The insatiable appetite for data in our digital world is forcing a reassessment of the way data centers are designed, built, and operated. 

The power crunch—intensified by the colossal energy demands of AI—is stretching the existing data center infrastructure to its limits.

Cummins Power Generation is working in lockstep with the data center sector to provide power solutions that can address the high intensity and unpredictable fluctuations associated with AI.

Average Rack Density Rising…But Slowly

The most common rack density range, according to Uptime Institute's Global Data Center Survey 2025, is 5-9 kW per rack. The figure has remained fairly constant over the past five years, and is a long way short of the widely predicted surge to 20kw+ for the industry in general.

Still, Uptime expects rack densities will rise steadily over the coming years as more data centers adopt servers in the 10kw-30kw range. 

The big question for data center operators will be whether their existing infrastructure can handle the future capacity and power demands of upcoming hardware generations, like new NVIDIA GPU systems, or if they even want to go down that route knowing it will be hugely expensive.

The introduction of Nvidia’s Blackwell GPUs and GB200NVL72 rack designs in 2024 saw peak rack power density rise to132 kW, and future models like Blackwell Ultra and Rubin AI servers will require between 250 and 900 kW, with up to 576 GPUs per rack by 2026-2027. 

Top-end AI servers are projected to break through the mythical 1MW barrier with the arrival, slated for 2028, of Nvidia’s Rubin Ultra AI GPU and HBM4 memory. 

Such advancements are underscoring the expansion of data center campuses from tens of megawatts to hundreds of megawatts, with some planning for gigawatt-scale power consumption.

Big Names Making a Move

At the recent Pennsylvania Energy and Innovation Summit, several major players in data center and energy infrastructure announced huge and symbolic investments: 

Blackstone and PPL Corporation announced a $25B joint venture investment to develop natural gas-fired plants tailored for data center needs; CoreWeave is launching a $6B data center project to meet GPU/AI demand; Google and Brookfield Renewable are working to secure 3GW of hydro power to ensure sustainable supply for their infrastructure; and Amazon Web Services (AWS) will spend $20B on two data center complexes in Pennsylvania, including one it is building alongside a nuclear facility with the idea that it will essentially plug directly into the power plant. 

Picking up the nuclear theme, engineering giant Rolls-Royce is confident the small modular reactors (SMRs) it builds for nuclear submarines can be adapted to play a role in the data center business, notwithstanding concerns over the supply of water for the reactor’s cooling systems.

With major data center operators recognizing the gargantuan need for power for the AI ecosystem, the next challenge becomes how to deliver that power to the servers demanding it.

Higher Voltage, Greater Efficiency

In a conventional data center setup, AC power is converted to DC, and from that point there can be multiple DC-DC conversions (ie 300V down to 48V). At each conversion point, energy is lost and heat generated, increasing operational costs and environmental impact. 

Centralizing the rectification process means converting AC to DC closer to the source, then distributing high-voltage DC directly to the racks, reducing losses and boosting Power Usage Effectiveness (PUE).

It is widely anticipated hyperscalers like Meta, Google and Microsoft will move to deploy medium voltage (MV) distribution – up to 13.8kV – and higher DC voltage 400VDC and 800VDC architectures to take advantage of the greater power density of AI server racks. 

With the lower currents required, they can expect to recover huge amounts of previously lost energy while enjoying significant savings in the amount of copper needed for cabling. 

Moreover, HV architecture also allows for more power to be delivered in a smaller physical footprint, freeing up precious rack rack space and increasing the computing density of the data center. 

While companies like NVIDIA are driving the development of new ecosystems for 800VDC, there is a need for a wider range of commercially available hardware, including solid-state transformers (SSTs), high-efficiency DC-DC converters, and specialized power shelves.