How do you lift six buildings with 36 different floor levels? Not by defying the odds, but through carefully calculated engineering.   Lazari's new film captures the floor jacking at 63 New Bond Stree...

When Wilkes County lost 60% of its childcare facilities, Wilkesboro United Methodist Church stepped in to turn dusty classrooms into a safe haven for 80 children.

Mayor McCarty said the city will reorganize its economic development team, create a concierge-style permitting service and adopt ordinances to streamline development and adaptive reuse—aiming to make Sacramento more competitive for housing and business projects.

The commission approved the design for a major interior and façade rehabilitation at 701 State Street (Paseo Nova), endorsing reuse of existing steel arched windows, a lowered parapet and a revised tower, and asking for refinements to awning depth, cornice proportions and select cornice/window details.

The Mission City Planning Commission voted 4–3 to recommend City Council approval of the Mission Fox Ridge Preliminary Development Plan (case 26-15) after reconsidering traffic-safety evidence, an EIFS cladding deviation with mineral wool, and mixed-use feasibility; the commission added a stipulation on tower elevations and forwarded the file to Council.

How Milford Hills United Methodist Church discovered that their empty space was the answer to their city’s housing crisis.

In the years after World War II, as highways extended outward and housing followed, the shopping center became a defining building type of the American suburb, and the fully enclosed regional mall …

Willington selectmen reviewed a draft charge and application to form a School Building Committee to study reuse and disposition of Hall and Center schools; questions focused on timing with the SDC, access to floor plans for adaptive reuse, and funding for consultants.

