About the role
Crusoe is on a mission to accelerate the abundance of energy and intelligence . As the only vertically integrated AI infrastructure company built from the ground up, we own and operate each layer of the stack — from electrons to tokens — to power the world's most ambitious AI workloads. When you join Crusoe, you join a team that is building the future, faster. We're in the midst of the greatest industrial revolution of our time. The demand for AI compute is boundless, and power is a bottleneck. We're solving that — with an energy-first approach that makes AI infrastructure better for the world and faster for the people innovating with AI. We're looking for problem-solving, opportunity-finding teammates with a sense of urgency, who believe in the scale of our ambition and thrive on a path not fully paved — people who want to grow their careers alongside a team of experts across energy, manufacturing, data center construction, and cloud services. If you want to do the most meaningful work of your career, help our customers and partners advance their AI strategies, and be part of a high-performing team that believes in each other, come build with us at Crusoe. Conductor is the platform we're building: a self-driving fleet. Most infrastructure teams watch dashboards and replace dead nodes after they fail. We're building something different — a control plane that predicts, decides, and acts on its own, keeping tens of thousands of accelerators doing useful work at the highest goodput in the industry while optimizing power and cost in real time. The mandate for this role is to make the fleet self-driving. This is a Principal Engineer role reporting into Cloud Availability, working directly on one of the highest-priority technical charters in the org. Your Charter — The Problems You'll Own Unified observability plane. One pane that correlates GPU, networking (InfiniBand/RoCE), storage, orchestration, and workload signals — so any engineer can diagnose and recover jobs fast. The fleet as one logical computer. Tens of thousands of accelerators across sites behaving as a single programmable system: one health model, one scheduler, one source of truth. Closed-loop autonomy, not alerting. Diagnose, decide, and remediate with no human in the loop — drain, checkpoint, replace, and resume a live job automatically. The hard part: earning enough trust to act on a running job. Goodput as an objective function. Don't just measure useful compute — continuously maximize it, trading scheduling, placement, and maintenance decisions against it as the north-star metric. Predict failures hours ahead, not seconds after. Forecast GPU, NVLink, optics, and thermal degradation before it stalls a job, and pre-emptively migrate work. An ML problem on noisy hardware telemetry at fleet scale. Straggler & silent-failure detection. One slow GPU stalls an entire distributed job — isolate the exact rank, GPU, and node from collective-operation signals and surface root cause fast. Energy-aware compute. Because we own the power stack, schedule, throttle, and place workloads against real-time energy availability, cost, and thermal headroom. The problem you can work on here and nowhere else. A digital twin of the fleet. Simulate failures, scheduling policies, and remediation logic before they touch production — the foundation that makes real autonomy safe. Agentic operations. An agent that doesn't just answer "why is this job slow" but proposes and executes the fix, with guardrails and a full audit trail. Zero-trust, fully auditable multi-tenancy. Every action — human or autonomous — identity-scoped, policy-checked, and streamed to customer audit systems in real time. Self-qualifying hardware. New and repaired nodes prove themselves through automated burn-in before taking customer load. Fleet growth becomes a software-gated pipeline. Why This Role Is Different Scale almost no one gets to work on — distributed systems and control theory across tens of thousands of accelerators. A problem unique to Crusoe — we own the stack from energy to compute, so the platform can treat power as a control input, not a constraint. Greenfield authority — you define the operating standard for how we run AI at scale. You don't inherit one. What You'll Bring 10+ years building infrastructure-layer systems at scale — fleet management, distributed control planes, scheduler internals, or hardware lifecycle automation. This is a systems-builder role, not a consumer of managed cloud services. Deep experience with distributed systems design: consensus, state reconciliation, closed-loop automation, and systems that make autonomous decisions against live production infrastructure. Hands-on fluency with GPU/HPC infrastructure — GPU health telemetry, NVLink/InfiniBand/RoCE fabrics, thermal and power behavior at the hardware level. Track record of designing and shipping large-scale observability or telemetry platforms that correlate signals across compute, network, and storage layers. Comfort operating in ambiguity and defining the architecture and standards for a system that doesn't exist yet — this is a 0→1 charter, not a maintenance role. Strong software engineering fundamentals in at least one systems language (Go, Rust, C++, or similar) and the judgment to know when to build vs. adopt existing tooling. Experience applying ML/statistical methods to noisy operational telemetry (failure prediction, anomaly detection) is a strong plus. Prior exposure to zero-trust or policy-based multi-tenancy architectures is a plus but not required.
