The purpose of this final installment is not to predict, but to describe. Here is what is already being designed, funded, and built, the developments that will shape data centers over the next several years, based on what is on the table today.
Nuclear Power Comes Back
The most striking shift is in how these buildings will be powered. Every major technology company has now signed deals for nuclear electricity to feed its AI data centers, more than a dozen projects committing roughly ten gigawatts in the past year alone. Some are restarts of existing reactors: Microsoft has contracted to revive a unit at Three Mile Island, expected to deliver power around 2027 to 2028. Others bet on small modular reactors, a newer, factory-built kind of reactor that is smaller and faster to deploy than a traditional plant. Google, Amazon, and Meta have all backed small modular reactor projects, though most are not expected to deliver power until after 2030. It is a striking reversal for an energy source that had been in retreat for decades, driven now by technology companies that need enormous amounts of reliable, low-emission power and are willing to sign long contracts to get it.
Chips Keep Getting Denser
The hardware is not standing still. Each new generation of AI chip packs more power, and produces more heat, than the last. The most advanced AI racks shipping today draw around 132 kilowatts, and the next generations already on the public roadmap are designed to draw far more, with some future racks aimed past a megawatt each. This steady climb in density is the engine behind almost every other change in this list, because each more powerful chip demands more power to run and more effort to cool, and every step up forces the rest of the building to be rebuilt around it.
Cooling Goes Liquid by Default
Because of that density, the cooling methods described in Part 2 are quickly becoming the default. Direct-to-chip liquid cooling and full immersion, once rare, are now being designed into new AI facilities from the start. The expectation across the industry is that air cooling alone will soon be the exception rather than the rule for AI hardware, and some operators are going further, capturing the heat their buildings produce and reusing it rather than simply throwing it away.
Power Gets Closer and Smarter
Alongside nuclear, on-site generation and large batteries are being built directly at data centers to get around the grid wait, and operators are signing long-term deals for wind, solar, and gas together. The emerging model is a data center that draws from several sources at once and can lean on its own generation when the grid is stretched. The grid itself is being upgraded in parallel, with new transmission and large battery installations, though that work tends to move more slowly than the data centers waiting on it.
Bigger, and Built Differently
The scale keeps climbing. Individual campuses approaching a gigawatt of power, enough for a small city, are under construction, and the largest operators now design and build them as single integrated systems rather than one building at a time. At this scale, building a data center starts to resemble building a power plant and a small factory at the same time, and global capacity is on track to roughly double by 2030.
Efficiency Moves to the Center
As the footprint grows, so does the engineering effort behind it. Operators are cutting water use through reclaimed sources and closed-loop cooling systems that recirculate rather than draw fresh supply, capturing waste heat and routing it to warm nearby homes and offices, and matching more of their power consumption with clean generation. Water use in particular has become a specific design target, with newer facilities engineered to use a fraction of what older designs required. Each of these efforts pays for itself in lower operating costs over time, which is part of why they are spreading as fast as the hardware they support.
What It Comes Down To
None of this is speculation; it is the set of projects and designs already in motion. The through-line is consistent with the whole series: computing keeps growing, power remains the binding constraint, and the industry is reaching for nuclear reactors, denser chips, liquid cooling, and gigawatt-scale campuses to keep up. Where it settles is the story Blueprint will keep following.
This is the fifth and final article in Data Centers 101, a Blueprint Data Centers series explaining how modern data centers work, from the hardware inside them to the power that runs them. Blueprint is an independent data center platform developing greenfield data centers designed with flexibility to support a range of use cases including high-performance computing, AI and other advanced workloads. Follow Blueprint for more on the infrastructure behind the digital economy.
