Asks for

The ad says“You've owned infrastructure as code and observability for a production fleet, provisioning with Terraform or Ansible, deploying with Helm and ArgoCD, and monitoring with Prometheus, or their equivalents.”
The ad says“You've owned infrastructure as code and observability for a production fleet, provisioning with Terraform or Ansible, deploying with Helm and ArgoCD, and monitoring with Prometheus, or their equivalents.”
The ad says“You're a strong programmer in at least one language that infrastructure is commonly written in, such as Python, Go, Rust, or C++, and your automation and services are maintained as shared code.”
The ad says“You're a strong programmer in at least one language that infrastructure is commonly written in, such as Python, Go, Rust, or C++, and your automation and services are maintained as shared code.”
The ad says“You've owned infrastructure as code and observability for a production fleet, provisioning with Terraform or Ansible, deploying with Helm and ArgoCD, and monitoring with Prometheus, or their equivalents.”
The ad says“You've run production Kubernetes for GPU workloads with a batch layer on top (Slurm, Kueue, Volcano, or similar), including quotas, priority and preemption, and node health.”
The ad says“You have 3+ years in systems or infrastructure engineering on production Linux, running GPU, HPC, or large-scale batch platforms, and you've owned at least one system from design through operation.”
The ad says“You've owned infrastructure as code and observability for a production fleet, provisioning with Terraform or Ansible, deploying with Helm and ArgoCD, and monitoring with Prometheus, or their equivalents.”
The ad says“You're a strong programmer in at least one language that infrastructure is commonly written in, such as Python, Go, Rust, or C++, and your automation and services are maintained as shared code.”
The ad says“You're a strong programmer in at least one language that infrastructure is commonly written in, such as Python, Go, Rust, or C++, and your automation and services are maintained as shared code.”
The ad says“You've owned infrastructure as code and observability for a production fleet, provisioning with Terraform or Ansible, deploying with Helm and ArgoCD, and monitoring with Prometheus, or their equivalents.”

Values

The ad says“Distributed storage: VAST, Weka, Lustre, Ceph, or object storage at scale; checkpoint I/O.”
The ad says“Cluster security: admission control, RBAC, node and container hardening, sandboxed runtimes (gVisor, Kata, Firecracker), and isolating autonomous agents on shared infrastructure.”
The ad says“Cluster security: admission control, RBAC, node and container hardening, sandboxed runtimes (gVisor, Kata, Firecracker), and isolating autonomous agents on shared infrastructure.”
The ad says“Distributed training infrastructure: multi-node PyTorch and NCCL debugging, the NVIDIA node stack (drivers, GPU Operator, DCGM), InfiniBand or RoCE fabrics, topology-aware placement.”
The ad says“Cluster security: admission control, RBAC, node and container hardening, sandboxed runtimes (gVisor, Kata, Firecracker), and isolating autonomous agents on shared infrastructure.”
The ad says“Distributed storage: VAST, Weka, Lustre, Ceph, or object storage at scale; checkpoint I/O.”
The ad says“Distributed training infrastructure: multi-node PyTorch and NCCL debugging, the NVIDIA node stack (drivers, GPU Operator, DCGM), InfiniBand or RoCE fabrics, topology-aware placement.”
The ad says“Distributed training infrastructure: multi-node PyTorch and NCCL debugging, the NVIDIA node stack (drivers, GPU Operator, DCGM), InfiniBand or RoCE fabrics, topology-aware placement.”
The ad says“Distributed training infrastructure: multi-node PyTorch and NCCL debugging, the NVIDIA node stack (drivers, GPU Operator, DCGM), InfiniBand or RoCE fabrics, topology-aware placement.”
The ad says“Distributed training infrastructure: multi-node PyTorch and NCCL debugging, the NVIDIA node stack (drivers, GPU Operator, DCGM), InfiniBand or RoCE fabrics, topology-aware placement.”
The ad says“Distributed storage: VAST, Weka, Lustre, Ceph, or object storage at scale; checkpoint I/O.”
The ad says“Distributed storage: VAST, Weka, Lustre, Ceph, or object storage at scale; checkpoint I/O.”

About Us

FAR.AI is a non-profit AI research institute working to ensure advanced AI is safe and beneficial for everyone. Our mission is to facilitate breakthrough AI safety research, advance global understanding of AI risks and solutions, and foster a coordinated global response.

Since our founding in July 2022, we've grown to 50+ staff, published 40+ academic papers, and convened leading AI safety events. Our work is recognized globally, with publications at premier venues such as NeurIPS, ICML including a Best Paper Honorable Mention in 2026, and ICLR, and features in the Financial Times, Nature News, Wired Magazine and MIT Technology Review. We conduct pre-deployment testing on behalf of frontier developers such as OpenAI, independent evaluations for governments including the EU AI Office, and publish the AI Security Leaderboard based on our red-teaming expertise. We help steer and grow the AI safety field through developing research roadmaps with renowned researchers such as Yoshua Bengio; running FAR.Labs, an AI safety-focused co-working space in Berkeley housing 40+ members; and supporting the community through targeted grants to technical researchers.

About the Team

Foundations is FAR.AI's infrastructure and engineering team. Our remit is broad: we run the compute platform, build the tools and frameworks researchers work in, automate research workflows, and help teams scale experiments well past what they'd manage alone. Our job is to accelerate the research. We do so by working directly with researchers through embedded engagements and day-to-day consulting, and building systems that can scale with the organization as it grows.