In the years after World War II, as highways extended outward and housing followed, the shopping center became a defining building type of the American suburb, and the fully enclosed regional mall became the template for a generation of construction. Victor Gruen designed Southdale Center in Edina, Minnesota, for the Dayton Company, and it opened on October 8, 1956, as the first fully enclosed shopping mall. Malls of this kind were designed around a single purpose, and although many had glazed entrances, skylit courts, and elaborate common spaces, the type was dominated by deep floor plates, largely windowless department-store anchors at the ends, and an ocean of parking around the perimeter. No one designing them imagined what they would become when retailing changed, and retailing did change. The peak came around 1990, when 19 new malls opened in a single year; the decline began in the late 1990s, the last new enclosed mall was built in 2006, and 2007 was the first year since the 1950s in which none opened. The Great Recession and the growth of e-commerce accelerated the trend, and Green Street, which tracks the industry, estimates that about 150 enclosed malls closed after supply peaked in 2008, leaving roughly 950 — with the pandemic adding further pressure. What has followed is a long, improvised experiment in adaptive reuse, and some of the outcomes would have been unimaginable to the original developers. The most striking is that dead malls have been turning into data centers. In Indianapolis, Lifeline Data Centers purchased the 41-acre Eastgate retail mall in 2008 and opened its first data center building there in 2009; Eastgate, built in 1958, had been the city's first mall. Other examples have accumulated since: a former Boscov's department store at Marley Station Mall in Glen Burnie, Maryland, is now an AiNET data center, and the former Bon Marché Mall in Baton Rouge has become the Bon Carré Technology Center. The same CoStar article notes that the conversions are not trivial — mall floors may need structural reinforcement to carry the weight of data-center equipment — although the large spans, the existing utility connections, and the absence of windows that once made a mall a poor candidate for most other uses turn out to suit servers reasonably well. The pattern is still running, although the most recent example is provisional and changing week to week. In June, KG Growth Partners, a Maryland firm working for an unnamed prospective buyer, inquired with the city about converting 845 Market Street — the roughly 350,000-square-foot former Bloomingdale's portion of San Francisco's shuttered Centre mall, separately owned by Macy's — to data center use. No permit application has been filed, and no sale has been completed; separate reporting has identified another investor, B3 Investors, as negotiating for the Bloomingdale's parcel, and it is not clear whether KG's client remains in the running. What is interesting for my purposes is how the proposal is framed: KG's chief executive insists it is not a data center in the traditional sense but a "regenerative urban energy center," occupying about a third of the complex, with the remainder given to retail, culture, sports, and free education and six acres of green space on the roof — a hybrid reuse rather than a server farm. It has arrived at an awkward moment. The Board of Supervisors unanimously approved a 45-day citywide moratorium on data center approvals on October 6, which Mayor Lurie signed on October 16, and as of this writing the lenders who took the mall back through foreclosure have reportedly selected a different buyer, Seven Equity Group, for the majority of the property, excluding the Bloomingdale's parcel. Whatever happens to this particular proposal, it shows how naturally the dead mall and the data center now appear in the same sentence. ## The office building follows the same path The office building is on a parallel trajectory, and here the scale is larger. Remote and hybrid work have left many downtowns with persistent vacancy, and conversion of office space to housing has moved from a curiosity to a product type. According to RentCafe, 90,300 apartments were in the office-to-residential conversion pipeline at the start of 2026, up 28% from the year before and almost four times the number in 2022; Philadelphia, with 2,697 units in the pipeline, is among the leading cities. There is a detail in those numbers that matters for the argument I want to make. In Chicago, future office-to-residential conversions are expected to come from buildings that average 90 years old. Older office buildings, with narrower floor plates, operable windows, and floor-to-window distances that let daylight reach the interior, convert to apartments more readily than the deep, sealed floor plates of the late twentieth century. In other words, the buildings that are easiest to reuse today are largely the ones whose designers — through constraint rather than foresight — happened to make choices compatible with uses they never contemplated. Convertibility was an accident, and where it was absent, the cost of conversion has been high or prohibitive. Data centers, as it happens, have a history of being built inside other buildings. Sean Farney of JLL has observed that the first data centers were themselves adaptive reuse projects, built in former telecommunications hubs such as 111 Eighth Avenue in Manhattan or in former industrial plants such as 350 East Cermak in Chicago. And at least one has already gone around the cycle in the other direction: in downtown Denver, a 1960s data center was converted by Gensler and partners into a 12-story office building, The Link, which reused the existing steel, concrete, and masonry and embodies 68% less carbon than an equivalent new building. ## The present buildout We are now in the middle of one of the most consequential booms in a specialized building type in decades, and much of it is being designed with the same single-purpose logic. It is worth separating the question of obsolescence into layers, because the parts of a data center wear out — or, more accurately, become economically obsolete — on very different time scales. Stewart Brand, in _How Buildings Learn: What Happens After They're Built_ (Viking, 1994), building on the architect Frank Duffy's observation that a building is really several layers of longevity of built components, described six "shearing layers" — site, structure, skin, services, space plan, and stuff — each changing at its own rate, with structure lasting decades to centuries and building services turning over on the order of one to two decades. Brand's argument was that buildings adapt well when the slow layers do not obstruct changes to the fast ones. Data centers are an extreme case of that framework, because their fastest layer turns over far faster than anything Brand had in mind. **The computing hardware** turns over fastest. Data center GPUs physically last five years or more, but are typically replaced on economic grounds every two to four years because of rapid performance