Bloomberg Breaks Down the Cost of an AI Data Center in Texas

Bloomberg has published an exclusive breakdown of the estimated costs of building an AI data center in Texas. The takeaway: developers are budgeting billions of dollars per project, and the reason lies in the structure of the estimate, not in a single expensive line item. We do not have access to the line-by-line figures from the announcement, so below we walk through the cost logic using public industry estimates. Nvidia CEO Jensen Huang has cited a benchmark of $50–60 billion per gigawatt of capacity, of which about $35 billion goes to his company's hardware.

Where the money goes

The heaviest line item is compute hardware. A GB200 NVL72-based rack costs around $3 million by market estimates, and a large campus holds thousands of them. Accelerators, servers, switches and optics take more than half of the budget. The building, substations, power generation and mechanical and electrical systems split the remainder.

The second-largest group of costs covers everything that turns electricity into compute. A modern AI rack draws more than 100 kW, and air can no longer handle that much heat. Operators need liquid cooling, chillers, complex hydraulics and backup power. Every megawatt of IT load drags along additional cooling capacity and losses, so builders have to overprovision.

Why Texas

The state attracts developers with cheap land, fast permitting and access to gas and wind generation. The largest AI campuses, including Stargate in Abilene, are being built there. But this advantage has a downside. ERCOT's grid infrastructure is running at its limit, the queue for connecting large consumers is long, and new rules for loads of 75 MW and above add disclosure and load-management requirements. As a result, some projects build their own generation, which adds hundreds of millions of dollars on its own.

Land is almost invisible in the budget. What costs money is everything that has to be brought to it: high-voltage lines, substations, gas pipelines and water for cooling systems.

Cost benchmarks

• Total cost: $50–60 billion per 1 GW, by Huang's estimate.
• Nvidia's share: about $35 billion per gigawatt.
• GB200 NVL72 rack: around $3 million, with density above 100 kW.
• Other line items: building, substations, power generation, cooling, networking.
• Main risk: the time it takes to connect to the grid, not the price of concrete.

How to read these estimates

When evaluating data center announcements, first check what the number covers: the cost of the building and infrastructure, or the full budget including hardware. The difference is several-fold, and it is why figures from different companies look incomparable. The second question is whether the grid connection is secured and whether the project has its own generation. Without that, the billions remain commitments on paper.

For compute renters, the conclusion is simple. The price per GPU-hour rests on capital costs and the price of electricity, so don't expect a noticeable price drop without growth in available capacity. For investors, the key metric is the payback period of the hardware, which becomes obsolete faster than the building around it.

For the detailed line-by-line breakdown, see the full version of the Bloomberg piece.