AI Network Performance and Reliability Engineer

AMP PBC; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

AMP PBC is building the largest independent AI fleet. You will own the east-west network across clusters, bringing up the fabric from switch access to production-grade performance for high-scale GPU training workloads.

This founding role requires hands-on deployment, tuning, and leadership to ensure robust, scalable network infrastructure. You will work with RoCEv2/RDMA at scale, diverse switch platforms, and NCCL benchmarking, shaping the fabric for frontier AI labs.

Qualifications

• Experience bringing up and tuning large GPU training fabrics.
• Direct hands-on RoCEv2/RDMA in production for 1,000+ GPUs.
• Deep knowledge of switch platforms (NVIDIA Spectrum-X, Arista, Cisco, Juniper).
• Comfort with NCCL and benchmarking.

Responsibilities

• Own east-west fabric bringup and tuning across clusters, starting with a B300 deployment and extending to the full Grid
• Take clusters from racked and cabled to production performance: switch access, fabric configuration, validation, and handoff to customers running real training workloads
• Tune lossless Ethernet end to end, including PFC, ECN and DCQCN congestion control, buffer allocation, QoS classes, hashing and traffic isolation
• Diagnose and fix the failure modes that quietly destroy training throughput: link health, packet loss, congestion collapse, low-entropy and bursty AI traffic patterns
• Benchmark and validate fabric performance against real collective operations, and defend the numbers our customers depend on
• Work across heterogeneous hardware by design. Different providers, different sites, different switch vendors and different silicon, with NVIDIA reference architectures as a floor rather than an answer
• Build the tooling and runbooks that make cluster onboarding fast and repeatable, so the fleet can scale without the process scaling with it
• Partner directly with frontier AI labs on the performance of the compute they are training on
• Hire and lead the network team as the Grid grows

Skills

RoCEv2 RDMA

Switch platforms depth

NVIDIA Spectrum-X

Leaf-spine topology

NCCL benchmarking

Team leadership

Education

Bachelor's degree

Tools

Arista EOS

Cisco Nexus

Juniper

SONiC

Cumulus Linux

Job description

About AMP PBC

AMP PBC is the AI infrastructure partner to independent teams building at the frontier. AMP is a public benefit company comprised of two major divisions, which together deliver both compute and capital under management.

About AMP PBC

AMP PBC is the AI infrastructure partner to independent teams building at the frontier. AMP is a public benefit company comprised of two major divisions, which together deliver both compute and capital under management.

AMP's technology division, AMP Infra, is building the independent AI Grid: pooled, automated infrastructure orchestration, across clouds and other compute providers, to give frontier teams on-demand access to the highest quality compute at any scale. Having built internal solutions for the world's largest hyperscalers, the AMP team is now creating a global, silicon-agnostic infrastructure network so that any team has the compute resources to build at the frontier without giving up their independence.

AMP's venture arm, AMP Foundry, partners with the world's leading researchers and scientists, incubating ideas and deploying strategic capital into frontier labs and other key areas of the AI infrastructure stack. With over $1 billion under management, Foundry operates at the pace the frontier requires, and provides fuel to help the best teams push the scaling laws.

AMP is backed by world-class investors, and partnered with leading labs, hyperscalers, research institutions, chipmakers and compute providers. We offer deep expertise, genuine ownership, and a relentless drive to maximize the world's frontier output.

The role

We are building the largest independent AI fleet in the world, and the fabric is what decides whether it works. You will own the east-west network across our clusters: the GPU-to-GPU interconnect, RoCEv2 and InfiniBand, from switch access on day one through a tuned cluster that trains at the performance our customers paid for. This is the founding role on our network team. The industry has undervalued how hard it is to bring up a cluster properly, and networking is the most undervalued part of it. A deployment can be enormous and enormously expensive, and if the fabric is wrong it was all for nothing. We treat this as a core competency of the business, not a support function, and it is top of mind for the founders.

What you’ll do

• Own east-west fabric bringup and tuning across our clusters, starting with a B300 deployment and extending to the full Grid
• Take clusters from racked and cabled to production performance: switch access, fabric configuration, validation, and handoff to customers running real training workloads
• Tune lossless Ethernet end to end, including PFC, ECN and DCQCN congestion control, buffer allocation, QoS classes, hashing and traffic isolation
• Diagnose and fix the failure modes that quietly destroy training throughput: link health, packet loss, congestion collapse, low-entropy and bursty AI traffic patterns
• Benchmark and validate fabric performance against real collective operations, and defend the numbers our customers depend on
• Work across heterogeneous hardware by design. Different providers, different sites, different switch vendors and different silicon, with NVIDIA reference architectures as a floor rather than an answer
• Build the tooling and runbooks that make cluster onboarding fast and repeatable, so the fleet can scale without the process scaling with it
• Partner directly with frontier AI labs on the performance of the compute they are training on
• Hire and lead the network team as the Grid grows

What we’re looking for

• You have done this before. There is no substitute. Adjacent experience is interesting, but we are looking for someone who has personally brought up and tuned a large GPU training fabric
• Hands‑on experience with RoCEv2 or RDMA over Ethernet in production, at 1,000 GPUs or more. InfiniBand experience is valuable alongside it, but Ethernet is where our hard problems live
• Either of these shapes works, and both are strong:
  • You owned the full stack at a smaller company or neocloud, from switch access through to a cluster that customers could actually use
  • You went deep on a specific domain of fabric performance at a hyperscaler, at 10,000 GPUs or more
• Real depth on switch platforms, for example NVIDIA Spectrum-X, Arista EOS, Cisco Nexus, Juniper, SONiC or Cumulus Linux
• Fluency with leaf‑spine and rail‑optimized topologies, and the judgment to know when the reference design is wrong for the deployment in front of you
• Comfort with NCCL and collective benchmarking, or the appetite to own it quickly
• A bias toward being in the room. We move fast, change direction inside a day, and solve problems in person
• You want to be the person the fleet depends on, and eventually the person who builds the team that does

Mechanics

• Minimum education: bachelor's degree or an equivalent combination of education and training or experience
• Location: Our office is located in San Francisco
• Visa sponsorship: Yes, we sponsor visas and will do everything we can to assist in this process with our legal support.

If you're driven to build the infrastructure that lets the world's best teams push the frontier, you belong here.