gains. Wang, Zhang, Tzachor, and colleagues, writing in _Nature Computational Science_ (2024, doi:10.1038/s43588-024-00712-6), estimated that generative AI — specifically large language models — could generate 1.2 to 5.0 million metric tons of e-waste in total by 2030, depending on the rate of adoption, with an annual stream of up to 2.5 million tonnes by 2030 under the aggressive scenario. They also estimated that circular-economy strategies — extending the life of existing equipment, reusing modules, and recycling — could reduce that e-waste by 16% to 86%. Those numbers are contested in both directions. Alex de Vries-Gao, constraining the estimate by what the chip supply chain can actually produce and using AI server lifespans of four to six years rather than three, arrived at 131 to 225 thousand tonnes per year of AI-server e-waste by 2030 — roughly an order of magnitude below Wang's high-scenario annual figure for the same year (_Resources, Conservation and Recycling_ 229 (2026), 108872). Part of the difference is that accelerators retired from frontier training are often moved to less demanding work rather than scrapped. Others argue that both studies count far too little of what is in the building. **The mechanical and electrical plant** — cooling, power distribution, backup batteries and generators, networking — is the part that the server-based estimates leave out. In September 2026 the Basel Action Network (BAN), an environmental advocacy organization that tracks the international waste trade, released the first part of a four-part white paper, _The Coming AI Waste Wave_ (Jim Puckett, "Part I: How Big Is the AI Waste Wave?"), which builds the estimate from the bottom up. For a reference 100 MW AI data center, BAN estimates about 7,000 tonnes of equipment, of which computing hardware — accelerators, servers, and racks — is only about 13% by mass; cooling (about 35%), power supply and distribution (about 34%), and backup power (about 15%) make up most of the rest. That corresponds to roughly 70,000 tonnes of equipment per gigawatt of capacity (a range of 62,000–77,000), a figure broadly consistent with the World Economic Forum's independent estimate that each megawatt of capacity embeds 60–75 tonnes of minerals, mainly in power and cooling systems rather than servers. BAN also assigns each category its own replacement cycle — from 2.5 years for accelerators to 8 years for power distribution — which works out to about 14,000 tonnes of equipment retired per gigawatt per year. BAN's long-range totals are highly sensitive to its assumed growth in capacity (8.8% a year through 2050, which BAN itself calls provocative) and should be read as scenarios rather than forecasts. But its more important result here is much less assumption-dependent: most of the material in a data center is in its power, cooling, and backup systems rather than its computing hardware, a distribution the World Economic Forum's independent estimate supports. The facility layer is also being pushed toward early obsolescence by power density. Across the industry as a whole, density is rising slowly: Uptime Institute's 2026 survey found that the most commonly reported rack density reached 11 kW for the first time, with the average (excluding facilities purpose-built above 30 kW) at 7.8 kW. But 13% of operators now report peak densities of 50 kW per rack or more, and the AI-training halls at the top of that distribution, which require liquid cooling, bear little resemblance in power and cooling to the typical hall of only a few years ago. A building whose electrical and mechanical plant was sized for the average cannot simply be refilled with the newest equipment. The operator may accept less usable IT capacity, dedicate only part of the hall to high-density loads, or undertake a substantial retrofit — and BAN notes that moving from 5–15 kW to 50–140 kW racks generally means ripping out power distribution rather than uprating it in place. Either way, the mismatch can accelerate the economic obsolescence of the facility plant, not just the hardware inside it. **The building shell** is the slowest layer, and the one that will outlast everything inside it. That is where the longer-term question sits, and it helps to be precise about what kind of obsolescence is at issue. The hardware becomes technologically and economically obsolete; the mechanical and electrical plant becomes functionally obsolete when power density and cooling requirements change; and the building can become economically obsolete _as a data center_ while remaining physically sound. Meanwhile the site itself may remain valuable, because an existing grid interconnection can be difficult and time-consuming to replace. Structural life and functional life come apart. ## When the shell itself becomes obsolete At some point, some of these buildings will no longer be needed as data centers. I do not know when, and I do not think anyone does. The demand for centralized computing could keep growing; efficiency gains in chips and models could reduce the infrastructure required per unit of computation, although growing demand could absorb those gains; or a substantial share of inference could migrate toward smaller, distributed, or on-device systems, leaving some large central facilities stranded. Each of these is plausible, and the history of computing — from the mainframe, to distributed personal computing, back toward centralization in the cloud — suggests that the pendulum does not stay in one place. The historical analogy that seems most apt to me comes from Buffalo. The concrete grain elevators along the Buffalo waterfront, which Reyner Banham made the subject of _A Concrete Atlantis: U.S. Industrial Building and European Modern Architecture, 1900–1925_ (MIT Press, 1986), were enormous, single-purpose, highly engineered structures, and their purpose disappeared with a change in the transportation system rather than any failure of the buildings themselves. Once the St. Lawrence Seaway opened in 1959, grain could move directly from the upper Great Lakes to ocean-going ships; before the Seaway, more than 200 million bushels of grain arrived in Buffalo in a good year, and afterward the total never exceeded 100 million. Half a century later, 13 abandoned elevators still stood on the Buffalo waterfront, because there was little else such structures could be used for. It is fitting that the University at Buffalo's current Banham Fellow, Nicolas Barrette, is studying the architectural obsolescence of data centers, arguing that if these structures are obsolescent by design, then decommissioning, material recovery, electronic waste, and the brownfields they leave behind need to be thought through now, with adaptive reuse in the conversation from the start. So some people are thinking about this. My concern is that the thinking is happening in schools of architecture and planning rather than at the point where it would change anything — the siting, design, and permitting of the facilities now under construction. ## What would designing for reuse mean? From an environmental engineering perspective, a