A 1988 Data Center Becomes a 20 MW AI Asset
The second presentation, Converting a 1990 Enterprise Data Center to 20MW AI Workloads, featured Coe alongside Ari Chernikoff, vice president of development at Legacy Investing, and Dan Amundson, project manager at Gardner Builders.
The project centered on 1001 Third Avenue South in downtown Minneapolis, a financial-services computing facility originally constructed in 1988. The building had supported earlier generations of enterprise computing before portions were converted to commercial office space and smaller colocation operations. Although its remaining operating data center load had declined to approximately 500 kilowatts by the time the AI redevelopment began, the property retained substantial infrastructure inherited from its original mission-critical design.
Tall ceilings, robust floor loading, raised access floors, existing medium-voltage substations, and established utility connections made the building an unusually strong candidate for conversion. These characteristics, Chernikoff explained, can determine whether an existing building offers a practical path to high-density computing or proves economically unsuitable for redevelopment.
Legacy Investing had identified a market opportunity in what Chernikoff called the industry’s “missing middle”: facilities generally ranging from 20 MW to 80 MW, rather than the 100 MW-plus campuses increasingly associated with hyperscale and frontier AI development. The company viewed existing buildings with available power and favorable infrastructure characteristics as potential shortcuts to deployment, particularly where demand for AI inference and other latency-sensitive workloads favored proximity to major urban markets.
Following an AI tenant commitment in summer 2024, the development team faced an aggressive initial requirement: Deliver the first 5 MW of capacity by the end of that year. The first phase was ready for service approximately four months after Gardner’s involvement became formalized, followed by another 5 MW in 2025 and an expansion program ultimately reaching 20 MW over roughly 20 months.
All of this had to be accomplished while keeping existing data center customers and commercial tenants operational.
For Coe, the complexity was well matched to his professional experience. He’s a data center “OG” in the best sense of the term—a respected industry veteran with approximately 31 years of design experience, reaching back to an era before highly standardized facility templates became commonplace, when individual projects often demanded specialized approaches to equipment integration and infrastructure design. Coe had already demonstrated that perspective during DCF’s coverage of Vertiv’s Peter Panfil at 7×24 Exchange Spring 2026.
“Because data centers used to be one of a kind, I’m one of the few specialists left that can take unique buildings and find parts and pieces and the best optimum way to upgrade them,” Coe told the audience in San Antonio last Tuesday.
That institutional knowledge became an essential resource as the team confronted a building containing multiple generations of electrical and mechanical infrastructure.
The original facility included aging UPS equipment, an assortment of generators, eight medium-voltage substations, and chillers approaching or exceeding their intended service lives. Rather than replacing everything in a single construction campaign, the project required successive operating configurations that could deliver new AI capacity while maintaining existing loads.
Chernikoff described the team’s willingness to depart from conventional design and procurement sequences as another important element of its strategy.
“We always try to let procurement wag the design dog,” he said.
Instead of establishing a fixed equipment specification and waiting for manufacturers to deliver every component, Legacy aggressively sourced available equipment and worked with Syska Hennessy and Gardner Builders to adapt the designs accordingly. The project incorporated a mixture of existing generators, used Caterpillar equipment, new MTU units, and alternative electrical distribution components selected partly because they could meet the required schedule.
Coe described using Square D panelboards in place of conventional busway because busway lead times were incompatible with the deployment timetable. Existing generator arrangements were reconfigured, older systems were integrated with newer UPS equipment, and electrical distribution evolved through successive phases as additional capacity became available.
The result was not an exercise in abandoning engineering discipline. Rather, it demonstrated how experienced engineers could accommodate nonstandard equipment combinations while maintaining the necessary system-level reliability.
Amundson acknowledged that the speed of the first phase sometimes created problems requiring additional work during subsequent stages. Integrating new chillers with existing equipment involving different water temperatures, flow rates, and operating characteristics proved especially challenging, as did modifying the building’s older electrical systems while portions remained energized and in service.
The process demanded constant collaboration among the owner, engineers, construction managers, subcontractors, equipment suppliers, and building operations personnel.
It also provided a concrete illustration of Mahmood’s argument that the ability to finance or announce an ambitious infrastructure project is not equivalent to the ability to deliver and operate it.
