A small research team is uncovering the rapid expansion of artificial intelligence (AI) datacenters in the United States, using publicly available data and satellite imagery. The report by 404 Media highlights the work of Epoch AI, a non-profit research institute dedicated to understanding the scale and pace of AI development. Their project is particularly significant as it provides transparency in an industry that is expanding faster than public discourse allows.
Mapping the Hidden Infrastructure
Datacenter construction has emerged as a contentious issue across the country due to their vast energy and water requirements. Many communities are often unaware of these facilities until construction is well underway. Epoch AI has created an interactive map that visually represents these datacenters, linking each marker to satellite images and project specifics. For instance, a green circle on the map indicates Meta’s “Prometheus” datacenter complex in New Albany, Ohio, which has already cost an estimated $18 billion and consumes 691 megawatts of power.
The map not only identifies the locations of these datacenters but also allows users to explore their development over time. Through satellite imagery, users can observe how the complex evolves, with new buildings and cooling systems added as the demand for AI capabilities increases.
Understanding Power Consumption Estimates
A significant aspect of Epoch AI’s analysis focuses on the cooling systems used in modern AI datacenters, which generate substantial heat. The team employs open-source intelligence to assess the cooling units, often located outside buildings or on rooftops. According to the organization, “Modern AI datacenters generate so much heat that the cooling equipment extends outside the buildings.”
By counting fans and measuring their size, researchers can develop estimates of energy consumption. This data is essential for inferring the compute capacity and construction costs of these facilities. Jean-Stanislas Denain, a senior researcher at Epoch AI, emphasized the importance of cooling analysis, noting, “We focus on cooling because it’s a very useful clue for figuring out the power consumption.” However, the model does carry uncertainty, as cooling configurations can vary widely, leading to estimates that may be significantly higher or lower than actual consumption.
Despite the comprehensive nature of the mapping project, it remains incomplete. Variations in state and local disclosure laws, as well as the tendency for some projects to operate without publicity, create blind spots. Epoch AI estimates that the current dataset represents approximately 15 percent of global AI compute delivered by chipmakers as of November 2025.
Markers on the map also indicate other significant projects, such as xAI’s Colossus 2 project near Memphis, Tennessee. Notably, the company has installed natural gas turbines across the Mississippi border, likely to facilitate faster regulatory approvals. This particular project reportedly has 110,000 NVIDIA GB200 GPUs operational, according to earlier communications from Elon Musk.
Even with detailed mapping, significant gaps in understanding remain. Epoch AI acknowledges that while they may have a thorough analysis of a datacenter’s infrastructure, the actual usage and the entities involved may remain unclear. The organization’s goal is to expand its search globally, aiming to bring visibility to critical infrastructure that shapes the future economy, often hidden from public scrutiny.
