Rent GPU Cloud Compute
Rent H100, H200, L40S, and RTX 6000 Ada GPUs by the hour. Prebuilt environments for ML training, LLM inference, and content creation. Launch in minutes with transparent per-hour billing.
╔═══════════════════════════════════════════╗ ║ ▓▓▓ 4U TRAINING POD ▓▓▓ ║ ╠═══════════════════════════════════════════╣ ║ ┌───────────────────────────────────┐ ║ ║ │ ◉ PWR │ ◉ NET │ ◉ SYS │ ▓▓▓▓▓▓ │ ║ ║ ├───────────────────────────────────┤ ║ ║ │ █GPU█ █GPU█ █GPU█ █GPU█ │ TEMP │ ║ ║ │ █GPU█ █GPU█ █GPU█ █GPU█ │ 72°C │ ║ ║ ├───────────────────────────────────┤ ║ ║ │ ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ │ ║ ║ └───────────────────────────────────┘ ║ ║ 8x H100 SXM │ NVLink │ 10kW │ 640GB ║ ╚═══════════════════════════════════════════╝
Fast Provisioning
Most instances reach SSH in a few minutes, with first-boot setup tracked separately from billing
Pay Per Hour
No commitments — the advertised rate is the billed rate, metered in ~5-minute windows
13 Launch Environments
PyTorch, TensorFlow, CUDA, Jupyter, and creative stacks — plus snapshot-and-resume
Root Access
Root SSH with your own key, plus an always-on browser terminal
Choose Your GPU
L40S 48GB
48GB GDDR6 VRAM
$2.04
/hour
Inference, image and video generation, and rendering
RTX 6000 Ada
48GB GDDR6 VRAM
$2.04
/hour
Inference, diffusion models, fine-tuning, and rendering
H100 80GB
80GB HBM3 VRAM
$4.41
/hour
Large model training, high-throughput inference, and LLM fine-tuning
H200 141GB
141GB HBM3e VRAM
$4.47
/hour
Long-context LLM inference, memory-bound training, and frontier model serving
8x H100 80GB
640GB HBM3 VRAM
$31.10
/hour
Distributed training, massive LLM workloads, multi-GPU inference
Launch Environments
Every instance boots Debian with the NVIDIA driver, Docker, and the NVIDIA container toolkit — docker run --gpus all works out of the box. The environment's container image is pulled by a background job and is ready a few minutes after the instance goes running. Nothing auto-starts: each running instance card shows copy-paste quickstart commands for its environment.
Frameworks & Training
PyTorch + CUDA
PyTorch training environment
PyTorch 2.2 + CUDA 12.1 (cuDNN 8) container, pulled automatically after boot
TensorFlow
TensorFlow GPU environment
TensorFlow 2.15 GPU container, pulled automatically after boot
NLP / LLM (PyTorch + CUDA)
PyTorch, Transformers, PEFT, LoRA, Accelerate
PyTorch 2.2 + CUDA 12.1 (cuDNN 8) container, pulled automatically after boot
Computer Vision (TensorFlow + CUDA)
TensorFlow, OpenCV, Albumentations, Timm
TensorFlow 2.15 GPU container, pulled automatically after boot
LLM Training
LoRA/QLoRA and distributed training stack
PyTorch 2.2 + CUDA 12.1 (cuDNN 8) container, pulled automatically after boot
General ML
Flexible CUDA base image for custom stacks
CUDA 12.1 + cuDNN 8 devel container (Ubuntu 22.04), pulled automatically after boot
Creative & Media
Dreams.fm Director
FLUX Schnell, Depth Anything V2/V3, Qwen2.5-VL, BiRefNet — all pre-loaded for real-time creative pipelines
Boots from a prebuilt snapshot — models preloaded under /opt/dreams-fm/models
Stable Diffusion
CUDA base ready for diffusion stacks (ComfyUI/A1111 install in minutes)
CUDA 12.1 + cuDNN 8 devel container (Ubuntu 22.04), pulled automatically after boot
Audio / ASR
Whisper, torchaudio, TTS, librosa stack
CUDA 12.1 + cuDNN 8 devel container (Ubuntu 22.04), pulled automatically after boot
Audio / Music Generation
CUDA base tuned for music-gen and audio diffusion stacks
CUDA 12.1 + cuDNN 8 devel container (Ubuntu 22.04), pulled automatically after boot
Video Processing
FFmpeg and CUDA-ready media pipeline
CUDA 12.1 + cuDNN 8 devel container (Ubuntu 22.04), pulled automatically after boot
Notebooks & IDEs
Jupyter Lab
Notebook-first environment with common ML libs
Jupyter TensorFlow notebook container — start it with one command from the instance card
VS Code Server
Remote VS Code server environment
code-server container — VS Code in your browser, started with one command
Multi-GPU Training
Need more power? The 8x H100 box is a one-click self-serve launch whenever capacity exists — no sales call required. NVLink-connected for maximum throughput.
