Software Engineer, Compute (GPU)
Fluidstack · San Francisco, CA
About this role
**ABOUT FLUIDSTACK** We exist to make humanity more free. Powerful AI will be the biggest lever for human choice—but only if models are aligned with what humanity actually wants. Whoever deploys frontier compute infrastructure fastest will decide whether AI expands human freedom or shrinks it. Fluidstack is singularly focused on delivering **10 to 100s of GWs of compute faster than anyone else**, rethinking every layer of the stack. We acquire power, design and build data centers, and operate them—with teams spanning hardware and software. --- **HOW WE OPERATE** - **Be a barrel:** Full autonomy. Own things end to end; take on scope without being asked. - **Insane urgency:** We drive everything forward as fast as possible. - **Reason from first principles:** Challenge every assumption; the best idea wins. - **Love of the game:** Long hours at high intensity to push the frontier forward. - **Build something that actually matters:** Spend your time on work that matters to the world. --- **THE PRODUCTION ENGINEERING TEAM** Examples of key problems the team is working on: - **Build the repair pipeline** for a fleet of **10s to 100s of GWs**: a GPU failure isn’t a ticket—it’s a throughput problem. Automate chip fault detection → triage → RMA → return to service. - **Qualify every new GPU generation** inside a **6-month build window**: burn-in, performance baselining, and NPI execution. Define “production-ready” before sites go live. - **Migrate live compute at construction speed:** convert clusters across production sites and make Kubernetes-orchestrated bare metal sustainable at multiple GW annually. - **See and own the entire fleet in real time** at any scale: build the observability + orchestration layer to make hyperscale AI compute operable. --- **ROLE SCOPE** - **Own compute fleet health end to end:** build metrics pipelines, alerting, and a unified health view across **Kubernetes-orchestrated workloads and bare metal**. - **Turn deployment/repair into a pipeline:** automate compute failures from detection → triage → parts management → return to service (no one-off scripts). - **Design and expand the GPU qualification platform:** burn-in, performance baselining, and NPI execution for every new GPU generation. - **Own Redfish and BMC tooling:** firmware-level telemetry, log collection at fleet scale, and low-level access layers used by repair automation and health tooling. - **Own reliability, scalability, and operations** of the compute fleet at scale—enabled by aggressive automation, tooling, and incident discipline. --- **WHAT WE’RE LOOKING FOR** This is a starting point—if you don’t fit perfectly, tell us where you would. - Treat **toil as a bug**; manual steps belong on a backlog, not in your workflow. - Strong **hardware instincts**; comfortable reasoning about failure modes at firmware/silicon level. - Move toward **ambiguity**; build the map and explain it to others. - Learn quickly; reach real competence in unfamiliar domains. - **Pager-ready:** run incidents, write postmortems, fix systemic causes, and move on. - Fluent with **AI tooling** (LLM APIs, MCP servers, agentic frameworks; use Claude Code/Cursor or similar daily). - Shipped **production automation** that others depend on; comfortable using AI coding tools in any language. - **Bonus:** hardware lifecycle management and RMA automation; BMC/Redfish/IPMI tooling; GPU qualification/burn-in frameworks; workflow/orchestration engines (Temporal, Cadence); metrics/ale
Listing freshness
CronJobs last confirmed this listing 1h ago. If its source stops confirming the opening for seven days, this page is removed from active inventory.