Engineering AI Into an Occupied Downtown Building
The physical constraints of the Minneapolis project were particularly formidable. The building occupies an entire downtown block, with essentially no adjacent land available for new generators, cooling equipment, or other major infrastructure. Expansion therefore required repurposing existing floors and installing additional equipment within the building or on its roof.
Five generators were eventually installed on a rooftop structure using an enormous mobile hydraulic crane transported from Iowa. According to Amundson, the crane required 21 truckloads and two days of assembly, with the lift plan accounting for different generator weights, a roughly 185-foot reach from the sidewalk to the installation area, and underground vaults and basement structures near the crane setup locations.
The mechanical work was equally demanding. To install three new 1,500-ton chillers, the team temporarily removed sections of the building’s fifth-floor curtain wall and used detailed three-dimensional modeling to establish equipment paths, piping routes, and construction sequences. Existing cooling systems had to continue serving operational data halls throughout the installation, requiring a carefully phased approach in which new equipment was brought online as additional electrical capacity became available.
The computing equipment introduced further complications. The deployment incorporated Cerebras AI systems with substantial liquid-cooling requirements, including cabinets drawing approximately 55 kW each, with most of the heat load removed through liquid cooling. Associated networking infrastructure relied on a different thermal approach.
Chernikoff explained that Cerebras had encountered humidity and condensation challenges at other deployments where liquid-cooled AI equipment and air-cooled networking systems operated in close proximity. The Minneapolis building’s layout allowed the engineering team to separate those environments, placing AI computing equipment on one floor and associated networking infrastructure on another.
The strategy reflects a broader push to extract AI capacity from existing facilities. As DCF reported in its coverage of Corvex’s liquid-cooled AI retrofit strategy, operators are exploring alternative thermal architectures to deploy high-density accelerators within buildings originally designed around conventional air cooling, sometimes avoiding extensive facility-wide mechanical reconstruction.
The Minneapolis installation incorporated coolant distribution units, buffer tanks, rear-door cooling equipment, and extensive flow-control instrumentation. Coe described the use of hundreds of Belimo valves to provide detailed monitoring and control of water temperatures and flow rates serving the computing systems.
Even the choice of piping materials revealed the complexity of adapting conventional construction practices to specialized AI requirements. The team selected carbon-steel piping with Victaulic fittings to accelerate installation, then encountered stringent requirements for iron content in the cooling-water system. Working with chemical treatment specialists, the contractors developed flushing and passivation procedures intended to satisfy Cerebras’ water-quality specifications.
These challenges reinforced Mahmood’s earlier observation that the future of cooling will not be determined by a single favored technology. The effective solution depends on the workload, available infrastructure, maintainability, and the ability to deliver dependable thermal performance under actual operating conditions.
But perhaps the most impressive accomplishment was maintaining uninterrupted service throughout the conversion.
Amundson explained that the project’s later phases required construction immediately adjacent to operating data halls, including modifications to power distribution and chilled-water infrastructure supporting active equipment. During the second major deployment phase, the team issued more than 100 maintenance advisories involving work that affected operating systems or their redundant feeds.
“We didn’t cause any outages, but we did touch the redundancy over 100 times,” Amundson said.
Coe described the corresponding electrical challenges, including changes to UPS configurations, power distribution panels, and individual cabinet circuits as the building evolved toward the required eight-path power architecture. The work relied on the redundancy built into the installed computing systems to keep workloads running while portions of the supporting infrastructure were modified.
With approximately 28.5 MW of utility capacity available for the 20 MW IT deployment, operating efficiency was also essential. Coe reported that the facility generally achieved a PUE of approximately 1.30 to 1.31, while Chernikoff identified a further improvement target of 1.26 to 1.27.
The achievement illustrated the physical counterpart to Mahmood’s warning about autonomous infrastructure management. Whether decisions are made by experienced engineers or intelligent control systems, critical infrastructure must be understood as a collection of interacting systems in which changes to one component can have consequences throughout the operating environment.
For the Minneapolis team, the responsibility was to execute an unusually aggressive construction program without allowing those interactions to interrupt service.
The Grid, the Community, and the Social License to Operate
Power availability was central to the Minneapolis redevelopment strategy. Chernikoff explained that Legacy worked with Xcel Energy to secure incremental capacity from existing downtown electrical distribution infrastructure, rather than waiting for entirely new generation or transmission facilities to be developed.