- ✓8x H100 80GB box: $31.10/hr, launchable from the dashboard
- ✓640GB total HBM3 memory
- ✓NVLink-connected GPUs
╔═══════════════════════════════════════════╗ ║ ▓▓▓ 42U RACK ▓▓▓ ║ ╠═══════════════════════════════════════════╣ ║ ╔═════════════════════════════════════╗ ║ ║ ║ U42 │▓▓▓▓▓▓▓▓│ ◉◉◉◉◉◉◉◉ │ █ NVLINK ║ ║ ║ ║ U38 │▓▓▓▓▓▓▓▓│ ◉◉◉◉◉◉◉◉ │ █ NVLINK ║ ║ ║ ║ U34 │▓▓▓▓▓▓▓▓│ ◉◉◉◉◉◉◉◉ │ █ NVLINK ║ ║ ║ ║ U30 │▓▓▓▓▓▓▓▓│ ◉◉◉◉◉◉◉◉ │ █ NVLINK ║ ║ ║ ║ U26 │░░░░░░░░│ STORAGE │ ░ 100TB ║ ║ ║ ║ U22 │▒▒▒▒▒▒▒▒│ NETWORK │ ▒ 400Gb ║ ║ ║ ╚═════════════════════════════════════╝ ║ ║ PWR: 480V/3Φ │ COOL: Liquid │ 200kW ║ ╚═══════════════════════════════════════════╝
Popular Use Cases
LLM Fine-Tuning
Fine-tune Llama, Mistral, or Qwen on your data. LoRA and QLoRA for efficient training.
Recommended: H100 80GB for 70B models, RTX 6000 Ada for 7B-13B
Stable Diffusion
Run SDXL, FLUX, or train custom LoRAs. Batch image generation at scale.
Recommended: RTX 6000 Ada for single images, H100 for batch
Research
Experiment with new architectures. Full root access, and the Jupyter environment puts a notebook in your browser with one command.
Recommended: RTX 6000 Ada or H100 for maximum flexibility
Video AI
RunwayML alternatives, video generation, frame interpolation.
Recommended: H100 80GB for longer videos
Speech & Audio
Whisper transcription, TTS, voice cloning with XTTS.
Recommended: RTX 6000 Ada for most audio tasks
Production Inference
Deploy models for production. Install vLLM, TGI, or Triton on a root box.
Recommended: H100 for highest throughput
Frequently Asked Questions
How fast can I deploy a GPU instance?
Most GPU instances go from launch to running in a few minutes. First boot installs the NVIDIA driver and reboots once, and the environment's Docker image is pulled by a background job — it's typically ready a few minutes after the instance shows running. Billing only starts once the GPU runtime is validated as ready.
What GPU types are available for rent?
We offer NVIDIA L40S 48GB ($2.04/hr), RTX 6000 Ada (48GB, $2.04/hr), H100 80GB ($4.41/hr), H200 141GB ($4.47/hr), and 8x H100 640GB ($31.10/hr) — all self-serve, one-click launches whenever capacity exists. For reserved capacity or multi-node clusters, contact us.
How does billing work?
Pay-as-you-go from a prepaid credit balance, with no minimum commitment. The advertised hourly rate is exactly what the meter runs at. Usage is billed in roughly 5-minute windows, and only for time the GPU runtime is validated as ready — boot and driver setup aren't billed. You get a low-balance warning email; if your balance reaches $0 the instance is destroyed (disk erased), so top up or snapshot first. Optional auto-destroy deadlines are enforced server-side. Terminate anytime and only pay for what you used.
Can I use my own Docker images?
Yes. Every instance is a root box with Docker and the NVIDIA container toolkit preinstalled, so `docker run --gpus all <your-image>` works out of the box. Connect over root SSH with your own key — paste your public key in the launcher — or use the always-on browser web terminal, which needs no key setup at all.
Can I save my work and come back later?
Yes. Snapshot a running instance from the dashboard (up to 10 snapshots per account, snapshot storage currently free), then launch a new instance from that snapshot whenever you want. We also email you when an instance becomes ready, when your balance runs low, when an instance is destroyed, and when a credit purchase lands.
Do you support multi-GPU training?
Yes — the 8x H100 640GB box ($31.10/hr, NVLink-connected) is a one-click self-serve launch whenever capacity exists. For reserved capacity or multi-node clusters, contact us and a human will scope it with you.
Start Training Without Long Commitments
Create an account, choose your GPU, and launch. No contracts, no commitments, pay only for what you use.