16× NVIDIA GB300 | 288GB ×16 = 4.6TB | FP8 | 1T+ parameter models | The frontier ceiling | $184 | $0.0511 | ▾ |
Click any row for the full spec sheet · every preset deploys in one click — open the deploy wizard →
Per-second billing — scale to zero and you pay $0 for idle time.
$28,800
12 h/day × 30 days · 50% uptime
Per day
$960
Per week
$6,720
Per GPU of the node
$3,600
Idle (scaled to zero)
$0.00
$80/hr · $0.0222/sec · autoscale + scale-to-zero included
Deploy this preset →1× T4
Scale-1 — 1× T4
Cheap inference, transcoding, and CUDA experiments
1× L4popular
Scale-1 — 1× L4
Best price/perf for LLM inference up to ~13B and video
4× L4
Scale-4 — 4× L4
Multi-worker inference farms on one VM
1× RTX PRO 6000
Scale-1 — 1× RTX PRO 6000
96GB Blackwell workstation-class VRAM, single GPU
1× A100
Scale-1 — 1× A100 80GB
Ampere workhorse for training and dense inference
8× A100
Scale-8 — 8× A100 80GB
Full-node A100 training
8× H100
Scale-8 — 8× H100 80GB
Hopper full node — frontier fine-tunes
Every GPU card deep-links into the launch wizard with the SKU pre-selected. Choose Spot vs On-Demand, Standard vs Premium network tier (billed egress), boot-disk size/type with IOPS, and local NVMe in the wizard. Open the full wizard →
Boot-disk IOPS control, local NVMe scratch, CUDA/PyTorch images — launch a VM instance →
Deployments bill per second with scale-to-zero. Start with $5 free.