The Breaking Points: Power Emerges as AI’s Defining Limit

extract

The Breaking Points: Power Emerges as AI’s Defining Limit As AI workloads scale, power limitations are increasingly driven by infrastructure timelines and system complexity, rather than generation alone. Our second in a three-part series examines how power is becoming the defining limit for AI infrastructure, reshaping deployment timelines and site selection. A developer signs off on a new AI deployment. The models are ready, the GPUs are secured, and the site is selected. Then the timeline slips – not because of compute, but because power isn’t available, and won’t be for years. As AI infrastructure scales, access to electricity has become one of the most immediate constraints on deployment, making delays like the scenario above increasingly inevitable. Chips still matter, but power now determines when and where new capacity can be built. From Power Availability to Power Delivery Grid interconnection backlogs have become a national bottleneck for AI-era power delivery. In 2023, Lawrence Berkeley National Laboratory found nearly 2,600 GW of generation and storage capacity seeking transmission interconnection – more than twice the size of the existing US power plant fleet at the time. That surge in demand is driving a new wave of infrastructure investment. US utilities and power providers are expected to invest roughly $1.4 trillion through 2030 to expand generation, transmission, and grid capacity, according to estimates from Deloitte and industry analysts. Even so, the challenge is no longer just how much power exists, but how quickly it can be delivered. Global demand for AI infrastructure is accelerating faster than utilities and grid operators can match. In many regions, interconnection timelines now stretch into years, even where generation is ample. Curt Holcomb, vice chairman with JLL’s global Data Center Solutions practice, said the industry was already under pressure before the current AI wave. “We were already running out of power coming out of the last expansion cycle,” he said. “Then AI requirements came in at two to three times the power levels.” The surge has left little excess capacity in key markets. “That low-hanging fruit is pretty much gone,” Holcomb said. The constraint is visible on the ground. “Everyone I talk to says if I wanted to build a new facility today, I can’t get power,” said Kevin Wollenweber, senior vice president and general manager of Data Center and Internet Infrastructure at Cisco. Holcomb said timelines for new capacity are stretching significantly as infrastructure requirements increase. “In many cases, you’re looking at 36 to 48 months from commitment to delivery," he said. "The first bottleneck is figuring out how to get the power to the site." Interconnection backlogs have turned from a planning hurdle into a gating item for AI deployment. (Image: Alamy) A Geographic Reset in Site Selection As backlogs intensify in established markets, development is shifting toward regions with better access to energy and faster approvals. Recent data from Synergy Research Group indicates that Texas and Midwestern states are poised to capture more than half of the new US hyperscale capacity, as operators prioritize power availability over historically dominant factors like network proximity. John Dinsdale, chief analyst at Synergy Research Group, said power availability has become the dominant factor in site selection. “With the extremely rapid growth in demand for AI technology and infrastructure, availability of power has now become the most dominant decision criteria,” he told Data Center Knowledge. Additional analysis from ABI Research reaches a similar conclusion. In The Great Data Center Divide (April 2026), the firm argues that grid access and permitting now outweigh chip supply in determining where AI capacity can be deployed, accelerating a shift toward regions with available energy and faster approvals. Texas remains one of the clear exceptions, but not a clean escape. Berkeley Lab has noted that some regions, including ERCOT in Texas, process interconnection requests more quickly than many other markets, helping to explain Texas's growing share of new hyperscale development. Beyond (and Behind) the Grid With interconnection timelines extending, some operators are exploring alternatives to traditional grid dependence. On-site generation, hybrid energy systems, and large-scale battery energy storage systems (BESS) are entering the conversation as ways to accelerate deployment and de-risk delivery. Holcomb noted that “bring-your-own-power” strategies are gaining traction, though they are not viable in every market due to regulatory and environmental constraints. Wollenweber said this is introducing a new layer of complexity. “You’ve got solar, wind, traditional generation, and batteries all being coordinated together,” he said. “It’s a complex system.” Rather than a single power source, operators are increasingly managing multiple inputs to balance reliability, cost, and sustainability. However, even as generation capacity expands, questions remain around how much energy storage is required to support next-generation AI workloads. Battery systems are becoming critical as operators incorporate intermittent renewables and attempt to smooth demand spikes, but the industry is still working through how to size and integrate them. Developers may know how much power they need, but have less clarity on how much storage is required to maintain consistent performance under dynamic AI loads. Luke Witmer, vice president of software engineering at Wärtsilä Energy Storage, said early-stage storage designs often begin with two-hour battery systems sized to expected load volatility, then refine based on workload profiles and UPS design. Hybrid portfolios blend intermittent renewables with storage and grid power to balance reliability and cost. (Image: Alamy) System-Level Complexity: Charging Ahead The growing complexity of power infrastructure is reshaping how data centers are designed and operated. What was once a relatively linear system – generation, transmission, distribution – is becoming more dynamic, with multiple inputs and real-time coordination requirements. Wollenweber said this shift is forcing operators to rethink infrastructure. The challenge is no longer just securing power, but also managing it as part of a broader system that includes storage, generation, and fluctuating demand from AI workloads. Witmer said this shift fundamentally changes how power systems are operated. “The battery isn’t just a backup generator on standby anymore,” he said. “You need active state-of-charge management, load forecasting, automated dispatch, and reserves management.” Power constraints are no longer theoretical. They are delaying projects, reshaping site selection, and redefining how data centers are designed and operated. As AI workloads grow, the ability to deliver and manage power is becoming the critical factor in bringing new capacity online. The challenge is no longer generation; it is delivery at scale. In our next installment, we will delve into how cooling systems are evolving to meet the demands of AI at scale.