Want to know if this job is worth applying to?
San Francisco, California, United States
Onsite
Lithos builds the fastest agentic inference on the planet: frontier open models at full precision on standard GPUs. Our engine currently holds top spots on Artificial Analysis for output speed, time to first token and end-to-end latency on models like DeepSeek and Kimi. We serve inference to growing startups as well as FT500 companies through a public API or on their own hardware.
Lithos was founded by Carnegie Mellon computer science professors who spent years turning systems research on inference and GPU virtualization into a product. We are a small team, and this is one of our first ten hires.
For an inference company, capacity is existential. Every promise we make on speed, price and uptime depends on having the right compute, in the right place, at the right price, before the demand arrives.
You will own that. This is the most important commercial seat at Lithos after the founders, and it reports directly to them.
Own the multi-year capacity plan. Translate demand forecasts and the model roadmap into GPU, network and data-center requirements, and stay ahead of demand rather than chasing it.
Run procurement across vendors. Source and negotiate GPU capacity, networking and colocation across neoclouds, colo providers, OEMs and hyperscalers. Build a vendor bench deep enough that no single supplier can hold us hostage.
Own the commercial terms. Lead contracts end to end: commit profiles, drawdown schedules, pricing, acceptance testing, uptime definitions, service credits, network fabric performance guarantees and SLA remedies. You will take over live, in-flight deals on day one.
Build a rolling price book and vendor scorecard. Track market pricing, delivery reliability and real-world performance by vendor, region and GPU generation.
Model total cost of ownership (TCO) per unit of performance. Measure cost per token and cost per GPU-hour against what the hardware actually delivers, and make that the basis for every capacity decision.
Partner with engineering. Work with our systems and inference teams to make sure the hardware we buy is what our software runs best on, and that acceptance tests prove it.
7+ years in compute capacity, infrastructure procurement or data-center commercial roles at a neocloud, colo provider, OEM or channel partner, or a hyperscaler capacity team.
You have personally negotiated and closed seven-figure-plus infrastructure contracts, including commit structures, drawdowns and SLA remedies, and can walk through one in detail.
You think and speak in numbers: $/GPU-hour, utilization, power density, delivery lead times, effective cost per unit of performance.
Working knowledge of current GPU generations, high-speed networking (InfiniBand/RoCE) and what separates a good cluster from an expensive one.
Comfortable as a team of one at first: building the spreadsheets, the processes and the vendor relationships yourself.
Nice to have: experience with capacity for AI inference or training workloads, or on the vendor side selling GPU capacity to AI labs.
