The honest answer is "it depends on what you build", so here is the spread for a 50 MW project in Texas, before GPUs, from the SCR AI Build Configurator dataset:
| Build type | Per MW | 50 MW total | Build + grid |
|---|---|---|---|
| Hybrid mining + AI | US$6.1M | US$305M | 16 months |
| Bitcoin miner retrofit | US$10.0M | US$500M | 20 months |
| Powered shell (lease the megawatt) | US$11.3M | US$564M | 30 months |
| Modular / prefabricated | US$12.2M | US$610M | 11 months |
| Neocloud / GPU-as-a-service | US$14.3M | US$715M | 20 months |
| Edge / micro deployment | US$20.8M | US$1.04B | 8 months |
| Hyperscale AI factory | US$23.5M | US$1.18B | 38 months |
Then decide who buys the GPUs. B200-class accelerators add about US$30M per MW, which is why a 100 MW AI factory in Georgia models at US$2.35B for the building but US$5.33B with GPUs. For five fully worked projects, see the cost to build a data center, 5 MW to 100 MW.