According to the project team, the incremental nature of those capacity additions allowed the development to avoid regional transmission organization studies associated with certain larger power requests. The approach was particularly well suited to Legacy’s focus on facilities within the 20 MW-to-80 MW range, where available distribution capacity and strategic site selection can offer advantages over larger greenfield developments.
The building’s downtown location also created significant regulatory and construction challenges. The installation of lithium-ion UPS batteries above the 75-foot building-height threshold, for example, required extensive coordination with Minneapolis officials over code compliance and fire protection. The team ultimately incorporated multiple safeguards, including temperature and gas detection and specialized battery protection equipment, while working through an iterative approval process with the city’s fire marshal.
These experiences connected to a broader argument Mahmood had advanced about the industry’s relationship with the communities hosting its infrastructure.
One of the Fifth Compute’s greatest challenges, he suggested, may be social rather than purely technical. The industry has historically designed data centers to be largely invisible, allowing the public to use the services they enable without much awareness of the buildings, power systems, and resources behind them.
That invisibility becomes a liability when data center development begins affecting local electricity systems, water supplies, construction activity, land use, or neighborhood expectations.
Mahmood argued that developers should treat communities as partners in infrastructure planning rather than stakeholders to be engaged only after major project decisions have already been made. He emphasized that a government permit and sustained public acceptance are not the same thing, and that community trust requires ongoing transparency about a facility’s benefits and impacts.
“A community may grant us permission to build,” Mahmood said, “but we have to earn the privilege of remaining.”
One of his more provocative proposals was what he called the “nutritional internet”: a standardized disclosure framework that would make the characteristics and impacts of data center developments more understandable to the public.
Borrowing from the familiar nutritional labels on packaged foods, Mahmood envisioned disclosures describing a facility’s energy demand and sources, water consumption, potential grid impacts, opportunities for waste heat recovery, and contributions to employment, education, and local economic development. Such a framework, he argued, could give communities and policymakers more consistent information with which to evaluate proposed projects and monitor commitments over time.
The concept was not presented as a finished regulatory standard, but as a challenge to an industry that has become highly sophisticated at measuring infrastructure performance while remaining less effective at communicating its broader physical footprint.
For Mahmood, the Fifth Compute will require the industry to become better integrated with the systems surrounding its facilities, rather than treating power, water, community acceptance, and public policy as external concerns.
The Minneapolis redevelopment offered a narrower but tangible example of the value of existing infrastructure and local coordination. Its conversion ultimately also produced a commercial outcome: In January 2026, Cloud Capital and Arcapita announced the acquisition of the facility, describing a 21 MW property with plans for a further 10 MW expansion.
For Legacy Investing, the transaction demonstrated the potential investment value of repositioning older computing facilities for new AI workloads. For the broader industry, it underscored the commercial importance of sites capable of delivering power and computing capacity without requiring an entirely new development cycle.
From Orbital Data Centers to the Next Generation of Engineers
Mahmood’s keynote ultimately ranged well beyond terrestrial infrastructure.
He explored the possibility that computing could eventually become a critical component of humanity’s expansion into space, with orbital platforms, lunar systems, and interplanetary networks supporting autonomous operations, scientific research, and communication across increasingly distant environments.
The rationale was not simply to move conventional data centers into orbit. Mahmood identified several potential drivers, including extended access to solar energy in certain orbital configurations, processing data closer to space-based sensors and spacecraft, and providing intelligence capable of supporting activities where communication delays make constant human intervention impractical.
The engineering challenges, however, remain formidable. Heat rejection in space differs fundamentally from terrestrial cooling, radiation introduces new reliability requirements, launch costs constrain equipment design, and conventional maintenance assumptions collapse when technicians and replacement components cannot readily reach the installation.
Mahmood acknowledged those limitations and resisted treating orbital data centers as an inevitable substitute for facilities on Earth. Instead, he argued that emerging space activities could eventually create specialized computing requirements that justify developing new infrastructure beyond the planet.
Those possibilities also illustrated a larger point about the evolution of computing: Each new environment requires an underlying ecosystem of power, communications, operations, and reliability before increasingly sophisticated applications can be supported.
In that respect, the engineering challenges of a lunar computing installation and the conversion of an aging Minneapolis data center share a fundamental concern. Both require the integration of computing capabilities with physical infrastructure adapted to the demands of a particular operating environment.
