We already know which will be the most expensive data center in the world. If Bill Gates paid it, it would be almost zero

Already in 2024 we saw that infrastructure spending for AI was being insane. The trend has not relaxed, quite the opposite. Big tech continues to burn money as if there was no tomorrow (literally) and most of that spending is going to most valuable asset in the AI ​​race: data centers. How much do they really cost?

Data centers in numbers

Epoch AI has published Frontier Data Centersa complete database about data centers being built in the United States. Through satellite images, public documents and permits, they have obtained information about the estimated construction cost, as well as energy consumption and computing power.

The award for the most expensive data center goes to Microsoft Fairwater, whose total cost It could reach $106 billion when completed in 2028. To put it in context, Bill Gates’ fortune is estimated to be 107 billion dollars. It would be fair to pay it. The forecast for Microsoft Fairwater even surpasses Meta Hyperion, the data center that It will be as big as the island of Manhattan which would cost 72,000 million.

Next on the list is Colossus 2, by xAIwhose estimated cost is 44 billion dollars. It is closely followed by Meta Prometheus with 43 billion and the Amazon and Anthropic data center in New Carlisle with 39 billion.

Epoch AI has collected more data, such as how much computing power each facility will have. This data is measured using the NVIDIA H100 GPUs for reference. They have also calculated the energy demand and who will be the main user of each of them. Below we leave you a table with the key information:

Dizzying climb

Looking at the case of Microsoft Fairwater, and always according to Epoch AI’s forecast, in March 2026 the investment will be $18 billion. A year later, in February 2027, it rises to 35,000 million, just four months later it shoots up to 71,000 million, to reach 106 billion in 2028.

The price increase is dizzying and responds to several factors. The first is that the computational cost of training models has been increasing. For example, GPT 4 cost OpenAI over 100 million and rumors before the release of GPT-5 pointed to training rounds of 500 million each. Epoch AI also did an analysis on this and they estimated that the cost of training has multiplied by 2.6 year after year.

On the other hand, there is the demand for GPUs, necessary for training the models and the most expensive component of all. An NVIDIA H100 GPU costs 25,000 dollars and its successor, the NVIDIA B200 also known as Blackwell, could be between 30,000 and 40,000 dollars. And this is just the GPUs, many are needed more components to get a data center up and running, such as power generators, high-speed networks or refrigeration, among others.

The initial bottleneck was the shortage of GPUs, but it has been overcome by a more fundamental constraint: there is not enough power for so many chips. data centers They consume a lot of energy, Seriously, a lot. To put it in context, in 2024, data centers were already the 4% of United States electricity consumption and it is expected that Demand will double in the next five years.

Nobody wants to live near a data center for one reason: mass consumption is raising energy prices up to 267% in nearby areas. Power supply has become a new choke point for the industry. Microsoft is already considering producing its own energy by creating nuclear power plants and others like Google and Amazon are considering taking data centers into space.

Image | Microsoft

In Xataka | AI data centers are an energy hole. Jeff Bezos’ solution: build them in space