Network Engineer, Capacity and Efficiency

Job Application for Network Engineer, Capacity and Efficiency at Anthropic

Network Engineer, Capacity and Efficiency San Francisco, CA | New York City, NY

About Anthropic

Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.

About the team

The Capacity & Efficiency team sits inside Anthropic’s Compute organization and owns the cost, utilization, and attribution story for non-accelerator infrastructure — the network, compute, and storage backbone that moves petabytes between training clusters, inference fleets, and object storage across clouds and regions. The scale is real, the spend is large, and the efficiency levers are still mostly unpulled.

We work alongside the Systems Networking team (who build and operate the fabric) and the Observability team. This role lives at the intersection: you’ll use deep networking knowledge and rigorous measurement to figure out where and how bandwidth, latency, and dollars are being used, find optimization opportunities and land them.

About the role

We’re looking for a network engineer who thinks in metrics first. You understand spine-leaf fabrics, BGP, SDN overlays, and cloud interconnect products well enough to build them. You will instrument them, model their cost-per-bit, and squeeze out the inefficiency, while ensuring we can move the bits to the right places in the most efficient manner. You’ll own the observability and efficiency surface for Anthropic’s network: building intelligence using telemetry, to understanding how workloads use the network, to cost attribution that tells a research team exactly what their checkpoint sync is costing.

This is a hands-on IC role. You’ll write code (Python, Go), build dashboards, model capacity, and work with networking teams to help meet the needs of the workload owners. You’ll also influence architecture: when the data says a traffic pattern is pathological, you’ll be in the room root causing it and fixing it. You will be working across multiple areas: network telemetry and observability, and cost modeling and attribution. We expect you to be strong in at least two and willing to grow into the third. If you're a telemetry-first engineer who's never built a chargeback model, or a traffic engineer who hasn't shipped eBPF probes, apply anyway and tell us which axis you want to grow on.

What you’ll do

Workload network profile development: characterize how each major workload actually uses the network: bandwidth, latency sensitivity, cross-cloud, cross-region traffic patterns, topology dependencies. This is the observability foundation everything else builds on.

Build the network observability stack. Build or use telemetry pipelines, sFlow/IPFIX, gNMI streaming, eBPF host probes, to turn packet counters into per-flow, per-tenant, per-workload cost and utilization data.

Usage monitoring, attribution & cost model : Use network telemetry to attribute end-to-end usage, egress, and interconnect transit costs back to workloads & teams. Collaborate on designing a cost data model for network usage.

Capacity sizing & forecasting: use telemetry, growth drivers, forecast interconnect, egress, intra-DC bandwidth needs and feed procurement & contract teams ahead of demand.

Hunt for efficiency. Analyze inter-region traffic patterns, identify hot links and stranded capacity, and quantify the dollar impact. Build the models that tell us whether we should buy more capacity, or move the workload.

Influence decisions you don't own . A large fraction of this role is convincing other teams to act on what your data shows: making the case to research that a traffic pattern needs to change, to finance that an interconnect tranche is worth buying, to Systems Networking that a QoS policy needs rewriting. You'll partner closely with Systems Networking on fabric architecture and Observability on telemetry platform integration, but the cost and efficiency wins will come from moving teams that don't report to you.

Automate. Extend our intent-based network configuration systems and write the tooling that turns your efficiency findings into safe, reviewable, and impactful changes.

You may be a good fit if you

Have 5+ years operating large-scale production networks — data center fabrics (spine-leaf, Clos), backbone/WAN, or hyperscaler-adjacent environments.

Understand how traffic moves through the network even if you don't know the specifics of how.

Know at least one major CSP’s networking model well AWS (VPC, TGW, Direct Connect, Gateway Load Balancer) or GCP (Shared VPC, Interconnect, Cloud Router, Network Connectivity Center)

Have built or operated network telemetry at scale: streaming telemetry (gNMI/OpenConfig), flow export (sFlow, IPFIX, NetFlow), or eBPF-based host-side instrumentation. You can reason about sampling, cardinality, storage tradeoffs, and enrich telemetry to build intelligence and actionable insights.

Comfortable writing Python or Go to build tooling, telemetry pipelines, infrastructure-as-code, config management for network devices and automation, that you’ll ship to production.

Think quantitatively by default. You reach for a notebook or a Grafana query before you reach for an opinion, and you can turn messy counter data into a defensible cost model.

Communicate crisply. You can explain to a finance partner why a 10% egress reduction matters, and to a network engineer why a specific ECMP imbalance is costing real money.

Strong candidates may also have

Background on a cloud provider's networking team or a cloud networking product team — building or operating the interconnect, backbone, or SDN control plane from the provider side, not just consuming it as a customer.

Familiarity with AI/ML infrastructure traffic patterns like collective communication (all-reduce, all-gather), checkpoint/weight transfer, inference serving, and how these stress networks differ than traditional workloads in terms of burst behavior, flow synchronization, and bandwidth symmetry.

Background in traffic engineering for large backbones and the operational judgment to know when TE is worth the complexity.

Hands-on time with multi-cloud connectivity: cross-cloud peering, private interconnect products, and the billing models that come with them.…