GPUs for agents, loved by humans.
GPU and CPU sessions that keep your code and files between runs. Stop compute when you're done. Resume when you need it.
Live rates
T4$0.47/GPU-hL4$0.64/GPU-hA10G$0.88/GPU-hL40S$1.56/GPU-hRTX PRO 6000$2.42/GPU-hA100$2.00/GPU-hH100$3.16/GPU-hH200$3.63/GPU-hB200$5.62/GPU-h
Products
One supercomputer. Every kind of work.
For developersGive your agents GPU boxes that bill only while they run.Stop a box and its files stay while billing stops. One GPU to a fleet in one command.For scientistsAutoresearch: an experiment loop that runs while you sleep.Choose the metric, the one file it may change and a time budget. Every result is charted.For businessesPreviewAsynchronous inference, fine-tuning and RL, priced by deadline.Pick a deadline of one, six or twenty-four hours. The slower lane costs less, and late is free.For teamsForward deployments and custom training, with our engineers in the room.We embed with your team, set up dedicated GPUs and run custom training end to end.
Examples
Serious work, already running on supercomputer.
8× H100 · language modelsThe nanoGPT speedrunTrain GPT-2 with modded-nanogpt to a 3.28 validation loss.Result3.276 val loss after 43 s of trainingOpen example →1× H100 · biologyDesign a GLP-1 receptor binderBoltzGen designs peptides for the groove the natural hormone binds.Result100 designs at about 10.6 s eachOpen example →1× H100 · roboticsFine-tune a robot policyVLA-JEPA fine-tuned on the LIBERO object suite.Result0 → 20 of 20 episodes, about $10Open example →1× L4 · reinforcement learningHigh-throughput RLTrain a Breakout agent with PufferLib's GPU-vectorized environments.Open example →1× L4 · simulationLegged-robot rollouts on the GPUMuJoCo MJX steps a batch of Ant robots in parallel on one GPU.Open example →1× H100 · generative 3DOne image to a textured 3D meshTRELLIS.2 turns a single picture into a ready-to-use mesh.Open example →1× L4 · 3D captureGaussian splatting on one L4nerfstudio splatfacto on a real capture, rendered as a fly-through.ResultAbout 460k Gaussians in 15–25 minOpen example →8× L4 · JAXMulti-process JAX on eight GPUsOne JAX process per GPU, joined with jax.distributed.Open example →1× L4 · inferenceChat with Qwen3.5-9BThe largest dense Qwen3.5 that fits one L4, its weights kept on a volume.Result18.0 of 24 GB at peak, about 11 tokens/sOpen example →
Get started