We're building the VMware for GPUs. Most GPUs in the cloud sit idle a lot of the time because they're bolted to one machine over PCIe. We decouple them.
Our virtualization layer is a userspace shim, loaded through LD_PRELOAD, that intercepts CUDA calls and ships them over the network to a host with a physical GPU somewhere else in the data center. Workloads run unchanged. When a process goes idle, the GPU detaches and gets reassigned, and reattaches in tens of milliseconds when it's needed again. Think Ceph for GPUs: the GPU becomes a network resource a cluster-wide scheduler can pool and pack.
The hard part is making that fast. We're within ~10% of native on most AI workloads, and it lets us serve ~1.8x more users on the same fleet. We run it in production as our own GPU cloud and sell it to enterprises with underutilized fleets.
Team is from Citadel Securities, Aquatic, Old Mission, AWS, and Bain. Series A, $17.5M raised from Matrix and Y Combinator.
Open roles
- Software Engineer, C++ Systems ($215K-$325K + meaningful equity). Low-level, latency-sensitive work on the virtualization layer. Quant dev / HFT background preferred: - Software Engineer, Infrastructure ($150K-$250K + meaningful equity). Go, Kubernetes, reliable systems at scale: - Business Operations Lead ($150K-$225K + meaningful equity): If you're cracked at building systems software and want to do something more meaningful than creating liquidity, reach out.
Email me at [redacted] and mention HN. I read every email.