Virya Compute Risk Explorer

The Virya Compute Risk Explorer models what an AI data center earns or loses per GPU-hour in 86 economies, at site sizes of 1 MW, 100 MW, and 1 GW. It sets GPU rent revenue against electricity, GPU payback, building, staff, and a country risk charge. A summary then shows the base case, a combined downside, money needed up front, and deployment status, with break-even conditions and illustrative protection costs.

A compute spread is the revenue a site earns per installed GPU-hour minus the modeled costs of producing that hour, idle hours included. Revenue is the GPU rent earned in the hours the GPUs are in use; costs are electricity, GPU payback, building, staff and admin, and the country risk charge. A positive spread is what is left over; a negative spread is a loss.

For each economy and site size, the model builds a year-by-year path of GPU rent, utilization, power use, and costs over the GPU life, then levelizes it to one figure per installed GPU-hour. It solves for the utilization, GPU rent, and power price at which the site breaks even, and reruns the site under the downside case. The full methodology is proprietary.

No. GPU rent is the price a compute customer pays to use one GPU for an hour, and it is the site’s revenue. Land, the building, power and cooling equipment, and the grid connection are the owner’s costs, carried in the building cost line and paid back over the building’s 20-year life. Neither is counted inside the other.

Each size is grid power: PUE covers cooling and other facility load; the rest is IT load, divided by server power per GPU. At a PUE of 1.3, a 1 GW site has 769 MW of IT load and about 603,318 H100 GPUs, 100 MW about 60,332 and 1 MW about 603. A B200 site holds about 430 GPUs per MW. Hotter economies have a higher PUE and fewer GPUs.

The explorer covers 86 economies across the Americas, Europe, Asia, the Middle East, Africa, and Oceania: the economies the Virya Electric Grid Index scores. Each uses its own electricity tariff, construction and labor price levels, climate, and country risk. Select one on the globe, or type its name or two-letter code in the search box.

Each economy uses its industrial or large-commercial tariff, excluding VAT and recoverable taxes, from Eurostat size bands where published and from national regulators and utilities elsewhere. Where World Bank reliability data shows outages, that share of hours runs on diesel backup, priced at US$300/MWh; documented overrides apply where newer national data differs, up or down. Cooling and other overheads enter through PUE, and the electricity slider scales the grid tariff.

Each economy has a financing rate taken from the Virya Electric Grid Index; the page never displays it. The rate sets the cost of paying back the GPUs and the building, so a higher rate raises GPU payback and building cost per GPU-hour and narrows the spread. The Δ cost of capital slider shifts it up or down in percentage points to test the sensitivity.

Small sites carry more fixed cost per GPU. In the model a 1 MW site costs more per MW to build than a 1 GW site, and it needs 8 staff for about 603 H100 GPUs against 400 staff for about 603,318 (at a PUE of 1.3). Where Eurostat bands apply, a 1 MW site also buys power in a smaller, usually dearer band.

A site passes when its base-case spread beats zero by more than 10% of revenue and US$0.05 per GPU-hour, and stays above zero with GPU rent 20% lower and one year less GPU life at the same time. A base case within those margins, or one that fails that combined downside, is marginal; a bigger loss fails. A switch tests each stress alone, a looser rule.

The model’s list GPU rents and capital costs are calibrated to representative public figures, and a 1 GW site spreads building and staff costs over the most GPUs. Fewer pass the combined downside. The model does not assess whether a site can get a grid connection, permits, and committed demand; the delay input prices only the wait. A positive spread is a screen, not an investment case.

The workload sets utilization, GPU rent, and power draw. Training runs 15 points above base utilization but earns 80% of list GPU rent; fine-tuning runs 8 points above and earns 90%; inference runs at base, earns full list GPU rent, and draws less power. Base utilization is 70% everywhere, with a two-year ramp. Inference leading training follows from these presets, not from local demand evidence.

A contracted share of capacity is billed in full at a contract price, by default 85% of year-1 market GPU rent, for the term you set. A fixed price holds when market GPU rents fall, so the stress tests, break-even GPU rent, and GPU rent floor all treat that revenue as locked in. An indexed contract follows the market instead. Contracted capacity in use sets only power use.

The months until first revenue input delays the first billed GPU-hour. During the wait, market GPU rent keeps falling and the money already spent carries its financing cost; the utilization ramp starts only once compute is ready. The summary shows the longest delay the site can absorb before its margin turns negative. Delays are illustrative scenarios, not forecasts of local grid connection times.

Yes, as an experimental option. Its inputs hand a share of staff roles to humanoid robots, priced from a Japanese-built industrial humanoid listed at ¥18 million before tax, plus an import premium where humanoids are not built at home. Each card then shows the robot count and the margin with an all-human staff. Robots matter most at 1 MW, where staff weigh most per GPU, and little at 1 GW.

The illustrative risk lab in the summary PDF prices four floors (utilization, GPU rent, joint revenue, and resale), outage cover, and a power hedge for the selected site size, as benchmarks, not quotes. Its downside is the same combined downside as the screen. The country risk charge, customer credit plus political and transfer risk, is not insured separately. The risk engine is proprietary.

It applies when the site loses money at the assumed utilization or at today’s GPU rent, or when utilization does not raise the margin. An in-the-money floor can still be priced, with its intrinsic value in the premium; declining to price one on a loss-making site is this model’s pricing policy, not a rule of insurance. Such exposures are usually better addressed in financing or contracts.

Yes. Name a scenario, then copy its link: the link carries the inputs and the model version. Add up to five scenarios to a comparison, kept only in this browser. Each PDF is sent to your email once your details are verified, starts with the disclaimer, and charts the results: cost by site size, the stress tests, the risk lab, and year-by-year cash.

Exchange rates refresh twice a day, from European Central Bank (ECB) reference rates where the ECB publishes the currency and from ExchangeRate-API otherwise. Electricity tariffs, outage hours, price levels, wages, climate, GPU prices, and financing rates are reviewed periodically against their sources: the Virya Electric Grid Index, World Bank (including World Bank reliability data), ILOSTAT, Eurostat, IMF, Damodaran, and national regulators and utilities. Methodology details on request: contact@anshulsaxena.com.

The Explorer opens after you verify your email with a 6-digit code we send you. You then stay signed in on that browser for 7 days on all Virya sites. The privacy notice at anshulsaxena.com/privacy explains what is kept.

No. The explorer shows modeled scenarios, not investment advice. The figures come from public data, representative GPU rents, and stated assumptions, and they are a screening aid rather than a forecast, a valuation, or a quote.

Email contact@anshulsaxena.com with feedback, corrections, or questions. A correction to an economy’s data is most useful with its source named.