About
How NodalZen plans an AI data center
NodalZen is a free, open-source planner that follows an AI data center from the GPUs in a rack to the grid connection that powers them. Four steps share one project, so a decision in one, such as the site, updates the others: cooling, costs and carbon.
Who it is for
- Guided
- Students, newcomers and business users. Pick a typical cluster, answer a few plain questions, and get an explained summary.
- Explorer
- Analysts, product and sales teams. All four steps with live results, charts, a map and simulations.
- Expert
- Engineers, planners and researchers. Every assumption editable, formulas inline, and results as CSV.
All three modes use the same engine and the same project, so they give the same answers.
What it calculates
1. Size: power, cooling and backup
IT load comes from each system's published power and its average utilization. Heat is split between liquid and air cooling, and cooling efficiency changes month by month with the local climate, using NOAA climate normals. Transformer, UPS and PDU losses are added, giving power usage effectiveness (PUE): facility energy divided by IT energy. Transformers, switchgear, UPS, batteries and generators are sized for peak load in the hottest month.
2. Site: where to connect to the grid
Every US state is scored on electricity price, grid carbon intensity, interconnection queue wait, water stress and free-cooling hours, weighted by what matters to you. Missing data lowers a score's confidence rather than being filled in.
3. Source: equipment and lead times
Order-by dates are worked backward from your target operation date. The critical path is the equipment with the least slack: the items that decide when the site can open.
4. Run: energy and cost per token
Throughput comes from published MLPerf Inference results, or your own figure. From it come tokens per second, energy and cost per million tokens, and annual electricity cost and carbon.
Simulations then compare up to four scenarios, show which inputs move the results most, and test a grid outage, a late delivery and a phased build.
Where the data comes from
| Data | Source |
|---|---|
| GPU systems | Vendor datasheets and product pages (labelled vendor-reported) |
| Inference throughput | MLCommons MLPerf Inference v6.1 |
| Electricity prices | US EIA, Form EIA-861 (2025) |
| Carbon intensity | US EPA eGRID2023 |
| Interconnection queues | LBNL and GridTracker, Queued Up 2026 |
| Water stress | WRI Aqueduct 4.0 |
| Climate | NOAA 1991–2020 hourly climate normals |
| Equipment lead times | Wood Mackenzie and American Clean Power; US DOE |
| Defaults | LBNL 2024 data center report, DOE FEMP, ASHRAE, PNNL |
In the planner, every value shows its source and date, and every result shows the formula behind it. Catalogs are sorted alphabetically by vendor, and no vendor is preselected.
Questions
Is it free?
Yes. The code is open source under the Apache 2.0 licence, and the compiled data is available under CC BY 4.0, subject to each source's own terms.
Does it send my project anywhere?
No. It runs in your browser and saves your project there. There is no account, no cookies and no tracking. A share link carries the project inside the link itself.
Can I build from these numbers?
They are planning estimates for education and early-stage exploration, not engineering, financial or procurement advice. Check them with qualified professionals and vendors.
Does it work offline?
Yes. After the first visit, the planner and all its data work without a connection.