Instant provisioning · Zero configuration · Per-minute billing
Production GPU compute, ready in seconds.
Deploy NVIDIA H100, H200 and B200 instances with your frameworks and models preloaded. Transparent pricing, no egress fees, and built-in cost controls, typically at less than half the cost of major cloud providers.
Saved templates · one-click redeploy
Save instance type, workload, cached weights and cost controls as a reusable template.
Deployment history
Every launch attempt lands here with its applied configuration and outcome.
Quick-start configs
Region · real-time capacity
Availability and provisioning times below reflect the selected region and GPU count.
Tip: Pick the region closest to you or your users. Price is the same everywhere. A green dot means machines are free right now.
Service health
US East: All systems normal
The platform is healthy here. If your run fails, check its logs first. The cause is most likely in your code or data.
Recent incidents · US East
No incidents in the last 7 days.
Instance type · throughput, ETA & capacity
Estimate for:
LoRA on an 8B model, 50k samples, 1 epoch · relative throughput (H100 = 100)
Tip: Fine-tune: Teach an existing AI model your own data or style. For this, we recommend RTX 4090 (best value). Choose B200 if speed matters most.
Memory
Prototyping & SDXLBest value
Throughput score
30· ~5h · ~$2.20
142 free · ready ~20s
$0.44/hr
≈ $24.20/mo total
Fine-tuning & mid-size inference
Throughput score
55· ~2h 44m · ~$4.20
88 free · ready ~30s
$1.54/hr
$4.10 big cloud
≈ $68.20/mo total
Training & LLM serving
Throughput score
100· ~1h 30m · ~$3.63
64 free · ready ~45s
$2.42/hr
$6.88 big cloud
≈ $103.40/mo total
70B+ inference, long context
Throughput score
125· ~1h 12m · ~$4.22
18 free · ready ~1 min
$3.52/hr
$7.90 big cloud
≈ $147.40/mo total
Frontier-scale trainingFastest
Throughput score
220· ~41 min · ~$3.38
6 free · ready ~2 min
$4.95/hr
$10.00 big cloud
≈ $204.60/mo total
Spec comparison · networking, runtime, limits & total cost
Job ETA/cost reflect the selected workload (Fine-tune). OOM = model doesn't fit in VRAM. Monthly total = current usage settings on that GPU.
Pre-cached weights · zero cold start
Weights staged on local NVMe at no extra cost. No Hugging Face pulls, no cold start.
Tip: These are AI models loaded before your machine starts, so you don't pay while waiting for downloads. Pick the one your project uses. Not sure? PyTorch alone is fine for most code.
Runtime & networking
H100 type
SXM5 uses NVLink for scaled training. PCIe is offered as a single-GPU instance.
Runtime image
Default. Broadest library support
Network fabric
Standard. Fine for single-node jobs
Tip: Not sure? Keep the defaults (CUDA 12.4 and Ethernet). We'll flag anything that won't work with your GPU before you pay.
Compute & storage
Tip: Start with 1 GPU. Hours are a monthly budget estimate. You only pay for minutes actually used. NVMe volume is your disk: 100 GB fits a mid-size model plus a dataset. The deploy check will tell you if it's too small. Bandwidth is always free.
Cost & reliability controls
Tip: Keep both on. Idle auto-stop ends billing if you forget to shut down, and failover keeps your job running if a provider has problems.
Built to remove the friction developers report most.
01
Pre-warmed model cache
Llama, DeepSeek, ComfyUI and PyTorch weights are staged on local NVMe, so you never spend paid GPU time on 30-minute model downloads.
02
Idle protection
Instances pause automatically after a configurable period at 0% GPU use (15 minutes by default), so a forgotten session never runs overnight.
03
Verified-boot billing
Billing begins only after SSH and Jupyter pass health checks. Any instance that fails is credited back automatically.
04
No egress fees
Bandwidth is included with clearly published limits. Usage beyond them is rate-limited, never billed.
05
Multi-provider resilience
Capacity is sourced across independent providers, so a single-vendor outage doesn't interrupt your work.