Resume from saved checkpoints.
More AI from
the GPUs you
already have.
Train models. Run agents. Serve inference.
Built by engineers previously at
Used by
Stanford
VoiceStudio
One place to run your AI workloads.
Bring your code for training, agents, or inference. Nodus places the work on compatible GPUs and handles execution.
Run code. Manage files. Isolate execution.
Send requests. Stream responses.
Run your work. Share your idle compute.
BYOCompute runs your jobs on your own machines first. Liquid Capacity lets you lend eligible idle compute to other teams and earn Nodus credits.
Your workloads first. Idle capacity, shared.
Your workloadsLiquid Capacity jobs
Reserved Compute
Explore reserved compute
Capacity for the long run.
Dedicated GPUs and clusters for training and inference. Tell us what you need. We’ll help plan your reservation.
More work from the same hardware.
From our published research on GPU systems, runtimes, and scheduling.
01 / RUNTIMES
Read Nodus Research ↗Cold start in seconds.
“A warmed engine returned from a cold host in 7.158 s, 7.66× faster than the best stock boot.”
Measured on one A100 · 5 cold trials · reproduced at 7.65× by an independent verifier.
Read report 025 ↗ 02 / TRAININGEight RL jobs. One GPU.
“Eight customers' RL post-training jobs shared one GPU at 2.137× the steps per GPU-hour.”
Measured on one A100 · learning unchanged · 2.244× on fresh seeds.
Read report 024 ↗ 03 / GPU SYSTEMSFewer GPUs. Same SLO.
“Thirty-two tenants ran on 4 H100s instead of 32 dedicated H100s.”
Single-node, replayed traffic · 3 seeds · ≥98.5% SLO per class.
Read report 002 ↗