GPU workloads

Coming soon. GPUs aren't deployable yet; join the waitlist on the pricing page or with POST /v1/waitlist. When they launch:

Pick a GPU plan: gpu-l4, gpu-rtx4090, gpu-l40, gpu-a100 or gpu-h100 (billed per second; see pricing). Each GPU instance gets a dedicated node with the NVIDIA container runtime, and your container sees the GPU via CUDA (--gpus).

aloof deploy llm --image vllm/vllm-openai:latest --plan gpu-l4 --port 8000 -e MODEL=Qwen/Qwen2.5-7B-Instruct
aloof autoscale llm --min 0 --max 2 --idle 600

Planned capacity: Hyperstack GPU VMs (A100, H100, L40; Canada/Europe/US) and Runpod pods (L4, RTX 4090; many regions), tried in order of cost and availability. GPU nodes are released after 10 idle minutes. A cold start on GPU means provisioning a node, which takes a few minutes, so keep min 1 for latency-sensitive inference.