From Vacant Banks to Payments Innovation: Why Data Center Strategy Is Splitting in Two

Two data-center products, one skyline: greenfield AI campuses for training scale, and vacant bank buildings as the regulated last mile of authorization, fraud, and settlement.

The data center boom is no longer one market with one playbook. Hyperscale AI training wants cheap land and megawatts. Instant payments, fraud scoring, card authorization, and settlement want milliseconds, regulated premises, and fiber that already exists.

That split is the strategy problem. Most of the planned U.S. pipeline is racing into rural greenfield sites. Most of the payments workload still lives in cities, on bank cores, near card networks, and inside buildings designed to protect money.

Treat those as the same product and you overbuild the wrong thing. Separate them and you can own the next layer of payments infrastructure.

The market has already voted with a shovel

Existing U.S. data centers are still urban. Pew Research found 87% of operating facilities in urban areas. The pipeline is the opposite: about 67% of planned sites are rural, and 39% sit in counties that currently have none.

That flip is a power-and-land story, not a fashion.

• Greenfield construction in major markets is running roughly $10–16 million per MW. Adaptive reuse is often cited at 30–50% lower capital cost, with renovation in 6–12 months versus 24–48 months of new-campus construction.

• Land purchases for future data centers hit about $6 billion in the first half of 2026, up 79% year over year. In constrained markets, powered land has traded at several million dollars per acre. At least nine states have considered moratoriums.

Power remains the gate for hyperscale. Fiber is reasserting itself as a second gate once a campus leaves existing long-haul routes. Interconnection queues in some grids now stretch toward seven or eight years. Behind-the-meter generation is no longer an experiment; it is how operators buy time.

The industry default has been: leave the city, buy a farm, wait for the utility. That is correct for training clusters measured in hundreds of megawatts. It is a poor answer for the payments stack.

Two products, two geographies

Treat “data center” as one asset class and the analysis collapses. Split the workload and the site choice becomes obvious.

Workload	What it actually needs	Where it belongs
AI training / batch HPC	Cheap firm power, land for phased GW campuses, liquid cooling from day one	Rural or industrial greenfield / power-plant brownfield
Cloud origin / storage	Power + fiber + tax incentives; latency is secondary	Secondary markets and powered land
Instant payments, auth, fraud, FX, wallet balance	Sub-second SLAs, 24/7 uptime, carrier diversity, regulated physical security	Urban / metro edge, financial districts, existing FI sites
Exchange-adjacent trading	Microseconds to a matching engine	Specific colo buildings, not generic rural halls

Rural greenfield wins on scale. Adaptive reuse wins on time-to-revenue, interconnection, and politics. Payments innovation lives in the second column.

Rural greenfield: the case for building from scratch

Pros

1. Design freedom. New halls can be specced for 50–100 kW racks, liquid and immersion cooling, 5–6 meter clear heights, and 150+ psf floor loads. Existing offices and bank branches cannot.

2. Land assembly. A 200–800 acre campus can phase substations, generators, water systems, and future halls. A corner bank lot cannot.

3. Power adjacency. The best rural sites sit on 345 kV corridors, near generation, or on retired industrial switchyards. A retired coal plant’s interconnection can be worth more than the dirt.

4. Isolation. Fewer neighbors, fewer historic-district fights, easier generator and fuel logistics.

Cons

1. Time-to-power is now the product. Entitlements, transmission upgrades, and transformer lead times can erase the “cheap land” advantage. A two-year revenue delay on an AI hall is expensive. On a payments platform that already has to be live on FedNow, RTP, and card networks, it is fatal.

2. Community backlash is schedule risk. Ratepayer fights, farmland preservation, water rights, and local moratoriums are no longer PR problems.

3. Fiber follows last. A 10-mile lateral can take six months. A 200-mile route can take two to three years. Payments traffic is intolerant of a single long-haul path.

4. Wrong latency profile. Training can live 300 miles from users. Card authorization, account-to-account instant payments, and real-time fraud models cannot treat 20–40 ms of extra propagation as a rounding error when the customer SLA is “did the tap work.”

Rural greenfield is the right factory for compute. It is a weak branch for money movement.

Adaptive reuse of vacant commercial real estate

Vacant U.S. commercial stock is not one pile. A Class B office tower, a paper mill, a big-box store, and a former bank operations center fail or succeed for different reasons.

