Stratos, a proposed 9 GW AI and data center campus in Utah, backed by investors including Kevin O’Leary, has emerged as one of the most ambitious infrastructure projects of the AI boom.. Developed under Utah’s Military Installation Development Authority (MIDA), the controversial project combines hyperscale compute, large-scale energy generation, and accelerated development structures inside a single industrial zone built for AI.. But beneath the headline capacity is a broader pivot in how AI campuses get built and powered: developers are increasingly constructing around the grid, not merely connecting to it.. Building Around the Grid. For years, hyperscalers largely followed a grid-first formula: utilities built generation and transmission, while data centers tapped the grid as large industrial customers. The relationship occasionally strained local systems, but the overall structure remained intact.. Related:AI Inference Pulls Infrastructure Back into Metro Data Centers. AI has now changed the math. Training clusters and inference fleets now consume power at scales that collide with transmission bottlenecks, transformer shortages, interconnection delays, and utility planning cycles built for slower industrial growth. In major markets, developers now face waits measured not in months, but in years.. Those constraints now shape where and how large AI campuses get built. Instead of prioritizing fiber-rich metro corridors alone, developers increasingly hunt for places with fast permitting, available land, pipeline access, political alignment, and pathways to self-controlled generation.. Power availability now matters as much as geographic proximity, and Stratos is structured around that logic. The project is organized through MIDA, which enables streamlined processes and coordinated infrastructure development within a defined project area.. “The Project Area is intended to support the development of state-of-the-art energy generation; hyperscale data center(s); advanced manufacturing,” the MIDA project area filing states.. The filing frames the project around energy resilience, integrated infrastructure, and “secure, domestically controlled power and data capacity” aligned with national-security objectives.. At a proposed 9 GW, the campus would rival the power footprint of a large metropolitan area – more akin to industrial energy development than conventional data center expansion.. Grid operators themselves increasingly warn that AI-scale load growth is beginning to strain existing planning systems. In an April update to the Texas Senate Committee on Business & Commerce, the Electric Reliability Council of Texas (ERCOT) said it was tracking roughly 410 GW of large-load interconnection requests, of which approximately 87% were tied to data centers. The grid operator also described a “rush of requests for AI Data Center loads seeking firm service” as hyperscale projects flood interconnection queues. ERCOT further warned that existing interconnection structures increasingly force projects into repeated “restudy loops,” potentially delaying approvals by years.. Related:Texas May Have Accidentally Built the Perfect Grid for AI. Industrial Energy Economics. The Stratos proposal reflects those pressures almost perfectly.. “This always looked like industrial energy,” said Jigar Shah, former director of the US Department of Energy’s Loan Program Office (DOE LPO), in an interview with Data Center Knowledge. “This means that you need an offtake before you know what your economics are.”. That framing departs from the legacy model, in which utilities extended infrastructure to support enterprise and cloud growth on more predictable timelines.. Shah characterized xAI’s rapid Colossus deployment in Memphis as an unusual case made possible by aggressive capital deployment and unique market timing. “The only way for someone to copy Elon is to use pure corporate equity over infrastructure investment,” Shah said. The trade-off is capital risk: at the multi-gigawatt scale, project economics hinge on long-term offtake, sustained AI demand, and the ability to fully utilize massive power and infrastructure commitments.. Related:AI Data Center Boom Rewires US Power Supply Chain. Mechanics on the Ground: More Refinery than Server Farm. Trey Travis, VP of Operations at Southeastern Hose, said AI infrastructure deployments increasingly resemble large industrial mechanical projects rather than traditional data center builds. “People hear ‘AI data center’ and think of software, but what’s going on in the ground looks more like a refinery,” Travis told Data Center Knowledge.. Travis noted that high-density GPU deployments are reshaping cooling and fluid-handling requirements across new AI campuses, driving demand for large-scale cooling distribution unit (CDU) loops, high-pressure hose assemblies, expansion joints, and industrial-grade fluid systems rarely associated with cloud builds only a few years ago.. “Honestly, this looks much more like a large-scale industrial plant order than a data center order,” Travis said.. He added that many project teams are confronting mechanical and supply-chain challenges common to heavy industry, including vibration management, thermal expansion, specialty fittings, certification bottlenecks, and long lead times for polytetrafluoroethylene (PTFE)-lined assemblies and other specialized components.. Centralized vs. Distributed: The Infrastructure Debate. The Stratos model sits at one extreme of an emerging debate over AI infrastructure design. Some developers and investors continue to pursue massive centralized campuses tied to dedicated power systems. Others argue that future inference demand may spread into smaller, more distributed facilities closer to users and enterprise workloads.. HyperFrame Research CEO and principal analyst Steven Dickens said the AI infrastructure conversation has already shifted away from traditional data center metrics toward raw power availability.. “We don’t talk about racks, or even servers anymore – it’s all about power consumption,” Dickens told Data Center Knowledge.. Dickens said future AI infrastructure will likely include a mix of large centralized campuses for frontier training and regionally distributed inference facilities, particularly in Europe, where data sovereignty and latency requirements are decisive.. “Power availability will play a role in the DC placements,” Dickens said.. Shah questioned whether AI’s long-term trajectory will remain centered on multi-gigawatt campuses at all. “Very smart people I trust are now saying that we need no new central data centers beyond the ones already under construction,” he said. “The rest can be less than 100 MW data centers, down to 100 kW data centers. These can be deployed quickly.”. If Stratos represents one end of the spectrum – vertically integrated AI infrastructure tightly coupled with dedicated power systems – the distributed model represents almost the opposite approach: smaller facilities deployed closer to demand and existing infrastructure.. For now, however, the economics of frontier-scale training still push developers toward dense concentrations of power and compute.. Developers Become Energy Operators. As regulators tighten interconnection requirements and utilities struggle to accommodate large-load demand, developers are increasingly structuring projects to control more of the energy stack. At the same time, requirements around site control, financial commitments, and grid interconnection are getting stricter.. Stratos appears structured around avoiding some of those constraints. The project filing repeatedly emphasizes integrated generation, long-term power control, and accelerated development structures rather than relying solely on conventional utility expansion.. Stratos may ultimately prove less important as a singular megaproject than as an early example of where portions of the AI infrastructure industry are heading: toward campuses where compute infrastructure and energy infrastructure are beginning to converge into a single industrial system.
