A report published this week by JLL, a global real estate services provider, projects that even as global data center capacity increases 15% this year, overall demand will far outstrip available supply.
As the number of artificial intelligence (AI) applications continue to exponentially rise, demand for graphical processing unit (GPU) capacity will far exceed data center capacity for many years to come, says Andrew Baston, head of data center research and strategy for JLL. In fact, AI by 2030 will represent just under half of all demand for data center capacity even as the number of data centers being constructed continues to climb, according to the report.
Data center construction capable of providing an estimated 10 Gigawatts (GW) is projected to break ground globally in 2025, with an additional 7 GW that is likely to reach completion. Based on this current pace of under construction and planned developments, the global data center market will likely expand at a baseline 15% to possibly 20%, the report concludes.
The primary challenge, however, is not so much the building of the data centers as much as it is connecting them to transmission lines, says Baston. It can take as long as four years or more to have high-capacity power lines extended to new data center sites. Most of this delay is associated with securing easements and regulatory approvals, notes Baston.
In fact, the location for a data center is now evaluated based on available power capacity and proximity to transmission lines, rather than just the price of real estate and total acreage, according to the report. Countries around the world are now competing with each other to provide the real estate needed to construct data centers for AI applications, says Baston.
Additionally, electrical utilities are now more selective when providing Purchase Power Agreement (PPA) approvals because speculators are artificially driving costs up in anticipation of a site becoming more attractive to an actual builder of a data center, notes Baston. “Land speculation is adding noise to the system,” he says.
The core challenge, however, remains the amount of power being consumed by GPUs. Higher system rack densities now result in anywhere from 40 to 130 kilowatts (kW) being generated per rack, with next generation GPUs expected to consume as much as 250 kW per rack, according to the report.
That level of power consumption will require liquid cooling and eventually immersion cooling that partially or completely embeds servers in some type of fluid, the report notes.
Naturally, there is a lot of interest in nuclear energy sources to enable additional data center development. Multiple companies are developing small modular reactors (SMRs), but they are not likely to become commercially viable until 2030, according to the report.
Despite that aggressive increase in data center construction, the report forecasts that data centers will represent only about 2% of global electricity consumption in 2025 as demand for electricity created by, for example, electrical vehicles, continues to increase.
It’s not clear to what degree access to power sources might adversely impact the pace at which AI applications are built and deployed, but the one thing that is certain is that the cost of building and deploying those applications will be significantly higher than the average enterprise IT organization might be anticipating.
