AI Factory Network Architect

Runware

12h•Remote

About The Position

As AI Factory Network Architect, you will own the network architecture of Runware’s AI supercomputer, from the NVLink domain inside a server to optical links between pods across continents. You will set the design, build the first systems hands-on, and grow the team responsible for operating them. This is a hands-on leadership role involving fabric specification and validation on real hardware, troubleshooting production issues, and establishing standards to prevent recurring problems. You will work with hardware, platform, and inference engineering teams, NVIDIA, optical and switching vendors, and site teams deploying pods. The role covers the scale-up fabric (NVLink and NVSwitch domains), scale-out fabric (InfiniBand and RoCE), storage and management networks, and the pod-to-pod and inter-site backbone. You will design the optical layer, define a repeatable reference network design for Sonic Inference Pods, bring up initial fabrics, and tune them against inference throughput and tail latency. You will align network topology and buffer decisions with inference traffic patterns, manage capacity planning and vendor roadmaps, establish operating standards, and hire and lead network engineers. Success means supporting inference traffic across tens of thousands of GPUs at line rate, onboarding pods quickly, and making network failures invisible to API developers.

Requirements

15+ years designing, building, and operating high-performance networks, including significant experience in HPC or AI supercomputers at the scale of thousands of GPUs or more.
Deep hands-on InfiniBand experience with fabric design, subnet management, routing engines, adaptive routing, congestion control, and performance debugging at the port and queue-pair level.
Working knowledge of NVLink and NVSwitch domains and the interaction between scale-up and scale-out fabrics for multi-GPU and multi-node inference and training.
Optical networking expertise, including 400G and 800G transceivers and cabling, link budgets, DWDM and coherent optics for inter-site links, and operational management of optics at scale.
Hands-on RDMA experience across InfiniBand and RoCE, including GPUDirect, NCCL and collective behaviour, PFC and ECN tuning, and the trade-offs between InfiniBand and RoCE.
Track record of leading network teams through greenfield builds and production incidents, and serving as an escalation point for complex fabric problems.
Ability to automate using Python or Go, infrastructure as code, and telemetry pipelines.
Clear writing skills and ability to produce executable design documents for site teams in other regions.
Willingness to travel to pod sites in the US and Europe for bring-up and major changes.

Nice To Haves

Experience with NVIDIA Spectrum-X, Quantum-X800, Blackwell-generation NVLink systems, or an equivalent hyperscaler in-house stack.
Hands-on experience with co-packaged optics, linear-drive pluggables, or photonic switching evaluations.
Experience designing networks for modular or containerized data centers, edge sites, or deployments with constrained power and space.
Background in inference-specific traffic patterns, including disaggregated prefill and decode, KV-cache transfer, and multi-tenant serving across a shared fabric.
Contributions to SONiC, OpenFabrics, UEC, or similar open networking efforts.
Experience with sovereign or data-residency network designs, including inter-site encryption and segmentation.
Prior experience at a GPU cloud, hyperscaler, national lab, or AI lab building its own supercomputer.

Responsibilities

Own end-to-end network architecture for the AI Factory, including scale-up, scale-out, storage, management, pod-to-pod, and inter-site networks.
Design the optical layer, including transceiver and cable selection, evaluation of co-packaged and linear-drive optics, DWDM for pod-to-pod and metro links, and link budgets and reliability targets.
Define the reference network design for Sonic Inference Pods so each pod has a validated, repeatable fabric and can join the global network in days.
Build and bring up initial fabrics, including rail-optimized topologies, subnet manager and routing configuration, adaptive routing, congestion control, QoS, and GPUDirect RDMA tuning.
Measure fabric performance against inference throughput and tail latency rather than synthetic benchmarks.
Work with the inference engine team to ensure traffic patterns—including collectives, KV-cache transfer, and multi-node serving—inform topology and buffer decisions.
Lead capacity planning and vendor roadmaps with NVIDIA, switch and optics suppliers, and carriers; negotiate technical specifications and evaluate new fabric generations.
Set operating standards for telemetry, link-health monitoring, failure-domain design, automated fault isolation, and runbooks.
Hire and lead a network engineering team as the footprint grows.
Serve as the final technical escalation for fabric incidents across sites.
Travel to pod sites in the US and Europe for bring-up and major changes.