Foundations is growing quickly, and our infrastructure portfolio is growing fastest. We run FAR.AI's research on a mix of bare-metal and managed Kubernetes GPU clusters from multiple providers. We rent the hardware and operate the platform ourselves. The fleet has grown from dozens to hundreds of GPUs this year and it's continuing to grow quickly: we're adding providers, taking on users beyond our own researchers, and moving experiments onto frontier open-weight models. A large amount of research is now being done by AI agents working directly on the cluster, which is driving updates to our platform infrastructure and security.

Running it well now takes dedicated specialists, so we're standing up an infrastructure sub-team that owns the cluster fleet: adding capacity, designing and managing the networking and storage under it, infrastructure as code, and the security posture, plus some of the platform layer above it. It works directly with research teams as their needs change.

About the Role

You'd work across the whole infrastructure stack, from scheduling to storage to monitoring to security, and bring real depth in at least one part of it. We're particularly interested in experience with large-scale pre-training and post-training infrastructure and the network fabric under it, cluster security and sandboxing, distributed storage systems, and batch scheduling for large GPU clusters. Expertise in an adjacent area is also a good fit.

In frontier AI research, working out the infrastructure is often part of the science. You'd work directly with researchers and other engineers to keep our large-scale experiments performant and fault-tolerant.

We're also hiring a Tech Lead Manager, GPU Cluster Infrastructure for this team. If leading a small team while staying hands-on sounds like you, take a look there instead.

What you'll do

Operate the Kubernetes GPU fleet day to day. You handle node lifecycle, upgrades, driver and image rollouts, staged changes with safe rollback, and capacity planning.
Own batch scheduling and multi-tenancy, including queues, quotas, priorities, preemption, gang scheduling, and fair share across research teams.
Design and run the storage under the fleet, from high-performance shared filesystems for datasets and checkpoints to object storage tiers, quotas, and backups.
Keep multi-node training runs fault-tolerant. You own node health and automated draining, debug NCCL and fabric problems, track down stragglers and flaky GPUs, and build the checkpoint and restart patterns.
Harden the platform, covering identity and access, network policy, secrets, workload isolation, and sandboxing for the AI agents that run on the cluster.
Bring new capacity online. You acceptance-test providers on fabric, NCCL, and storage throughput, hold them to their SLAs, and integrate new clusters into the platform with infrastructure as code.
Work directly with research teams on their infrastructure problems and turn the recurring ones into platform fixes. Share the on-call rotation, runbooks, and postmortems.

Requirements

You have 3+ years in systems or infrastructure engineering on production Linux, running GPU, HPC, or large-scale batch platforms, and you've owned at least one system from design through operation.
You've run production Kubernetes for GPU workloads with a batch layer on top (Slurm, Kueue, Volcano, or similar), including quotas, priority and preemption, and node health.
You've owned infrastructure as code and observability for a production fleet, provisioning with Terraform or Ansible, deploying with Helm and ArgoCD, and monitoring with Prometheus, or their equivalents.
You're a strong programmer in at least one language that infrastructure is commonly written in, such as Python, Go, Rust, or C++, and your automation and services are maintained as shared code.
You write clearly for engineers, researchers, and providers, whether it's a design doc, an incident summary, or an escalation.

If you meet most of this and not all of it, we encourage you to apply anyway.

Additional skills we're excited about

Real depth in one or more of these makes you a compelling candidate:

Distributed training infrastructure: multi-node PyTorch and NCCL debugging, the NVIDIA node stack (drivers, GPU Operator, DCGM), InfiniBand or RoCE fabrics, topology-aware placement.
Distributed storage: VAST, Weka, Lustre, Ceph, or object storage at scale; checkpoint I/O.
Cluster security: admission control, RBAC, node and container hardening, sandboxed runtimes (gVisor, Kata, Firecracker), and isolating autonomous agents on shared infrastructure.
Scheduler internals: Kubernetes scheduler plugins or custom controllers, gang scheduling, fair-share and quota, and the utilization, fairness, and latency tradeoffs between them.
Multi-provider platforms: scheduling and storage across clusters at different providers so users see one system, including clusters with no shared network and uneven data locality.

Benefits*

Health Insurance - 94% of Insurance premium paid by Organization commencing within 1 month after your start date
Retirement - 401(k) plan with up to 2% match
PTO - 25 days Paid Time Off per year, accrued weekly and up to 10 days of paid sick leave per year
Paid Leave - Paid Bereavement, Family, Medical and Pregnancy Disability Leave
WFH Stipend & Equipment - Work computer and stipend provided for eligible employees
Catered Meals (Berkeley Office Only) - Catered lunches and dinners on workdays at our office

*(These benefits are currently only offered to full-time employees based in the US, we use an EOR for our international employees who have their own set of benefits for Health, Paid Leave, etc.)

Logistics

If based in the USA or Singapore, you will be an employee of FAR.AI (501(c)(3) research non-profit / non-profit CLG). Outside the USA or Singapore, you will be employed via an EOR organisation on behalf of FAR.AI or as a contractor.

Location: Both remote and in-person (Berkeley, CA or Singapore) are possible. We sponsor visas for in-person employees, and can hire remotely in most countries. For this role we prefer candidates whose working hours overlap with Berkeley.
Hours: Full-time (40 hours/week).
On-call: We don't run a formal on-call rotation yet. The team is spread across time zones and covers incidents during working hours. As the experiments we run get larger we expect to introduce one, and this role would take part in it.

If you have any questions about the role, feel free to contact us at talent@far.ai. Otherwise, if you don't have questions, the best way to ensure a proper review of your skills and qualifications is by applying directly via the application form. Please don't email us to share your resume (it won't have any impact on our decision). Thank you!