Cumulus Labs builds the software that turns raw GPU capacity into fast, cheap, production AI. We're a YC company (W26) taking on some of the hardest systems problems in inference.

We ship three products:

• Ion: a proprietary C++ inference engine with hand-written CUDA kernels, built for unified-memory NVIDIA Grace hardware (GH200, GB200, GB300). Sub-second model hot-swapping, custom attention scheduling, and split inference for physical AI. It outperforms leading inference providers on vision benchmarks, with numbers on our blog.
• Paladin: a GPU orchestrator that schedules training and inference across heterogeneous, multi-cloud fleets, allocates GPUs fractionally, and live-migrates running jobs between GPUs with no downtime.
• Talos: an agent platform that builds, runs, monitors, and continuously optimizes production agents on top of the stack.

Companies already run inference and agents on Cumulus in production, and we're backed by YC and many other investors.

We're small and early, so you own whole systems end to end rather than tickets. If you want to write kernels, build distributed schedulers, and squeeze every token per second out of the newest NVIDIA silicon alongside people who care about doing it right, this is the place.

We're based in San Francisco and work in person.