The GenAI GPU shortage is already causing a spike in demand, increased costs, and reduced availability. But another pressing problem looms: data centers are running out of space and power. This is especially problematic for small and medium-sized businesses offering high-performance computing (HPC) colocation services, where today's data centers are stretched to their limits.
A recent report from JLL, a real estate investment management company, predicts that AI-driven growth will continue, with data generation predicted to double over the next five years.
Additionally, data center storage capacity is expected to increase from 10.1 zettabytes today to 21.0 zettabytes by 2027, necessitating the need for more data centers. The power demand for generative AI is estimated to be 300-500 megawatts or more per campus, which also requires more energy-efficient designs and locations.
The power grid is reaching capacity
According to the report, the design of AI-specific data centers will differ significantly from traditional facilities, with operators planning, designing, and allocating power resources based on the type of data being processed and the stage of development of GenAI. There is a need. The massive increase in GPUs will exceed existing heat removal standards, prompting a shift from traditional air cooling to liquid cooling and rear-door heat exchangers.
In an interview with HPCwire, Andy Kubenglos, JLL's managing director for the U.S. data center market, emphasized the importance of planning. He explained that the power grid is reaching capacity, transformer lead times are over three years, and innovation is needed. The GPU squeeze is impacting small colocation deployments of 4-5 racks, which are finding it increasingly difficult to secure data center space due to the demands of hyperscalers.
Cvengros also emphasized that all major metropolitan areas are essentially at capacity, with secondary regions such as Reno, Nevada and Columbus, Ohio becoming prime locations for new data center construction. However, demand is expected to continue and the new data center will take him three and a half years to complete.
Global GenAI energy demands present both opportunities and challenges. Finding GPUs for HPC is only half the problem. As HPCwire points out, where to connect them can be a big challenge. This problem is particularly difficult for small operators, who may be forced out of the market by competition for resources.