How much does it cost to build a 1GW AI data center? A $38 billion entry fee, with 60% used to buy GB200.

Question

AIMPACT news, May 15 (UTC+8). According to Beating Monitoring by Dongcha, AI research organization Epoch AI has calculated the ownership costs of a typical 1GW capacity AI data center. Building such a data center requires an upfront capital expenditure of $38 billion, plus an annual daily operating expense of $900 million. If the capital expenditure is amortized over the asset’s lifespan, the data center’s annualized total cost is $8.5 billion. Among these, servers (with the model assuming that all use NVIDIA GB200 NVL72 systems) make up the absolute largest share, with annual depreciation costs reaching $5 billion—accounting for 60% of total spending. In comparison, daily operating expenses are negligible on the ledger. Even the energy costs, which account for the largest portion, are only $600 million per year. This breakdown is highly sensitive to the depreciation period of the hardware. Epoch currently assumes an IT equipment lifespan of 5 years and a data center facility lifespan of 14 years. If the IT equipment lifecycle is shortened to 3 years, the data center’s annualized total cost will surge to $12 billion; if extended to 7 years, it can be reduced to $7 billion. (Source: BlockBeats)

AlphaAfterTea · Accepted Answer

Only 600 million of the 900 million operational expenses are for energy? Are manpower and bandwidth costs being eaten up?

ReflectiveKey · Answer

If the data from BlockBeats is accurate, then all AI valuation models would need to be recalculated now.

TransparentDomeCity · Answer

3-year lifespan vs 7-year lifespan, an annualized cost difference of 5 billion, depreciation policy is the profit switch

GateUser-423f10e3 · Answer

Server depreciation accounts for 60%, and the hidden costs of AI infrastructure are too high.

Miner&apos;sOldKeyboard · Answer

What is the amount per watt-hour when annualized at 8.5 billion? Has anyone calculated it?

SilverCubeInsomnia · Answer

This is not yet the one-time cost of model training; inference is just the tip of the iceberg.

MoonlightShellPool · Answer

Capital expenditure of 38 billion, how many tokens need to be sold to break even?

GateUser-f49a50d4 · Answer

The fluctuation range of 7 to 12 billion leaves too much room for storytelling to investors.

PerpNightwatch · Answer

As the chip iteration speed slows down, a 7-year depreciation period becomes reasonable; now, even 3 years feels too long.

StakingDaydreamer · Answer

1GW scale, equivalent to the electricity consumption of a small city, and energy costs are surprisingly not the main expense?

How much does it cost to build a 1GW AI data center? A $38 billion entry fee, with 60% used to buy GB200.

Trending Topics

TradfiTradingChallenge

PlatinumCardCreatorExclusive

DailyPolymarketHotspot

GateSquarePizzaDay

SpaceXOfficiallyFilesforIPO

Pinned