AI Infrastructure Operations, Demand Planning

Anthropic runs one of the largest and fastest-growing infrastructure fleets in the industry, across multiple accelerator families, CPU families, clouds, neoclouds, and on-prem sites. Capacity Engineering owns the data, tooling, and systems that let Anthropic plan, measure, and maximize utilization of that fleet: we partner on supply deals, wire telemetry from day zero, own the canonical capacity data layer, and build the planning and enforcement tools every research and product team relies on. This role sits in the Planning pillar, on the Demand Planning team, and works daily with research engineering, pretraining, inference, compute supply, finance, and external vendors.

You own the tranches. The job has two halves that feed each other. Upstream, you take the Demand Planning forecast and turn it into per-tranche requirements — shape, interconnect, region, supporting resources, date — and carry those into sourcing negotiations and data center build reviews so we contract for capacity we can actually use when we need it. Downstream, you own the integrated schedule and system of record for every tranche in flight — from contracted through reserved, ingested, in-cluster, healthy, and occupied — and you drive the owners of each hop to their dates. Every slip you see downstream becomes a contract-language fix, an automation, or a correction fed back to the forecast.

• Turn the forecast into per-tranche requirements. Take the Demand Planning forecast plus direct input from research, pretraining, and inference planners, and convert it into concrete accelerator, interconnect, region, supporting-resource, and date requirements for each tranche. Represent those in sourcing negotiations and data center build reviews, including which contractual terms actually move delivery dates.
• Qualify tranches for deliverability before signature. The Capacity Planner signs fit-to-forecast; you sign whether the shape can land schedulable, healthy, and instrumented in that region on that date, with storage, egress, identity in place.
• Close the delivery loop. Track forecast-versus-delivered on shape, region, and timing for every tranche; publish the variance; and feed it back to Demand Planning and into the next contract.
• Own the bring-up system of record. Define the canonical contract-to-occupied state machine with explicit entry and exit criteria per stage, and make it a first-class object in the capacity data layer so every downstream tool sees in-flight capacity, not only what has landed.
• Run a portfolio of bring-ups in parallel — new cloud regions, on-prem sites, neocloud blocks — with one integrated schedule spanning provider milestones, cluster creation, network turn-up, storage readiness, health burn-in, and first-workload landing. 
• Drive readiness automation: All capacity systems are fully integrated for all new capacity, from contracted through ingested, automated and scaled.
• Instrument and publish the numbers that matter — time-to-occupied and paid-idle dollars per tranche — with executive-level reporting on status, tradeoffs, and risk across the portfolio.

• Significant experience delivering large-scale infrastructure — cloud regions, accelerator clusters, HPC systems, or bare-metal fleets — at multi-region scale or ≥10k accelerators (or CPU/storage equivalent).
• Technical range from through cluster orchestration and node health, up to the telemetry and planning tables on top — enough to debug where they disagree rather than route it.
• SQL and enough Python to answer your own questions and build your own reporting.
• A degree in a technical field or an equivalent engineering track record.

• Reserved-capacity onboarding, private offers, or capacity commitments with cloud or neocloud providers.
• Enough demand-planning exposure to challenge a forecast, translate it into per-tranche requirements, and feed delivery variance back into it.
• Data center or colocation delivery: power and space planning, network turn-up, site acceptance, vendor management.
• Accelerator health and burn-in, collective-communications sanity testing, or fleet-health SLOs — and a rigorous definition of "healthy."
• Systems of record or lifecycle services for infrastructure assets.
• Onboarding a new hardware generation into an existing scheduler and observability stack.