few things seem worth considering now, while the buildings are still on the drawing board. _Assets worth preserving._ A decommissioned data center will typically have something most redevelopment sites lack: large electrical interconnection and substation capacity, fiber, heavy floor-load capacity, long structural spans, and, depending on the cooling system, potentially substantial water and sewer infrastructure. These are infrastructure assets with value well beyond their original use, even if they are privately owned, and planning for their reuse — whether for industry, district energy, distribution, or uses not yet imagined — is a different exercise from planning for demolition. _Liabilities worth anticipating._ Data centers also contain a great deal that will need managing at the end. BAN's inventory for a 100 MW facility includes 40 to 56 backup diesel generators of 16–18 tonnes each, with their fuel, lubricating oils, and exhaust fluid; battery banks that for a 10 MWh lead-acid installation weigh 250–330 tonnes; chillers and air handlers charged with HFC refrigerants (EPA's technology-transitions rule limits new data-center cooling equipment to refrigerants with a global warming potential below 700 from January 1, 2027); glycol coolants whose inhibitors and biocides may require testing before discharge; and, in facilities that adopted two-phase immersion cooling, fluorinated dielectric fluids, some of which fall within broad definitions of PFAS — a technology the industry has largely backed away from since 3M ended production of its Novec fluids. Add cooling towers and their water-treatment chemistry. Most of this is managed material contained in equipment, not contamination; the risk is that leaks, spills, abandonment, or incomplete decommissioning turn it into a remediation liability, and it is far cheaper to design for clean separation and removal than to characterize and remediate later. _Separable layers._ If the computing hardware turns over every two to four years and the mechanical plant every decade or so, then the building should be designed so that those layers can be replaced or removed without demolishing the shell — mechanical yards that can be cleared, services that can be stripped, and structural systems that do not depend on the equipment they currently support. This is Brand's principle applied to a building type that tests it harder than any other. _A shell that could become something else._ The features that make the current generation of facilities efficient as data centers — windowless walls, very deep floor plates, minimal interior subdivision — are precisely the features that made the late-twentieth-century office building and the regional mall difficult to convert. Designing facade panels that can later be opened for daylight, floor-to-floor heights compatible with other occupancies, and vertical risers that could carry plumbing would add cost now. Whether that cost is justified depends on how one values the option of future reuse, which is a question that the market, by itself, has historically answered poorly. _End-of-life planning, and financial assurance where warranted._ These are separable steps, and a jurisdiction could take the first without the last: require an end-of-life plan and a material inventory at permitting, set a decommissioning standard, and require financial assurance only where the scale or the liabilities warrant it. A particularly relevant precedent comes from utility-scale renewable energy. North Carolina, for example, now requires owners of utility-scale solar projects to submit a decommissioning plan and establish financial assurance (Session Law 2023-58), and Maine requires a performance bond, surety bond, letter of credit, or other acceptable financial assurance for the total cost of decommissioning (35-A M.R.S. § 3495). Older and more stringent models exist further afield: hazardous waste facilities under RCRA must demonstrate financial assurance for closure and for post-closure care, with post-closure cost estimates based on a 30-year period (40 CFR Parts 264/265, Subpart H); under the Surface Mining Control and Reclamation Act, a coal mining permit may not be issued until the operator has posted a performance bond covering reclamation (30 U.S.C. § 1259); and nuclear licensees must provide reasonable assurance that funds will be available for decommissioning (10 CFR 50.75). A data center is not a hazardous waste facility, a mine, or a reactor, and any obligation should be proportionate. But if we already ask this of solar farms, it is not obvious why a facility of this scale, with its own material flows and potential site liabilities, should be permitted with no end-of-life plan at all. ## Who should be asking? The postwar mall and the late-twentieth-century office tower were not designed for reuse, and we are now paying the cost of that — in conversion projects that are expensive, in buildings that cannot be economically converted at all, and in the embodied carbon of what gets demolished instead. This time we are committing extraordinary amounts of material, infrastructure, and embodied carbon to long-lived physical assets that serve an unusually fast-changing technology, and that is true whether demand for AI computing grows for fifty years or collapses in ten. We also have the advantage of having watched the earlier cycle play out, and of seeing the current wave of data centers being built in some cases inside the very buildings the last cycle left behind. The open question is whose responsibility it is to ask what these buildings will become: the developers and operators, whose investment horizons may give little weight to a reuse option decades away; the local jurisdictions that site and permit them, which are only beginning to impose conditions; or the engineering and planning professions, which ought to be able to describe what a reusable data center would look like before the next generation of empty buildings makes the case for us. ## Coda The author outlined the arguments and approach. Research and drafting were performed by Claude (Anthropic). The draft was reviewed by ChatGPT (OpenAI) and Gemini (Google). Final revisions were made by the author, who assumes responsibility for the work. ### Share this: * Share on X (Opens in new window) X * Share on Facebook (Opens in new window) Facebook * Share on LinkedIn (Opens in new window) LinkedIn * Email a link to a friend (Opens in new window) Email * Like Loading... ### _Related_

City staff presented options for relocating and reusing the May Hicks Curtis House, identifying two preferred sites (311 W. Cherry and 11 W. Cherry) and three primary reuse concepts (residence, small office, cafe/bar). Cost estimates approach $1 million and the Arizona 250 Commission has pledged $25,000 toward a downtown cafe/bar option if the city commits by March 31.

The Chester City Council moved ordinance B10 on its first reading to remove several zoning barriers — including an 800‑foot interstate buffer and a first‑floor commercial requirement — that staff said are hindering new housing and adaptive reuse.