Pros of reuse

• Faster entitlements on already-zoned commercial or industrial sites — often 3–6 months, not a multi-year rezoning war.

• Lower capex when the shell, slab, and utility laterals can be kept.

• Better politics and embodied-carbon math than paving farmland.

• Industrial sites already sit on substations. Urban commercial sites already sit on fiber laterals and carrier hotels.

• Inference, payments, and SaaS still want to be near users and peering.

Cons of reuse

• Structure fights density. Typical office floors are designed for roughly 50–80 psf. Modern data halls want 150 psf and up; AI zones can run several times that. Office ceiling heights are often a full story too short.

• Power is not “already there” just because the building has service. A 2,000-square-foot bank on 800-amp service is four houses, not a colo.

• Cooling retrofit is the expensive surprise. Rear-door heat exchangers can land. High-density liquid loops need structural review, roof rights, and often a new plant.

• Most office conversions fail the power / structure / height screen. The winners are industrial shells, purpose-built telecom buildings, and a narrow set of financial operations facilities.

The correct reuse thesis is not “turn every empty office into a GPU farm.” It is “match the building to the workload it can actually host.”

Focus: banks and other financial institution properties

This is the overlooked inventory.

Over the last decade, U.S. banks closed on the order of 9,000 branches and 22–36 million square feet of branch space. The national branch count fell from a peak near 82,000 in 2009 to roughly 69,000 by 2025. Closures have slowed — and even reversed slightly in some Sun Belt markets in 2026 — but the overhang is still real: standalone corner lots, drive-throughs, vaults, and a separate, more interesting class of assets — operations centers and retired bank data centers.

These are not the same product.

Retail branches: micro-sites, not megawatts

A typical closed branch is 2,000–4,500 square feet on a high-visibility corner. Comcast’s conversion of a vacant TD Bank in Clinton, New Jersey is the honest example: about 2,170 square feet, unmanned, 800-amp service, two emergency generators, condensers, a sound wall. That is a cable node or a tiny edge room, not a hyperscale hall.

What a branch actually gives you: hardened construction and a vault; alarm and access-control heritage; a parcel the town already treats as a financial/utility use; parking and a generator pad; location inside the deposit franchise — near the customers whose payments you want to clear.

What it does not give you: meaningful IT load, dual high-voltage feeds, raised-floor volume, or room for liquid cooling at AI density.

Use case: edge POP, CDN cache, ATM/ISO processing closet, card-network access node, or a quiet meet-me room for a community bank’s core vendor. Do not pitch it as an AI campus.

Operations centers and former bank data centers: the real prize

This is where the payments thesis gets serious.

• A former Truist / SunTrust facility in Durham listed as a 122,000-square-foot powered shell on 15 acres, with 10 MW available and a path toward 25 MW, dual Duke Energy feeds, dual fiber entrances, and generator pads already stubbed.

• Vault Data Centers bought a former downtown Lynchburg, Virginia bank building specifically to operate as a data-center services site.

• Flagstar, after absorbing multiple bank stacks, did not keep six captive halls. It collapsed into Equinix colocation — a reminder that running a hall is not the same as needing proximity to one.

Bank operations campuses were built for a regulated, always-on business. That heritage still shows up in the bones: higher floor loads than a generic office; redundant power and cooling plants, even if the gear is aged out; diverse fiber into downtown and suburban ops sites; physical security that already satisfies examiner language; and staff who already know change tickets, dual control, and evidence packages.

Those attributes map onto payments infrastructure better than they map onto training clusters.

Why FI real estate is a payments advantage, not just a cheaper shell

Instant rails changed the building requirement.

FedNow, The Clearing House RTP, Visa Direct, Mastercard Send, wallet ledgers, and ISO 20022 orchestration all assume 24/7 availability and end-to-end times measured in seconds, not batch windows. Modern payment platforms now talk about thousands of transactions per second with sub-second mean latency and near-zero data loss on node failure. Fraud and AML models have to score the same payment before it is released. That is inference at the edge of the transaction, not a model trained last week on a farm in another interconnection queue.

