The GenAI GPU shortage has already sparked a surge in demand, elevated prices, and diminished availability. Nonetheless, one other urgent situation is looming: knowledge facilities are working out of house and energy. That is notably problematic for small firms offering high-performance computing (HPC) colocation companies, who’re discovering present knowledge facilities maxed out.
A latest report from JLL, an actual property funding and administration agency, highlights the AI-driven development is anticipated to proceed, with knowledge era predicted to double over the subsequent 5 years.
Moreover, knowledge middle storage capability is projected to develop from 10.1 zettabytes now to 21.0 zettabytes in 2027, necessitating extra knowledge facilities. The facility calls for of generative AI, estimated at 300 to 500+ megawatts per campus, may also necessitate extra energy-efficient designs and places.
Energy grids are reaching capability
In keeping with the report, the design of AI-specialized knowledge facilities differs considerably from typical services, requiring operators to plan, design, and allocate energy assets primarily based on the kind of knowledge processed or stage of GenAI improvement. With the massive improve in GPUs, present requirements for warmth removing will likely be surpassed, prompting a shift from conventional air-based cooling strategies to liquid cooling and rear-door warmth exchangers.
Chatting with HPCwire, Andy Cvengros, managing director of U.S. Knowledge Heart Markets for JLL, emphasised the significance of planning. He defined that energy grids are reaching capability, and transformers have lead instances exceeding three years, necessitating innovation. The GPU squeeze is affecting small colocation deployments of 4-5 racks, who’re discovering it more and more troublesome to safe knowledge middle house because of the calls for of hyperscalers.
Cvengros additionally highlighted that each one main metro areas are basically maxed out, making secondary areas like Reno, NV, or Columbus, OH, prime places for brand spanking new knowledge middle development. Nonetheless, the demand is anticipated to proceed, and new knowledge facilities are 3.5 years out.
The worldwide GenAI power demand presents each alternatives and challenges. Discovering GPUs for HPC is simply half the issue; as HPCwire factors out, the place to plug them in might turn out to be a much bigger problem. This situation is especially difficult for smaller operators, who could also be pushed out of the market by the competitors for assets.
