at peak, while inference (answering user queries) consumes 10 times more power than traditional web searches. The impact is significant enough that the International Energy Agency (IEA) now tracks AI data centers separately, alongside steel, cement, and aviation. AI is becoming a defining global energy load at an unprecedented speed, reshaping how power is generated and delivered worldwide. THE GAUNTLET FROM CONCEPT TO LIVE SITE Even when the budget and ambition exist, turning a concept into a live facility requires running a gauntlet. Launching a modern AI data center can take three to six years. Most of that time is not construction—it is planning, paperwork, and approvals. Site selection is no longer as simple as finding the land. The location must have access to high voltage
delivery and optical fiber backbones that can grow with the site. Then comes permitting—environmental studies, zoning, community engagement, and shifting political winds. Navigating layers of local, state, and sometimes even federal government can easily stretch out over one or two years. Often, the most stubborn hurdle is grid interconnection. Utilities face their own constraints, and interconnection queues often run two to five years. High-capacity transformers and custom switchgear have lead times that can stretch beyond 48 months. EVOLVING ENERGY MIX As AI data centers continue to grow, so does the pressure to keep them running on clean, reliable power. Renewables are no longer a “nice to have” —they are the baseline. Most operators lean on power purchase agreements (PPA) with large wind and solar
farms. But there is a catch: renewables do not work on the data centers’ 24/7 schedule. That is where battery energy storage systems (BESS) become essential. BESS absorbs excess solar or wind when nature is generous, then releases it when the grid is stressed or demand spikes. Pairing on-site generation with storage increases resilience and reduces dependency on the local grid. Natural gas still forms a critical backbone for data centers, especially where renewable capacity is not ready yet. But new methods are emerging: carbon capture systems are coming online, often capturing up to 95 percent of emissions and transforming traditional power stations into credible transitional assets. In the future, advanced nuclear small modular reactors (SMR), represent a new generation of compact nuclear plants built in factories, shipped on trucks, and installed where power is needed. These offer 24/7, zero-carbon baseload electricity and unparalleled
resilience. Major projects are already under way—like TerraPower, backed by the U.S. Department of Energy—with the first deployments on the horizon. A balanced portfolio of renewables, nuclear and cleaner fossil generation will drive AI’s growth while reducing carbon impact. MIGRATION TO HIGHER SPEEDS The network landscape is evolving quickly to keep pace with AI. Not long ago major deployments standardized on 40G, then 100G links. 400G is already mainstream within AI clusters, and 800G is rapidly becoming the new benchmark for backend connectivity in hyperscale environments. But it does not appear to stop there. IEEE 802.3dj is working on standards development for 1.6 terabit (Tb) ethernet to support next-gen AI workloads, with early deployments expected by 2027. These leaps are essential to support massive parallelism, chiplet-based architectures, and faster training cycles.
Coal Natural gas Nuclear Solar Wind Other renewables Other
Network
Client
Spine switches
North-south tra c
Leaf switches
Servers
East-west tra c
FIGURE 3 : A stacked area chart of the projected global electricity demand by source from 2020 to 2035. Source: International Energy Agency
FIGURE 2 : A diagram that illustrates data traffic flows within a common data center network architecture. Source: AFL
I
I
32
ICT TODAY
January/February/March 2026
33
Made with FlippingBook - Online catalogs