Harbor AI

What it does

We’re building Harbor to be the operating layer that keeps GPU fleets productive. A fleet can be completely online and still be wasting compute. Training takes longer, or an inference service handles fewer requests than it should. We’re starting with software for AI teams running dedicated GPUs that figures out why workloads are failing or slowing down, makes supported fixes with approval, and checks whether performance recovers. We’re building a context engine that keeps track of the setup, what changed and what worked before, and developing models that use that understanding to predict which changes will help. Long term, a team should be able to tell Harbor what work needs to get done and what performance it needs. Harbor would figure out how to run that work across compute they own or rent, balancing speed, reliability and cost. As workloads and requirements change, Harbor would keep learning and adjusting how the fleet runs.