OpenAI Compute Infrastructure Engineer Interview Questions & Process (2026)

1. 1

   OpenAI's mission is ensuring AGI benefits all of humanity. How does that shape the way you'd think about reliability and capacity decisions for the compute that trains frontier models?

   How to answer: Connect infra decisions to research and safety outcomes: how checkpoint reliability protects months of training investment, how fair scheduling avoids starving safety-critical evals, how capacity planning affects the pace and predictability of research.

“Tell me about yourself”, “Why here?” and the other common questions, and how to prep each →

1. 1

   Tell me about a time you responded to a major outage or degradation in infrastructure you owned. What happened and how did you resolve it?

   How to answer: Use STAR. Cover detection, triage under pressure, the actual fix, and the follow-up — what changed afterward so the same failure mode couldn't recur silently.

2. 2

   Describe a time you disagreed with a decision about infrastructure investment or prioritization. How did you handle it?

   How to answer: Show you can reason from first principles about cost/reliability/velocity trade-offs and communicate them to non-infra stakeholders without just asserting authority.

3. 3

   Tell me about a time you significantly improved the efficiency of a costly or slow infrastructure process.

   How to answer: Quantify the improvement. Walk through how you identified the bottleneck (profiling, monitoring, or postmortem), what you changed, and how you validated the gain held up in production.

The six topics behind behavioral questions, and how to prep a story for each →

1. 1

   Walk me through how you'd approach bare-metal provisioning automation for a fleet that's growing by thousands of GPU nodes a quarter.

   How to answer: Discuss zero-touch provisioning, firmware/driver version drift across a heterogeneous fleet, health-check gating before a node re-enters the scheduling pool, and how you'd stage a rollout to avoid fleet-wide breakage from a bad automation change.

2. 2

   Explain the trade-offs between NVLink and InfiniBand for a distributed training job, and how network topology should influence scheduling.

   How to answer: Cover intra-node vs inter-node bandwidth characteristics, how topology-unaware scheduling silently degrades all-reduce performance, and why colocating a job's GPUs by network proximity matters for large-scale training throughput.

3. 3

   How would you extend a Kubernetes scheduler to do GPU-topology-aware bin packing for multi-GPU training jobs?

   How to answer: Discuss custom scheduler extenders or plugins, how you'd model GPU/network topology as scheduling constraints, and how to avoid regressing scheduling latency for the rest of the cluster while adding this complexity.

4. 4

   How would you use workload profiling and benchmarking data to influence a multi-quarter capacity planning decision?

   How to answer: Discuss what signals actually predict future demand (historical utilization, upcoming training runs, inference traffic growth), how you'd avoid over- or under-provisioning, and how you'd communicate uncertainty in the forecast to leadership.

5. 5

   A large training run's throughput has degraded 15% over the past week with no code changes. How do you diagnose it?

   How to answer: Show systematic diagnosis: check for network degradation, thermal throttling, noisy-neighbor contention, storage I/O bottlenecks on checkpointing, or scheduler placement changes. Propose the monitoring you'd want in place to catch this automatically next time.

1. 1

   How would you design a GPU cluster scheduling system that fairly allocates capacity between long-running training jobs and latency-sensitive inference workloads?

   How to answer: Cover preemption policies, priority classes, bin-packing for GPU topology awareness (keeping multi-GPU jobs on the same NVLink domain), and how you'd prevent a large training job from starving smaller inference jobs indefinitely.

2. 2

   Design a checkpoint storage system for training runs that must survive node failure without losing more than a few minutes of progress on a run spanning thousands of GPUs.

   How to answer: Cover checkpoint frequency vs. storage/network overhead trade-offs, sharded/async checkpointing, storage durability and throughput requirements, and how a training job resumes cleanly after a partial-fleet failure.

3. 3

   Design a capacity allocation system across competing training and inference workloads when total GPU supply is fixed and demand exceeds it.

   How to answer: Cover quota systems, priority tiers, spot-style reclaimable capacity for lower-priority work, and how you'd give teams visibility into utilization and queueing rather than a black-box allocator.

System design questions grouped by the concept they test, with a full walkthrough →