The AI Infrastructure Engineer Interview Process

A recruiter call to confirm the track and the level, then a technical screen (a practical coding problem or a technical conversation about a system you built), then an onsite or virtual loop of four to six rounds: a coding round of the systems kind, a design round for the track (an inference platform, a training scheduler, a cluster), a deep-dive on your own work, a domain round (a kernel, a distributed-training or a fleet-reliability conversation), and one or two behavioral rounds, with a values or mission conversation at some labs. Timelines run two to six weeks. Some companies replace the onsite's coding with a take-home or a paid work trial. The company pages describe each company's reported loop and the evidence behind it, and never invent a round where we have none.

The recruiter call is where the level is set: say which level you are interviewing for, because the interviewers are asked to listen for that level's signals and the panel discusses them afterward. The leveling page lists the signals for senior and staff.

Kernel: can you make one accelerator run a model fast, and can you say why it is slow? Tested by kernel problems, profiler conversations and roofline reasoning, most heavily at the labs' performance teams and the chip makers. Platform: can you run a fleet, with scheduling, quotas, health and rollouts, and can you tell an incident story that ends in automation? Tested by scheduler designs and behavioral rounds, most heavily at the labs' fleet teams and the GPU clouds. Distributed: can you plan a training run's parallelism, derive the traffic on each link, and keep the run alive through failures? Tested by parallelism-plan questions and slow-step investigations at the labs, the hyperscalers' training teams and the chip makers' cluster teams.

Design: can you draw the reference shape for an inference platform, a scheduler or a cluster, size it with a stated chain of arithmetic, and name the failure modes unprompted? Tested in the design round everywhere; it is the round that decides level most often. Cost: can you turn a workload into a bill of materials and a cost per token, and can you say what moves it? Tested inside every other round as the sanity check, and directly at the GPU clouds and the inference providers whose margin is the answer.

The company pages tag each company with the axes its loop weights most, and the readiness assessment scores you on the same axes so you know where to study.

Take-homes appear at some GPU clouds and inference providers: a small service (a batcher, a scheduler, a parser) built in a few hours and reviewed live. Work trials, a paid day or several days on a real problem with the team, are used by a few labs and startups for senior hires; they test collaboration and judgment on real code and are the strongest signal a company can collect. Pairing rounds, where you and an interviewer extend a codebase together, are common at the inference providers. The scoring in every format is the same: working code, tests, narration of trade-offs, and how you handle the second and third extension of the problem. The coding screen playbook covers the 45-minute structure.

Policies have split. Some companies ban assistants in coding rounds, say so in the invitation, and occasionally proctor; OpenAI is reported to ban AI in coding rounds. Others allow or expect an assistant, and some run rounds explicitly designed around one. The scored behaviours change with the policy: under a ban the round measures fluency, and with an assistant it measures specification, verification and judgment (writing the invariants and tests first, reading every line that comes back, catching the assistant when it is wrong on a concurrency edge case). Ask the recruiter which policy applies rather than guessing, and prepare for the one you get.

Frontier labs (OpenAI, Anthropic, Google DeepMind, Meta, Microsoft AI, xAI): the deepest domain rounds, a performance or fleet conversation that goes several layers down, a design round at frontier scale, and at some labs a values or mission conversation; the safety and mission rounds page covers what that round tests. Chip makers (NVIDIA, AMD, AWS Annapurna, Google TPU, Cerebras, Groq): architecture depth, CUDA or the vendor's stack, benchmarking methodology, and cluster validation; NVIDIA's new-grad loops are reported as a one-hour assessment and a multi-hour onsite. GPU clouds (CoreWeave, Lambda, Crusoe, Nebius, Together AI): fleet operations, Kubernetes at scale, incident stories, and the cost axis weighted heavily. Inference providers (Fireworks, Baseten, Modal): serving-engine depth, latency budgets, and pairing rounds on real code. Product companies with ML platforms (Uber, Netflix, Databricks, Snowflake): backend system design with GPUs as the resource, multi-tenancy, and the platform's customers as stakeholders.

India loops at the multinationals follow the parent company's format with local panels; India-headquartered companies run shorter loops with a heavier weight on hands-on GPU and Kubernetes experience. The companies guide lists who is hiring in each family.

Take the readiness assessment to find your weakest axis; work the napkin math track until the twelve formulas are reflexes; read the concept tracks for your target track in the order the study paths give; practice the five coding patterns cold; run three design prompts against the design-round playbook with a timer; write your incident, pushback and migration stories in five beats with numbers; and read the company page for each loop you have booked. The question bank is organized by the same categories, and every answer shows its arithmetic.