Methodology · Underwrite
Underwrite methodology
Underwrite answers one question: will this project pencil? Its numbers are modeled scenarios, not quotes or forecasts.
Compute Underwriter
Models what an AI data center earns or loses per GPU-hour (the compute spread) in 86 economies at 1 MW, 100 MW, and 1 GW of grid power.
- What it computes: GPU rent revenue per installed GPU-hour, idle hours included, minus electricity, GPU payback, the building (land, building, and grid connection), staff and admin, and a country risk charge in two layers (customer credit, and political and transfer risk).
- Stress tests: break-even utilization, GPU rent, and power price, and a combined downside of 20% lower GPU rent and one year less GPU life.
- Site sizes: GPUs follow from grid power, PUE, and server power per GPU. At PUE 1.3, 1 MW holds 603 H100s.
- Sources: the Virya Electric Grid Index, World Bank, ILOSTAT, Eurostat, IMF, Damodaran, national regulators and utilities, European Central Bank reference rates, and ExchangeRate-API.
- Refresh: exchange rates twice a day; other inputs are reviewed periodically.
Site Map
The US grid around a site: 94,619 transmission lines by voltage, power plants by fuel, 2,931 retail utility territories, holding company and planning areas, emission intensity, and ERCOT prices.
- Sources: public US government datasets (HIFLD, US EPA eGRID and facility registries, EIA-860, DOE, FAA, FCC, FDIC, USGS, USACE) and the ERCOT Public Data API.
- Emissions: EPA eGRID 2014 to 2023, every rate recomputed from eGRID masses and generation, plus a 2024 layer from a non-EPA run of EPA's eGRID code, flagged as such.
- Limits: the plant layer is the subset of EIA-860 plants matched in the EPA Facility Registry, not the full US fleet; ERCOT load zones are traced and approximate.
Scenario Lab
A day in a five-bus power market built on the PJM 5-bus test system. It commits units, then runs security-constrained economic dispatch for each of 24 hours and splits every locational price into energy, congestion, and losses. A teaching and analysis model, not a market forecast.
Modeled scenarios and analysis, not investment advice. Detailed model documentation is available on request: contact@viryaintel.com.