Decentralizing computing infrastructure, re-timing operational hours, and dedicated vs. public energy generation and grid connections are three decisions under consideration when dealing with overburdened grid draws and “below demand” energy distribution levels met, resulting from data center operations and artificial intelligence (AI). According to John Haw, CEO of Fidelity Energy, “ …AI workloads are highly power-dense and often require a continuous, stable supply. That means decentralisation doesn’t remove energy demand — it redistributes it. In practice, the success of these models depends on how intelligently workloads, storage, and local energy systems are coordinated.” This means that placing an end-to-end over different points will spread the load but not decrease the load demand. The “hub and spokes” models work for the required data center operations network efficiencies, but these models do not encourage a decentralized data center network approach. It is not realistic that AI decentralization saves enough energy combatting the load-drawing issues of running data centers. Eshan Jayamanne, Founder and CEO at Krane, stated, “Transmission losses over long distances are real, roughly 5% to 7% across the U.S. grid, but the bigger factor is grid congestion. Locating a 200-megawatt data center in a region with an already constrained grid creates bottlenecks that affect every other user on that circuit, regardless of proximity to the generation source. That's why we're seeing new data center campuses increasingly co-locate near dedicated power generation…”
Re-timing data center operations to coincide with "lower energy demand" hours in a region is somewhat feasible. David Keller, Co-Founder and CEO at InFlux Technologies, explains, “The ‘fail’ and ‘recovery’ of data centers during peak energy draw is determined by redundancy. A centralized compute network where a single authority handles resource provisioning could fail due to excessive energy draw, whereas a distributed compute network will have enough redundancy to maintain failover even during peak draws. Network operations are already optimized and timed for peak energy draw hours. Having a data center closer to centralized energy sources means faster execution and reduced latency.” The source dedication is more important than timing. This is an agreed-upon position within the data center and energy industry communities. Alex MacColl, Project Manager (EMEA) at DataMove, goes on to say, “Some workloads can be paused or scheduled to suit capacity requirements, and it is a prudent way to manage costs. For latency-sensitive work, however – for instance, financial transactions, betting platforms or real-time inference – these kinds of operations cannot be re-timed to suit, which explains some operators carefully planning their site locations to ensure the best operational conditions for their needs.”
Dedicated energy facilities set aside to meet only the needs of a major data center without sharing grid connections to other sites or customers are a hyperscale solution. Already, grids are being built solely for AI. Questions are rising about whether these builds can keep up with demand and if regulations will keep pace. The construction at these energy campuses connects from a generator directly to a data center and has no public grid interaction. Distributed cloud computing is also seen as a way to connect point-to-point hardware utilized for energy conservation. Eshan Jayamanne states, “If data center access becomes pay-to-play at the individual or community level, you'll see the same patterns we see in broadband access: wealthier communities with better infrastructure attract more investment, which further widens the gap.” Whitaker B. Irvin, Jr., CEO at Q Hydrogen, goes on to say, “How we answer the cost allocation question will ultimately determine whether distributed AI infrastructure becomes a broadly accessible resource or an amenity that concentrates further advantage among those who can already afford to build and operate their own energy supply. Large operators paying for dedicated generation and grid infrastructure while smaller players remain dependent on the public grid produces both an energy and technology gap, where the reliability and cost of power itself is a competitive differentiator.”
Data center and AI operational costs are enormous. The amount of power drawn from the grid to support the infrastructure and the connected point-to-point networks leads to grid overdraws, blackout hours, and depletion of energy storage reserves. As data centers and AI develop, the energy industry needs to keep up. While large, spiraling energy generation projects with out-of-reach estimates and independent backers have been put on the table, a more conservative approach to energy infrastructure development walking hand-in-hand with data center and AI needs is winning the day. Decentralizing computer infrastructure is realistic, but it only solves part of the energy distribution problem. Re-timing operational hours to accommodate data center or AI facility on-peak decisions for functions is part of policy in some already existing data center functions. Last, but not least, dedicated energy generation and grid power, not public power inclusion for data center and AI operations, is feasible and may re-scale a market, but demand might outpace scale.
Looking at these ideas opens a Pandora’s Box that needs to be explored. Without some answer for a greater data center or AI power draw from the grid than from any other customer, consistent with use and storage frequency, there is a future coming that holds necessities turning into luxuries (such as air conditioning, electric lights, etc.) simply because AI requires the “extra” kWhs. Data centers and AI are already using more energy than expected, and power solutions for grid overdraws or insufficient energy distribution are required. How data and energy infrastructure co-exist, who absorbs the overrun costs, and where the "conversation" can make a difference, are important questions to ask as the future of data centers and AI arrives.