About the job
We are building one of the world’s largest AI supercomputers from the ground up. As part of the Compute Infrastructure team, you will own both the raw GPU supercomputer and the platform layer that runs on top of it. You will work across the full stack — from low-level GPU kernel optimizations and Linux kernel internals to massive-scale orchestration and virtualization — to make training and inference at xAI as fast, reliable, and scalable as possible. This is a broad, high-impact role that combines hardcore supercompute and compute infrastructure work. Your contributions will directly accelerate Grok’s training speed and overall AI progress.
Responsibilities
Design, build, and optimize massive GPU clusters for extreme-scale training and inference workloads
Develop and tune low-level CUDA kernels (GeMM, Attention, etc.), using CUTLASS, Tensor Cores, and Nsight for maximum performance
Work on Linux kernel internals, scheduling, memory management, and resource isolation at cluster scale
Build custom container orchestration, virtualization layers (KVM, Firecracker, etc.), and distributed systems that go beyond standard Kubernetes
Profile, debug, and eliminate bottlenecks across GPU memory hierarchy, networking fabric, filesystems, and multi-GPU operations
Create and maintain infrastructure-as-code, automation, and tools that keep the entire supercomputer reliable and efficient
Collaborate closely with AI research teams to deliver production-grade performance and scalability
Qualifications
Minimum
No minimum qualifications listed.
Preferred
Deep low-level systems programming (C/C++ or Rust)
Experience building and operating high performance exabyte scale storage systems
Strong experience with large-scale GPU clusters or distributed compute infrastructure at production scale
Hands-on work with GPU kernel optimization (CUTLASS, custom kernels, Nsight profiling)
Experience with Linux kernel internals, scheduling, virtualization, or large-scale orchestration
Track record of building or running high-performance infrastructure for AI workloads (training or inference platforms)
Ability to reason from first principles and optimize for both memory-bound and compute-bound scenarios