NVIDIA is a pioneer in accelerated computing, known for inventing the GPU and driving breakthroughs in gaming, computer graphics, high-performance computing, and artificial intelligence. Our technology powers everything from generative AI to autonomous systems, and we continue to shape the future of computing through innovation and collaboration. Within this mission, our team, Managed AI Research Superclusters (MARS), builds and scales the infrastructure, platforms, and tools that enable researchers and engineers to develop the next generation of AI/ML systems. By joining us, you'll help build solutions that power some of the planet's most advanced computing workloads.
The MARS AI Data team owns how NVIDIA's AI research uses data: many petabytes of training data, checkpoints and datasets across multiple clouds and partners. We build the layer between storage and the workload. That covers data placement across file systems, object storage, and local NVMe. It also includes streaming loaders and caches that supply tens of thousands of GPU cores. Efficient checkpoint management. Quotas remain fair throughout numerous research groups. This is a builder's role on a team that is redefining what it does. You will help design MARS' data platform on Kubernetes from the ground up, working directly with the researchers training NVIDIA's own models. The work shows up as faster training steps, shorter restarts and fewer runs lost to complete storage volumes.
What you'll be doing
Design and build the AI data layer for a multi-cloud GPU fleet: how datasets, checkpoints and scratch data are placed, cached, tiered and moved between different storage backends.
Build high-throughput data paths for training and RL. That means streaming loaders from object storage, distributed caches, and checkpoint save and restore at thousands of GPUs, including peer-to-peer and local-first restore.
Make data a first-class part of the compute platform. Build Kubernetes-native APIs, operators and storage integrations so a job declares the data it needs, and the platform mounts, caches and accounts for it, on Slurm and on Kubernetes.
Own quotas, capacity and data lifecycle across hundreds of research teams: usage attribution, quota enforcement, retention, clean-up of stale data, and forecasting capacity before it runs out.
Measure how workloads use data. Build I/O telemetry and access analytics that show hot and cold data, metadata pressure and I/O stalls, and use them to drive placement and fixes.
Own reliability for what you build: on-call, incident analysis, and changes that stop whole classes of data incidents, such as full file systems, runaway metadata load and lost checkpoints.
What we need to see
BS or MS in Computer Science, Computer Engineering, Electrical Engineering or a related field, or equivalent experience.
5+ years building production systems or data-infrastructure software.
Deep experience with at least one of the following: parallel or distributed file systems (Lustre, WEKA, VAST, GPFS, Ceph or similar), object storage at scale (S3 or compatible), or large-scale data pipelines and data lakes.
A clear grasp of I/O performance: throughput, latency, metadata cost, caching and how access patterns interact with each.
Strong programming in Go, Rust, C or C++, plus Python for tooling and analysis.
A track record of owning a service in production and making it measurably better.
Clear written and spoken communication with engineers and researchers.
Ways to stand out from the crowd
Experience feeding large-scale AI training: data loaders, streaming datasets, and storage behavior at thousands of GPUs.
Experience with checkpointing at scale: asynchronous or distributed checkpoints, tiered checkpoint storage, fast restart after failure.
Built or ran a distributed cache or AI-oriented storage system, or used node-local NVMe as a managed tier.
Built a quota, accounting or capacity system that many teams depended on.
Knowledge of RL and agentic workloads and…