Where the $60 Billion Goes in One AI Data Center

One gigawatt of AI data center now costs about $60 billion. That is Nvidia's own number: on its August 2026 earnings call, Jensen Huang said a gigawatt went from about $30 billion five years ago to $60 billion today (see Sources, [1]). He also said that with Vera Rubin, Nvidia's revenue opportunity is about $40 billion of every gigawatt.

So where does the rest go? This article opens the bill line by line: the chips, the network, the power gear, the building, and the costs that keep coming after it opens.

What a gigawatt buys

A gigawatt is enough to power roughly a million homes, according to Epoch AI, which expects five AI data centers of a gigawatt or more to come online in 2026, each run by a different company (see Sources, [5]).

Estimates of the price differ, mostly because of which chips go in and who is counting:

Sources: [2], [4], [3], [1]. The honest answer is somewhere between $35 billion and $60 billion.

The chips

In Bernstein's breakdown, GPUs alone are 39% of the bill; nothing else comes close (see Sources, [2]). But Nvidia also sells the CPUs, the networking and full rack systems, so Huang counts its share per gigawatt in total: about $18 billion with Hopper, about $25 billion with Grace Blackwell and about $40 billion with Vera Rubin (see Sources, [1]).

That helps explain why Nvidia's data center revenue reached $89 billion in a single quarter (see Sources, [1]).

One rack

Foxconn assembles these systems. Its chairman said one Vera Rubin rack costs about $9.1 million, and a one-gigawatt site holds about 3,557 of them (see Sources, [3]). Multiplied out, the racks alone come to roughly $32 billion (our arithmetic).

Bernstein priced the earlier GB200 NVL72 rack at about $5.9 million: $3.4 million of compute hardware and $2.5 million of physical infrastructure around it (see Sources, [2]).

The network

Thousands of GPUs have to behave like one machine, which takes switches, optical links and a lot of cable. Bernstein puts networking at roughly 13% of the bill, the second-biggest single item after GPUs. Epoch AI's model puts networking at about $4.9 billion per gigawatt (see Sources, [2], [4]). CPUs are about 3% and storage only 1% to 1.5%.

Power and cooling

Bernstein says mechanical and electrical systems take close to one third of total spending (see Sources, [2]):

The building itself is surprisingly cheap. JLL forecasts construction at about $11.3 million per megawatt in 2026 (see Sources, [7]); for a full gigawatt that is about $11 billion (our arithmetic). Epoch AI's model gets almost the same facility cost, about $11.4 billion (see Sources, [4]).

The bills that keep coming

At 15 cents per kilowatt-hour, Bernstein estimates about $1.3 billion of electricity a year for a gigawatt site, and Foxconn's chairman gave the same figure. Epoch AI assumes cheaper power and gets about $0.6 billion (see Sources, [2], [3], [4]).

Electricity is not the biggest yearly cost. Foxconn's chairman said hardware depreciation runs about six times the electricity bill (see Sources, [3]), which is nearly $8 billion a year by our arithmetic. Epoch AI, spreading the IT equipment over five years, puts the total cost of ownership at $8.5 billion a year, with servers 60% of it (see Sources, [4]). AI chips behave less like a one-time purchase and more like a subscription.

Now multiply

Nvidia's CFO said OpenAI's existing and planned commitments represent about 12 gigawatts of Nvidia compute (see Sources, [1]). At $60 billion a gigawatt, that would be roughly $720 billion (our arithmetic), for one customer.

For a single real site, Meta's Hyperion campus has about $27 billion of development costs, and funds managed by Blue Owl own 80% of it (see Sources, [8]). Worldwide, JLL expects data center capacity to nearly double from 103 GW to 200 GW by 2030, needing up to $3 trillion of investment over five years (see Sources, [6]).

So where does the $60 billion go?

On Nvidia's own numbers, about two thirds goes to one chip company, for the chips, the CPUs and much of the network. Power gear and cooling take a big slice; the building is a small one. After it opens, billions more go out every year for electricity and for chips that age fast.

Frequently asked questions

How much does a 1 GW AI data center cost?

Between about $35 billion and $60 billion, depending on the chips and who is counting. Bernstein put a Grace Blackwell-era gigawatt at around $35 billion, Epoch AI's model at $38 billion up front, Foxconn's chairman a Vera Rubin site at up to $47 billion, and Nvidia's Jensen Huang at $60 billion today (see Sources, [1] to [4]).

What is the single biggest cost?

The GPUs. Bernstein puts them at 39% of total spending, three times the next item, networking, at roughly 13% (see Sources, [2]).

How much does it cost to run each year?

Electricity is about $1.3 billion a year at 15 cents per kilowatt-hour, according to Bernstein and Foxconn's chairman, or about $0.6 billion with Epoch AI's cheaper power assumption. Epoch AI puts total annual operating expenses at $0.9 billion, but counting the up-front spending spread over its useful life, the full cost of ownership is $8.5 billion a year (see Sources, [2], [3], [4]).

Why does the hardware cost so much every year?

Because it goes out of date fast. Epoch AI assumes a 5-year life for IT equipment and 14 years for the facility (see Sources, [4]), and Foxconn's chairman said hardware depreciation runs about six times the electricity bill (see Sources, [3]).

For the financing side of this spending, see Who Is Actually Paying for the AI Boom?. To estimate power costs for your own numbers, try the AI data center power and cost calculator.