The Economics of an AI Data Center: What a Gigawatt Costs

A few years ago, AI projects were measured in parameters and GPUs. Now they’re measured in gigawatts. When companies describe their ambitions — the OpenAI–Nvidia 10-gigawatt partnership, China’s $295 billion infrastructure push — the unit of ambition is electrical power, not silicon. So it’s worth asking a concrete question: what does a single gigawatt of AI data center actually cost, to build and to run? The numbers are back-of-the-envelope, but the shape is what matters.

Why a gigawatt is the unit

A gigawatt is a billion watts — roughly the output of a large nuclear reactor, or enough to power a mid-sized city. Framing AI capacity in those terms is the tell: the binding constraint is no longer chips you can order, it’s power you can secure. A gigawatt is now shorthand for “a complete, frontier-scale AI campus.”

The build: tens of billions of dollars

Building a gigawatt-class AI data center runs on the order of $30–50 billion, and the surprise for newcomers is where the money goes. It’s not mostly the building.

• The silicon dominates. The single largest line item by far is the AI accelerators — the GPUs and custom chips. A top-end accelerator costs tens of thousands of dollars, and a gigawatt facility holds hundreds of thousands of them. That’s why a roadmap like Nvidia’s Rubin platform or AMD’s MI400 moves the entire industry’s capex math.
• Memory is its own scramble. Every accelerator is paired with high-bandwidth memory, and HBM is scarce and expensive enough to have set off the HBM4 supply race and a broader memory supercycle. It’s a multi-billion-dollar slice of the bill on its own.
• The physical plant. Buildings, substations, transformers, and — increasingly — liquid cooling, because these chips run far too hot for air. Networking gear to lash the accelerators together accounts for billions more.

The power bill: half a billion a year, minimum

Now run it. A gigawatt drawn continuously for a year is straightforward arithmetic:

1 GW × 8,760 hours = 8.76 billion kWh — about 8.76 terawatt-hours a year.

At an industrial electricity rate of roughly $0.06–0.08 per kWh, that’s on the order of $500–700 million a year just for electricity — and that’s before cooling overhead. Data centers draw more than their IT load; the ratio is captured by PUE (power usage effectiveness). Even an efficient modern facility at a PUE near 1.1–1.2 pulls 10–20% extra, pushing the all-in energy bill toward $600–800 million annually. Add the water many cooling systems consume, and the operating footprint becomes a regional issue, not just a line item.

The depreciation treadmill

Here’s the cost that wrecks naive models: these assets don’t last. AI accelerators are typically depreciated over just a few years, and they fall behind faster than that as new generations ship. A gigawatt of chips bought today is, in accounting terms, melting — losing a large fraction of its value annually. That turns the build cost from a one-time check into a recurring one: to stay competitive, you’re re-buying a big slice of that $30–50 billion every few years. It’s the same dynamic that drives the broader capex boom, and it’s why a memory or chip supplier’s stock can swing on every demand signal.

Why power, not money, is the ceiling

You can raise $50 billion faster than you can get a gigawatt connected to the grid. Interconnection queues run for years, transmission is constrained, and utilities can’t add capacity on an AI timetable. That’s why securing power has become a strategic act in itself — the driver behind nuclear and small-modular-reactor deals, behind-the-meter generation, and site selection chosen for substations rather than tax breaks. The capital is the easy part; the electrons are the hard part.

The bull and bear case

The bull case is that a gigawatt is a money machine. Filled with accelerators serving inference to hundreds of millions of users, it can generate revenue that dwarfs its operating cost — and whoever owns the capacity owns a moat competitors can’t quickly replicate.

The bear case is timing and utilization. The build cost is spent now; the revenue is expected later. If demand softens or monetization lags, you’re left with a depreciating, power-hungry asset whose economics invert fast. Infrastructure booms — railroads, fiber — have a long history of overshooting before they pay off. The capacity usually does get used; the open question is who’s holding the asset when the cycle turns.

What to watch

The tells are in the disclosures, not the demos: capex guidance, utilization rates, the revenue actually attributed to AI, accelerator and memory lead times, and — most of all — power availability. A gigawatt is the most expensive machine humanity routinely builds now. Whether it’s a generational asset or an overbuild comes down to a single ratio: how much useful work it does per dollar of silicon and per watt of power, before the depreciation clock runs out.