Head of AI Infrastructure.

The company was moving from in-house prototypes used by a small research group to production deployments running training workloads for external partners. The transition required someone who had personally lived inside a training stack at frontier scale — interconnect, scheduler, checkpointing, failure recovery — and could read the gap between what worked at internal scale and what would survive in customer hands. The Head of AI Infrastructure seat was created to own that transition and to lead the engineering organisation underneath it.

The non-negotiables were tightly drawn. Production training-infrastructure experience at frontier scale, comfort across the full stack from networking to scheduler to user-facing tooling, and credibility with the research customers whose workloads would land first. The comp band was structured with significant equity at the deployment inflection. Geography was preferred US west coast or UK, with a willingness to consider Zurich, Munich or Tel Aviv for the right candidate. The CEO was explicit that they would not consider a candidate whose experience was research-supervisory rather than engineering-led.

Senior training-infrastructure leadership is one of the thinnest benches in AI. The work cuts across networking, accelerator systems, scheduler design and ML-platform engineering, and the practitioners who have shipped it at frontier scale sit inside a handful of labs and infrastructure companies. Reference networks are exceptionally tight — the bench is well-known to itself, and movement between organisations is observed across the sector within days. The pull from foundation labs has compressed availability further, particularly for engineers who have carried a training run end to end.

Our read was that the shortlist could only be built through direct outreach into that bench, with assessment grounded in the specific engineering work each candidate had owned. Search firms without practitioner depth tend to misread this segment badly — they cannot tell whether a candidate supervised a training run from a leadership distance or wrote the recovery code at three in the morning. We assessed on what each candidate had personally shipped, which incidents they had carried, and which design decisions had survived contact with production.

The shortlist composition was weighted toward engineers with end-to-end production training-stack ownership rather than research leadership with infrastructure adjacency. We held a strict line on demonstrable hands-on work and on customer-facing temperament, given the company's transition from internal-only to external workloads.

• Infrastructure engineer from a frontier-lab training team.
• Systems leader from a large-scale ML platform company.
• Practitioner with cross-stack networking and scheduler ownership.
• Engineer with public work on training failure recovery.

The hire came from inside a frontier lab's training-infrastructure group, where they had personally led the redesign of the scheduler and recovery path through a production training run. What made them right was the combination of depth and operating temperament — they spoke about the work with the specificity of someone who had carried the pager, and they engaged the company's research customers in the second interview as peers rather than as a vendor. The CEO and the founding research adviser read the same signal independently.

The close was the most considered of the engagement. The candidate was weighing a competing internal promotion at their lab and required time and structure on the equity question. We worked through the comp framing with the CEO across two weeks and held the candidate close on the deployment story rather than on cash. Brief to offer ran fifteen weeks, with the offer accepted in the second reference cycle.

“The engagement reinforced a discipline we have come to depend on for the thinnest segments of the AI bench: assess on the specific engineering artefact, not on the organisational chart. Candidates who had run training-infrastructure organisations without personally writing or reviewing the systems code read confidently in interview and broke down under reference. The hire was the candidate whose former colleagues described the design decisions in the same language the candidate had used unprompted — a small signal that is, in practice, the difference between a credible offer and a wasted one.”

— Peter Wood