A 2015 colocation facility of meaningful scale ran on a hundred to two hundred humans across operations and security. A 2026 hyperscale AI campus of comparable footprint runs on thirty to eighty. The campus designs already in glass for 2028 commissioning target fewer than twenty humans plus an agent stack. That is not a productivity improvement. That is a different operating model wearing the same physical shell, and the gap will keep widening because the inputs feeding it (sensor density, model quality, robotic dexterity, regulatory permission for autonomous patrol) are all on independent improvement curves that compound against each other.
The operational cost story is the obvious one and the least interesting. The interesting story is what happens to competitive position once a few operators have campuses that learn from every site they run, and most operators do not. Bespoke campuses stay bespoke. Campuses inside an operating data flywheel stop being individual assets and become nodes in a fleet that improves with every megawatt-hour of runtime.
What A 2015 Colo Actually Ran
A 2015 colocation site at fifty to a hundred megawatts of critical IT load was a labor-dense building. Twenty-four hour staffed operations centers. Two to four electricians per shift. HVAC and mechanical technicians around the clock. A dedicated security team running access control, badging, lobby, and patrol. Remote-hands crews for tenant work tickets. Yard logistics handled by a receiving team. Pest control, janitorial, landscaping all on contracts that required physical supervision. Cooling tower chemistry checked by a human with a clipboard. Battery string maintenance handled by a battery vendor on a quarterly cycle, in person, with a wrench.
The building management system existed but it was a passive observer. It flagged anomalies and a human read the flag and decided what to do. Setpoints were static. Load was balanced by spreadsheet at the rack and PDU level. Switching events were planned in conference rooms and executed by a team in a control room reading off a printed runbook. Predictive maintenance was a phrase in a sales deck. Actual maintenance was scheduled by calendar, not by condition.
A hundred to two hundred humans was not bloat. It was the staffing required to keep that building safe and within SLA given the tools available at the time. The labor was the runtime.
What A 2026 Hyperscale Facility Already Looks Like
Walk a current-generation hyperscale build and the staffing model has already shifted. Building management systems from Vertiv, Schneider Electric, Eaton, and ABB run continuous AI-augmented load balancing and thermal control across thousands of zones. Setpoints move in real time against weather, IT load forecasts, and utility tariff signals. Chillers stage themselves. Free cooling decisions happen at the second-by-second level without a human reading a graph. Cooling tower chemistry is monitored by inline sensors that flag deviation before a human would notice it on a clipboard.
Grid-edge software from vendors that grew out of utility automation (Vigilant Energy, Sentient Energy, GE Digital, Hitachi Lumada) coordinates the campus with the local distribution feeder. Battery energy storage systems run optimization layers built on platforms like Voltus, Stem, Tesla Autobidder, and AutoGrid (now inside Schneider after the 2022 acquisition) that arbitrage tariff windows, shave demand peaks, and participate in ancillary services without a trader watching the screen.
Predictive maintenance has moved from sales deck to deployed reality. Hitachi Lumada, Siemens MindSphere, Augury, and the successor stack to GE Predix monitor vibration, acoustic signatures, partial discharge, and thermal patterns on transformers, switchgear, generators, pumps, and chillers. Failure modes are flagged days to weeks ahead. The wrench-turning is still human, but the diagnosis is not.
Perimeter and infrastructure inspection has gone autonomous in pilot and limited production at multiple operators. Boston Dynamics Spot units patrol industrial yards under commercial deployments at BP, Ford, and OnePower among others, doing routes that a security guard or technician used to walk. ANYbotics ANYmal handles inspection in hazardous and confined areas where humans were the bottleneck. Skydio autonomous drones cover perimeter, rooftop, and overhead infrastructure inspection on programmed routes without an operator on sticks. Computer vision threat detection on the existing camera estate flags intrusions, climbing attempts, and anomalous loitering at a precision that retired the false-alarm problem that made the prior generation of analytics unworkable.
Thirty to eighty humans on a hyperscale AI campus is not a hollowed-out building. It is a building where each human is doing supervisory and exception work, not continuous monitoring. The agents handle continuous monitoring. The humans handle the cases the agents escalate.
The Stack Coming Online For 2028
The 2028 design point assumes everything in the 2026 list is table stakes and adds a layer of operational autonomy that does not yet exist in commercial deployment but is in field trial today.
Autonomous transformer and battery cell monitoring at the individual cell level. Cooling thermodynamics managed by models that own the full loop from chip junction temperature to evaporative tower exhaust, with no static setpoints anywhere in the path. Yard logistics handled by autonomous ground vehicles moving spare parts, replacement modules, and waste streams between buildings on a campus that was designed for that traffic pattern rather than retrofit for it. Robotic inventory inspection inside white space, walking aisle by aisle, reading asset tags, flagging missing or moved equipment, doing the audit that humans currently do quarterly.
