Data center · Cost estimator
What it costs to build and run a data center
Start from an AI or a conventional build, then move any of about 30 inputs. United States, 2026 dollars, land excluded. Defaults come from published 2025 and 2026 benchmarks; inputs with weak sourcing are marked so you can replace them with your own quotes.
Where the build money goes
What a year costs
Capital is spread over each asset's life at your cost of capital, then running costs are added. This is the figure to compare with rental or cloud pricing.
Cumulative spend over the facility's life
How your estimate compares with published projects
Each dot is an announced project or analyst model, in millions of dollars per MW, converted to IT load where the source allows. Developer announcements appear to cover the facility only, though only Meta and Cushman & Wakefield say so; program totals include chips. Hover or tap a dot for its name and basis.
Show the projects as a table
Per-GW headlines often count facility power, which is 1.2 to 1.5 times the IT load this estimator uses. Entries marked "power basis not stated" may therefore read low against your marker.
Where the defaults come from
Every default below is a starting point, not a quote. "Assumption" means no adequate public source exists and the value is a labeled judgment; replace it with your own figure where you have one.
Why the total looks large
Most published "cost per MW" figures of $10M to $18M cover the facility only. For an AI build, chips, networking and storage add roughly 1.5 to 3 times that again. Capacity here is IT load; per-GW headlines often count grid power, which is 1.2 to 1.5 times larger.
Platforms left out
No usable system price is published for NVIDIA's HGX Rubin NVL8, AMD's MI355X and MI440X, or Intel's Gaudi 3 PCIe card and Crescent Island, so no cost per MW can be calculated. NVIDIA's L4 is a low-power add-in for ordinary servers, where the host dominates the cost. Hyperscaler chips such as AWS Trainium are not sold on the open market.
The AI facility premium is disputed
The defaults put an AI facility about 50% above a conventional one per MW, in line with Cushman & Wakefield's 2026 all-in average. Turner & Townsend, comparing projects of similar IT capacity, finds a premium of only 7 to 10%. The conventional default rests on JLL's global shell-and-core average, the nearest published figure.
What is not modeled
Land, demand charges and minimum-take clauses in large-load tariffs, water, bandwidth, software licences, real-property tax on the facility, interest on IT equipment before it earns, spare capacity for on-site generation, and price changes when IT is replaced. Networking, storage and CPU servers are added as cost without taking power away from the accelerators, which overstates IT cost by perhaps 5 to 10%.
How region is applied
The regional index is derived from RSMeans data center building costs per square foot. It is applied in full to the building and at half strength to electrical and cooling, because most of that equipment is priced nationally. That weighting is an assumption.
How scale is applied
No published scale curve exists. The estimator interpolates between 1.35x at 1 MW and 0.90x at 1 GW around a 40 MW reference. Turn it off to use flat per-MW rates.
Interest during construction
Calculated, not sourced: the facility cost financed at your cost of capital and drawn evenly over the build period, as simple interest. IT equipment is assumed to arrive at the end and carries none.
Benchmarks were gathered and re-checked against their sources on October 7, 2026. This is an estimate for early planning; get quotes before committing money. How the numbers were sourced