Furnace
NVIDIA GB300 GPU rental by the hour. 288GB VRAM for your heaviest workloads.
$150/hour • No minimums • Auto-scaling
GB300 B300 Specifications
Built for ML workloads
Raw Compute Power
288GB VRAM per GPU. Run 70B parameter models entirely in memory. No sharding headaches.
Auto-Scaling
Start with 1 GPU. Scale to 1000+ automatically. Perfect for batch jobs or sudden traffic spikes.
Persistent Storage
Your datasets, models, and checkpoints stay warm between sessions. Pick up where you left off.
Pre-Configured Stacks
PyTorch, JAX, vLLM, TGI pre-installed. SSH in and start training. No setup scripts.
More VRAM, competitive pricing
Compare on a per-GB basis and Furnace delivers 3.6x more VRAM per dollar than standard H100 rentals.
Cloud pricing as of April 2026. Prices subject to change.
Perfect for
How Furnace works
Request Access
Sign up and request Furnace access. Approved within 24 hours.
SSH In
Get SSH credentials. PyTorch, JAX, CUDA all pre-installed.
Start Training
Run your workloads. Billed hourly. Scale up or down anytime.
Ready to fire up the Furnace?
Get access to GB300 GPUs for training, inference, and experimentation.