Nodus

Developers and researchers who train models, run coding agents, or do reinforcement learning need GPU compute that fits their specific code, container, and resource requirements. Finding compatible capacity, executing the job, tracking its cost, and recovering progress after an interruption are separate problems that normally fall on the user to solve themselves. Without this, a failed or preempted run can mean lost checkpoints and wasted compute spend. Coding agents like Claude Code, Codex, or Cursor also need a way to request GPU access, estimate cost, and retrieve results without the developer manually provisioning machines.

Nodus is a compute platform built around a bring-your-own-code workflow: users supply a container image, data, and resource requirements, and Nodus matches the job to compatible GPU or CPU capacity. It tracks cumulative workload cost across setup, runtime, and recovery against a spending limit, and if a run is interrupted it restores saved checkpoint files on replacement compute so the job can resume. The platform covers four workload types: training (pre-training, post-training, LoRA fine-tuning), agents (sandboxes, evals, long-running tasks with time and spending limits), reinforcement learning (environments, training, evaluation), and an OpenAI-compatible inference API (in beta). It also offers persistent project workspaces accessible via VS Code, Jupyter, or SSH, and integrates directly with coding agents so they can request and manage compute runs. Nodus separately lists reserved GPU capacity, an agentic research effort on lowering compute costs, and an early-stage concept for compute/inference financial infrastructure.