A payments leader does not win by owning the cheapest rural megawatt. It wins by placing authorization, ledger, and risk compute next to bank cores and processor hosts, on diverse metro fiber, in rooms an examiner will walk without a science project, and close enough to card networks, Fedwire access points, and ACH operators that the network hop is boring.

Former FI buildings already sit on that map. They also sit on a trust map. A vacant bank that becomes a payments edge node is a conversion the town already understands: quiet building, generators behind a fence, no retail traffic. A 500 MW hall on irrigated farmland is a different political object. Banks that finance data centers have already learned the reputational cost when the community links the brand to a power fight. Reusing the institution’s own stranded real estate flips that narrative.

How this makes banks the leaders in payments innovation

The strategic move is not “banks should become Equinix.” Most banks should not operate halls. Flagstar’s conclusion — no competitive advantage in owning the building — is right for a community or regional franchise.

The advantage is in controlling the topology.

1. Own the last mile of money movement.Whoever places low-latency compute between the merchant, the issuer, the network, and the core owns the service-level conversation. That is authorization, tokenization, account-to-account instant payments, request-to-pay, and disbursements. Rural training capacity can make the fraud model smarter. It cannot make the tap faster if the decision still hairpins through Virginia.

2. Productize “payments-grade” edge.A converted operations center — or a disciplined branch-to-edge program — can be sold as regulated-adjacent colo: PCI rooms, dual control, visitor management, private cages for processors, ISO sponsors, paymasters, and RWA platforms that cannot live in a generic public-cloud region. That is a product hyperscalers do not naturally sell.

3. Pair rural training with metro inference.Train and fine-tune fraud, credit, and AML models where power is cheap. Serve those models from metro FI sites and carrier hotels so the inference hop stays inside the payment’s latency budget. Keep the system of record and the settlement connection on diverse urban fiber. Trading firms have used this pattern for a decade: heavy research off-exchange, execution next to the matching engine. Payments can copy the pattern without copying the microsecond arms race.

4. Turn stranded bank real estate into a network effect.A national or super-regional franchise that closed 200 branches still has 200 corners in the deposit map. Ten of those, chosen for fiber and power, become a private edge mesh. Processors, fintech sponsors, and card-issuing programs will colocate to the mesh because that is where the banks already are. The bank or payments utility that organizes the mesh becomes the default interconnection point.

5. Win the examiner and the city at the same time.Reuse of a known FI building shortens two queues that usually kill projects: local land use and regulatory comfort. You are not introducing a new industrial use. You are changing the contents of a building that was already treated as critical infrastructure.

6. Keep the balance sheet honest.Large banks are already on both sides of this trade: they finance well over $100 billion a year of data-center construction, and they are consolidating their own halls into wholesale colo. The innovation prize is the third role — originating and anchoring payments workloads in reused FI sites — without pretending every closed branch is a 10 MW opportunity.

A practical checklist, not a slogan

Before anyone buys a vacant bank or an empty ops center, run a fail-fast checklist to evaluate:

1. Power path. What is available now, what is reservable in 12 months, and what requires a new feeder? If the answer is “800 amps and a prayer,” it is an edge closet, not a hall.

2. Fiber diversity. Two providers, two laterals, two conduits. One cable in a shared duct is not a payments site.

3. Structure. Floor load, slab-to-soffit height, roof rights for dry coolers, and a place to put generators without violating noise and setback rules.

4. Workload fit. Branch = edge. Ops center / former DC = metro colo. Power plant / mill = hyperscale. Do not mix the labels.

5. Flood, fire, and exam path. FI buildings in basements and river towns fail for the same reasons banks eventually left them.

6. Counterparty. Who is the tenant — a processor, a bank core, a card-network access node, an RWA settlement venue? The tenant defines the spec. The building does not.

   
   

The strategy, stated plainly

Build rural when the product is megawatts.

Reuse commercial real estate when the product is time, fiber, and proximity.

Reuse bank and financial institution real estate when the product is money movement.

The leaders in payments innovation will not be the firms that announce the largest campus in the least populated county. They will be the firms that put authorization, risk, and settlement compute back into the buildings the payment system already trusts — and treat the rural AI factory as a supplier, not as the storefront.

The vacant bank on the corner is not a hyperscale site. Used correctly, it is something scarcer: a regulated, local, low-latency door into the payment system.