Crusoe Staff Software Engineer at Crusoe to design and build the Dynamic Power Management system for AI infrastructure. Responsible for power telemetry, control loop design, and ensuring precise, observable, and dependable energy management.
Responsibilities
Power, not silicon, is the binding constraint on AI infrastructure. Crusoe is vertically integrated from electrons to tokens: generation, batteries, controls, and the cloud orchestration layer. That puts us in a rare position to treat energy management as a software problem and to get more useful compute out of every megawatt we bring online.
Dynamic Power Management (DPM) is the system that makes this real. It measures, forecasts, and orchestrates energy use across our data centers, from the site down to the individual GPU, so power goes where it does the most work. That lets us put more compute behind every megawatt, keep sites running smoothly as electrical conditions change, and deliver consistent performance to every customer.
This is a mission-critical system and a ground-up build: it runs underneath every customer workload, so it has to be precise, observable, and dependable from day one. As one of the founding engineers, you’ll work with the principal architect to turn product requirements into a production architecture, ship the first version, and help shape the team that grows around it. The work runs end to end, from GPU energy management and hardware telemetry up to capacity planning and placement, and it ties directly to revenue.
What You’ll Be Working On
The DPM control loop. Design and build the system that allocates power across site, row, rack, node, and GPU based on real demand, applies those decisions through GPU and chassis power management, and adapts within seconds as conditions change.
Power telemetry. Own the pipeline for high-frequency power data from GPUs, NICs, chassis and BMCs (Redfish/IPMI), PDUs, and facility systems, with the latency, accuracy, and handling of stale or missing data that a control system requires.
Power modeling and forecasting. Work out how much power each site, row, and rack really needs, where capacity is sitting idle, and how sure we are of those numbers, using live telemetry and simulation instead of nameplate ratings.
Workload energy profiling. Understand how real training and inference workloads use energy (throughput, latency, performance per watt), find where power is spent without improving results, and turn that into efficiency policies.
Simulation and trace replay. Build tools that replay production power and workload traces so policy changes can be validated offline before they control production power.
Shadow-to-production rollout. Run DPM in shadow mode first, computing recommendations in production and comparing them to actual usage, and turn on automated management only once the system has earned trust.
Safety and failure design. Define fail-safe defaults, behaviour on stale telemetry, limits on how quickly settings change, audit trails, and verification that every adjustment had the intended effect.
Capacity decisions. Connect DPM to capacity planning and VM placement
Qualification
GPU energy managementIntegration with BMCs (Redfish)
Preferred
GPU energy management: NVIDIA DCGM/NVML, NVIDIA DPS, or AMD equivalents.
Integration with BMCs (Redfish), PDUs, building management (BMS), or electrical power monitoring (EPMS) systems.
Experience running large GPU or HPC fleets, or working on data center energy or capacity planning.
Comfort with optimization and probabilistic methods for allocating scarce resources under risk, such as constraint solvers or failure analysis.
Experience putting a dollar value on infrastructure choices, such as stranded power, build-out timing, or hardware utilization.
Energy optimization software for generation, battery storage, or large grid-connected sites.
Workload-aware scheduling with Kubernetes or Slurm.