Yet Mahmood also identified a resource that cannot be manufactured or deployed at the speed of computing hardware: experienced people.
Citing the Uptime Institute Global Data Center Survey 2025, Mahmood noted that 46% of operators reported difficulty finding qualified candidates for open positions, while 37% reported challenges retaining personnel. The figures reinforced his argument that the expansion of digital infrastructure depends not only on securing equipment and electrical capacity, but also on developing the people capable of building, managing, and maintaining increasingly sophisticated facilities.
During the keynote Q&A, an attendee raised a related concern: If AI begins displacing junior engineers and other early-career technical professionals, what happens to the traditional learning process through which experienced personnel transfer knowledge to the next generation?
Mahmood’s response emphasized the responsibility of industry veterans to collect, preserve, and share the practical knowledge developed throughout their careers. He also encouraged the industry to seek talent beyond conventional educational and professional backgrounds, including veterans, skilled tradespeople, community college students, and career changers.
That argument created perhaps the strongest connection between Mahmood’s vision and the Minneapolis case study.
Coe’s decades of engineering experience enabled him to recognize the potential in equipment and building systems inherited from earlier generations of data center design. At the same time, Gardner Builders entered the project without extensive prior data center experience, requiring its team to work closely with established engineering and construction specialists.
Coe described that unfamiliarity as a potential advantage, saying he had enjoyed working with first-time data center general contractors because they approached individual challenges without assuming that conventional solutions were necessarily the only solutions.
“They don’t make any assumptions, they study every issue,” he said.
The Minneapolis project was therefore not simply a demonstration of legacy infrastructure being adapted for AI. It was also an example of experienced engineering judgment being combined with fresh perspectives, intense collaboration, and a willingness to depart from conventional approaches when circumstances required it.
That combination may become increasingly valuable as the industry confronts the overlapping demands of accelerated construction, new computing architectures, evolving thermal requirements, and the need to preserve reliability across increasingly complex environments.
The Fifth Compute Still Depends on First Principles
The contrast between Mahmood’s keynote and Coe’s engineering case study ultimately revealed a shared understanding of the industry’s future.
The Fifth Compute may introduce systems capable of acting with increasing autonomy, processing information through novel architectures, and operating in environments far beyond today’s terrestrial computing infrastructure. Those developments could fundamentally alter the relationship between technology and society.
But the resulting infrastructure must still be engineered, constructed, commissioned, maintained, and operated within real-world constraints.
Not every older building can become a modern AI data center. The Minneapolis project benefited from unusually strong structural characteristics, existing electrical infrastructure, available utility capacity, and a project team prepared to manage an exceptionally complicated conversion. Its success does not establish a universally repeatable shortcut around the industry’s power and construction bottlenecks.
What it does demonstrate is the continuing relevance of engineering adaptability, experienced judgment, and operational discipline in an era increasingly focused on standardized AI factories, gigawatt-scale campuses, and ambitious deployment schedules.
Mahmood’s central message was similarly grounded. The promise of increasingly intelligent systems does not relieve the industry of responsibility for their consequences. Greater autonomy requires stronger accountability. More powerful computing demands more sophisticated infrastructure planning. And continued technological advancement places a premium on the human knowledge and institutional experience required to make new capabilities dependable.
For an industry whose facilities increasingly underpin essential economic and social functions, those responsibilities extend beyond the technical boundaries of the data center itself.
Toward the end of his keynote, Mahmood offered a warning that captured this broader obligation: “Just because we can build something doesn’t automatically mean that we should.”
The question, he suggested, is not simply how much computing capacity the industry can create, but what humanity ultimately accomplishes with it—and whether that progress justifies the resources and responsibilities involved.
The Minneapolis conversion provided a practical answer to one part of that challenge. Coe and his colleagues showed how the accumulated expertise of earlier computing generations could be applied to a radically different set of requirements, creating new capacity without abandoning the reliability principles that made the original facility valuable.
In the end, the intersection of traditional and AI data centers may be less about replacing the industry’s past than about determining which of its principles, practices, and accumulated knowledge will remain essential as the boundaries of computing continue to expand.
The Fifth Compute may eventually take data centers beyond Earth. For now, its future remains closely tied to engineers who know how to keep critical systems running while rebuilding the infrastructure around them.