The vendor stack is mostly named already. Vertiv and Schneider on infrastructure intelligence. Boston Dynamics and ANYbotics on quadrupeds. Skydio on drones. Voltus, Stem, and the post-acquisition AutoGrid layer on energy markets. Augury on rotating equipment. The integration layer is where the moat sits, because the value is not in any single sensor or any single autonomous unit. It is in the operating data the campus generates and the models that get trained on that data and the next campus that ships with those models already running.
This is what makes a 2028 campus different from a 2026 campus with more robots in it. The 2026 campus uses autonomy. The 2028 campus is built around it.
The Tesla Caution
The over-automation lesson is worth carrying into this. Tesla's 2018 Model 3 ramp at the Fremont plant ran into the wall every operator should worry about. Elon Musk's original vision was the "alien dreadnought," a factory so automated that humans were the friction. The actual line had to be partially de-automated to hit production. Conveyors came out. Humans went back in. The reason was not that automation was wrong in principle. The reason was that the failure modes of the automated system were more expensive than the labor of the humans they replaced, because every novel failure required an engineering response and the line stayed down while engineering responded.
The correct read is not that automation does not work. The correct read is that the boundary between automation and human supervision is the design problem, and getting it wrong on the optimistic side is more expensive than getting it wrong on the conservative side. The campuses that ship in 2028 with sub-twenty headcount will have learned this. The ones that ship the same model with the same headcount target but without the operational data depth behind it will hit the Fremont wall in a different building.
The right framing is that autonomy compresses labor but expands the engineering and data layer behind it. The total cost of operations curve bends down, but the composition of that cost shifts hard from labor to software, sensors, training data, and the people who maintain the models.
The River Rouge Parallel
Henry Ford opened Highland Park in 1910 and the moving assembly line by 1913. Build time on a Model T fell from over twelve hours to roughly ninety minutes. Highland Park was the most studied factory on Earth. By 1928 Ford had moved the operation to the Rouge complex in Dearborn, a 1.5 square mile site where iron ore arrived on Great Lakes freighters at one end and finished automobiles drove out the other. Coke ovens, blast furnaces, a steel mill, a glass plant, a power plant, and final assembly all sat on a single integrated campus connected by mechanized transport between stations. The campus replaced the building because the operating model required it. Highland Park had hit the ceiling of what a single building could integrate. Rouge was the next architecture.
The relevant detail is not the scale. It is that each generation of automation at Ford compounded against the prior generation. Highland Park's moving line made Rouge's continuous-process integration possible, because the throughput discipline learned at Highland Park became the operating assumption at Rouge. A competitor trying to build a Rouge-equivalent from a craft-shop starting point could not skip Highland Park.
The Bell Labs and AT&T system from the 1930s through the 1960s is the same shape on the information side. Each generation of switching technology was built on the operating data from the prior generation. Crossbar switches were possible because step-by-step switches had been operated at scale for a generation and the failure modes were known. Electronic switching was possible because crossbar operations had been instrumented for a generation. The system compounded because the operator was the same.
Data center campuses are at the Highland Park to Rouge transition. The campuses being designed for 2028 commissioning are the next architecture. The operators that own multiple campuses and instrument them as a fleet are accumulating the equivalent of Ford's throughput data and Bell's switching data. That data is the asset. The buildings depreciate. The data compounds.
What Becomes True By 2028
Three things are likely to be true by 2028 that are not yet true.
First, autonomous operations will be a documented underwriting input. Lenders and equity will price campuses differently based on whether the operator has a real operating data flywheel or is running a bespoke site. Insurance will follow. The gap will show up in cap rates and in cost of capital before it shows up anywhere more visible.
Second, campuses without operating data flywheels will be increasingly dependent on the operators that have them. The dependency takes the form of management contracts, operating partnerships, and outright fleet acquisitions. The pattern resembles what happened in commercial real estate when institutional operators absorbed family-office assets through operating agreements rather than equity transactions. The asset stays on the original balance sheet. The operating intelligence does not.
Third, operating data itself becomes a tradable asset class. Data-rights deals on industrial campuses are starting to look like the early-stage version of what mineral rights deals did in oil and gas. The land has one set of cash flows. The operating data layered on top has another set. The two can be separated, and once they can be separated, they will be priced separately. This is an emerging category and the documentation is thin, but the structure is being negotiated in private transactions today.
The Aramco facility automation case from oil and gas and the Amazon fulfillment center automation curve from Kiva through Sparrow into the current autonomous mobile robot fleet are both proofs of the same pattern. The operators that compound operational data across a fleet pull ahead of the operators that do not, and the gap widens at a rate that is difficult to close once it opens.
The data center industry is converging on this architecture faster than the public discourse acknowledges. The 2028 campus is being designed now. The headcount target is set. The vendor stack is in field trial. The competitive sorting has already begun.