The opportunity
Anthropic's infrastructure fleet spans a growing set of clouds, neoclouds, and on-prem sites. Every part of it is expanding rapidly.
What you'll do
Build and own the medium-range multi-resource demand forecast: accelerators by chip/interconnect class, CPU by shape, storage by tier and access pattern, egress by path, managed services by SKU — driven by model roadmap, RL/inference growth, eval volume, and retention policy rather than trend lines.
Run the plan-vs-reality loop. Diff planned allocations against observed fleet: occupancy weekly, surface unrecorded trades and stale allocations, and drive variance toward zero with our planning-tools and data teams.
Qualify each incoming capacity tranche against the forecast before signature:: right shape, region, quarter, and supporting-resource envelope (storage, egress, CPU).
Partner with Finance and cost-efficiency teams to turn the forecast into core: drivers covering the large majority (≥80%) of non-accelerator spend, and to aim savings work where waste will appear next.
Have done capacity, demand, or supply planning for large-scale technical: infrastructure (cloud, HPC, hyperscale, or a large internal platform) and can articulate the difference between a forecast, a plan, and an allocation.
Build the model yourself rather than specifying it for someone else: SQL against a warehouse, Python/pandas or a proper forecasting/optimization stack.
What they're looking for
- Understand data-center resource classes well to know why storage and egress: don't forecast like GPUs, and why a contract's headline chip count is rarely the binding constraint.
- Prefer simple, inspectable models to clever opaque ones, and instrument your own forecast error.
- Direct experience with cloud or neocloud providers on reserved-capacity: onboarding, private offers, or capacity commitments.
- Demand planning or forecasting experience for accelerator fleets, including: translating research or product roadmaps into resource requirements.
