Stealthium

Stealthium builds runtime observability and security software for GPUs and AI workloads. Its platform serves infrastructure and security teams operating training, inference and shared GPU fleets. It connects device telemetry with the processes, containers and pods responsible for it, helping operators locate performance bottlenecks, investigate unauthorized workloads and maintain a record of what ran where. Its Hyperprints turn low-level metrics and kernel traces into workload-level insights, bringing performance analysis and security investigation into the same platform.

The company builds its monitoring software close to the hardware. Stealthium GPU Monitor queries NVIDIA drivers directly and handles monitoring calls inside the platform, with NVIDIA’s Management Library, NVML, retained where driver coverage is incomplete. Owning this layer lets Stealthium develop its telemetry collection independently; its engineering account describes the tradeoffs between richer driver-level information and the cost of collecting it. The platform also instruments the NVIDIA driver and CUDA runtime to observe memory activity and kernel execution.

Stealthium’s GPU security research connects that instrumentation to concrete threats. Extending earlier academic work, its team demonstrated timing-based inference of model characteristics on H100 hardware without reading a neighboring workload’s prompts or answers. Its detection approach captures compiled kernels and examines their instructions for probing behavior. Through its Rafay integration, operators can pair GPU provisioning and tenant policies with observations of actual execution, then route detections into existing security operations systems.