The sound inside a big data center is steady and tactile, like a wall of fast-moving fans; it’s the kind of noise that reaches your chest instead of simply your ears. The air is dry and chilly. Server rack rows vanish into the distance. The infrastructure is unglamorous. However, the air cooling architecture that characterized data centers for decades is no longer sufficient when you’re running the AI workloads that AWS is currently doing at scale. The chips are just too hot.
The density of heat produced by an AI GPU is too high for conventional air systems. Modern AI accelerators, such as Nvidia’s H100 series, might draw several kilowatts each, packed into racks with dozens of identical chips, whereas a typical web server could only draw a few hundred watts. Up to a certain point, blowing cold air past them is effective. After that, you need something in close proximity to the hardware, such as liquid, which is significantly more effective than air at conducting heat away from a surface.
Instead of using a commercial source, AWS developers created their solution internally. In order to remove heat from high-density GPU clusters without taking up the floor space that third-party systems usually use, the In-Row Heat Exchanger is a closed-loop liquid cooling unit that is mounted directly to server racks. When you operate at the scale that Amazon does, the mechanical efficiency improvement of 46% is not an abstract figure. Instead of being used for cooling overhead, that released power is put back into computation.
The tangible manifestation of where this is all going is Project Rainier. The $11 billion Indiana complex, which debuted in 2025, was built with heavy AI workloads in mind from the start. It supports about 500,000 of Amazon’s own Trainium chips, employs both external air and water cooling, and houses workloads for Anthropic, an AI safety startup in which Amazon has invested billions. The size of what AI infrastructure currently needs is extremely tangible when you walk around the perimeter of a campus like that, which spans hundreds of acres and pulls power at a level comparable to a small city.
The next stage is Louisiana. In acknowledgment of the fact that data centers of this magnitude have an impact on the communities surrounding them as well as the businesses that construct them, Amazon has committed $12 billion to Caddo and Bossier Parishes, with $400 million designated exclusively for local water and utility infrastructure. Particularly in areas where both are under strain, controlling water consumption is becoming just as apparent a priority as controlling energy consumption.
The most challenging aspect is the energy equation. Up to 10 times as much electricity is used by AI servers as by traditional web servers, and renewable energy alone isn’t always able to supply that demand with the dependability that a production data center needs.
Amazon has been open in investigating nuclear power, particularly small modular reactors that offer dense, low-carbon power output in a smaller footprint than a conventional plant but are still in the early stages of commercial development. It’s still unclear if that schedule corresponds with AWS’s infrastructure buildout. However, the fact that a tech company is actively preparing for nuclear capacity reveals something about the scope of the project.
