Rent a GPU for the afternoon
No card under the desk? Rent one by the hour, pair it like a machine of yours, serve the model, leave when you are done. Nothing about the box is special: it is a machine that happens to send an invoice.
What you need
- An account at a GPU cloud — Nebius, Vast.ai, RunPod, Lambda. An H100 (80 GB) for the big models, an L40S (48 GB) or a 5090 (32 GB) for a 27B.
- An Ubuntu image with the NVIDIA driver — every GPU cloud has one.
nvidia-smishows the card. - A workspace you can add things to.
Rent it and join it
Create the VM, open its console or SSH in, and run the same command as on any machine. The tile says gpu and which one.
$ curl -fsSL https://daslab.run/node | sh installed daslab 1.1.63 for x86_64 service enabled and started pairing code T6XK-3BVQ open daslab.run/pair waiting to be paired…
Type the code at daslab.run/pair on your phone, pick the workspace, and give it a name you will recognise on the bill.
paired with workspace labhaus as h100-afternoon
publishing: x86_64 · 26 cores · 10.0.0.0/24 · gpu nvidia
ready
Serve the model
$ pip install vllm $ vllm serve openai/gpt-oss-120b --max-model-len 131072 INFO Uvicorn running on http://0.0.0.0:8000
Scenes you pin to the box reach it at localhost:8000 today. next The tile lists the model and it appears in the picker of every scene in the workspace as h100-afternoon / GPT-OSS-120B; pick it and the agent runs on it, at the box's hourly rate and nothing per token.
One prompt instead of all of that next
With a Nebius account connected to the workspace, the agent rents the box, pairs it without the phone, serves the model and puts it in the picker:
Leave
On the box, daslab node leave — or delete the tile — then destroy the VM at the provider. Scenes pinned to it run on our machines from their next run; the model leaves the picker.