AI research institute Epoch AI crunched the numbers on what it would cost to build and run a typical 1-gigawatt AI data center. The upfront capital expenditure: a cool $38 billion. Annual operating expenses: another $9 billion.
Spread that capital over the expected life of the assets, and the annualized total cost comes to $8.5 billion. The servers — which the model assumes are all NVIDIA GB200 NVL72 systems — are the elephant in the room, accounting for $5 billion in depreciation per year, or 60 percent of the total.
By comparison, operating expenses look almost like a rounding error. Even the biggest cost, energy, is just $6 billion annually.
But these estimates are highly sensitive to how long the hardware lasts. Epoch currently assumes IT equipment has a 5-year lifespan and the facility itself lasts 14 years. If you shorten IT equipment life to 3 years, the annualized total cost soars to $12 billion. Stretch it to 7 years, and it drops to $7 billion.