.png){kind=logo}
There's something funny about it. The industry that measures everything — tokens, latency, GPU utilization down to the millisecond — spent decades putting its hardware inside buildings assembled from paper drawings, hallway conversations and a superintendent's memory. That gap is closing fast, mostly because it has to. Specialist modeling groups like SJS electrical VDC teams now sit inside the design process months before a shovel moves, and what they hand over isn't a picture of a building. It's a queryable dataset.
Racks that used to draw 6 or 8 kW now pull 50, 80, sometimes north of 130 in AI training halls. Cooling changed. Distribution changed. Schedules got shorter while the scope got heavier. You can't coordinate that with markups.
Five years ago a data center project was already complex. It just wasn't this compressed.
Hyperscale campuses are now planned in gigawatts, phased in halls, and repeated across sites with near-identical layouts. Liquid cooling is showing up in real projects, which reshuffles the mechanical picture and takes space that electrical assumed it had. Meanwhile the delivery date is tied to a compute commitment somebody already sold.
That's the real pressure. A month of slip on an office tower means a month of extra rent somewhere. A month of slip on an AI hall means racks of expensive silicon sitting in a warehouse doing nothing, while the tenant burns cloud spend elsewhere. The cost of being late is asymmetric, and everyone on the job knows it.
So the question stops being "can we build this" and becomes "how early can we see the problems."
Ask most people outside construction what BIM produces and they'll say a 3D model. Fine, but that's the visible part. The useful outputs are messier and more numeric:
The federated model, where each trade's geometry gets stitched together
clash reports, run weekly or nightly depending on how disciplined the team is
RFI and submittal logs, with dates and aging
procurement and long-lead tracking, tied to specific pieces of equipment
labor hours booked against zones and systems
laser scans and reality capture, comparing what got built to what was drawn
and, later, sensor and commissioning data from gear that's already energized
None of that is exotic. What's changed is that these streams now connect. A clash in zone 4 links to an RFI, which links to a submittal revision, which links to a delivery date, which links to a crew that has nothing to do next Tuesday.
This is the framing that makes sense to anyone who works with data. Every object in a coordinated model carries attributes — system, size, elevation, install zone, status, sometimes cost code. Geometry is one field among many. Once the attributes are consistent, you can run queries against the building.
That's the whole trick. And it's also where projects fail, because attribute discipline is boring work that nobody celebrates.
Clash counts behave like a defect curve. Plot them per zone, per week, and the shape tells you things a status meeting won't.
Falling counts mean coordination is converging. Flat counts mean people are arguing instead of deciding. Rising counts, three months into coordination, usually mean the design is still moving underneath everybody — a ceiling height got value-engineered, or the equipment vendor changed and nobody updated the footprint.
The hard clashes are easy. Conduit through duct, tray through beam, geometry hitting geometry. Software finds those. The expensive ones are soft: a run that's technically clash-free but leaves no room for a hanger, an access path to a panel that disappears once the ceiling grid goes in, a maintenance clearance that exists on screen and not in reality. Those need a human who's been on a jobsite. They're also the ones that turn into field improvisation and overtime.
Model quantities feed buyout. On data centers this matters more than usual, because switchgear, busway and transformers have lead times measured in quarters, not weeks.
When the model is trustworthy at 40% design, the procurement team can commit early with a small contingency. When it isn't, they either over-order or wait — and waiting is how a project loses its float before construction even starts. Early quantity confidence is worth more than a perfect model at 90%.
A data center is an electrical building with a roof on it.
Count the objects. Distribution rooms stacked with gear, busway runs feeding rows, redundant feeder paths that have to stay physically separated, cable tray corridors, grounding, lighting, fire alarm, the whole low voltage layer. Electrical usually has the highest object count in the federated model, the strictest clearance rules — NEC working space isn't negotiable — and the least ceiling space left over, because mechanical got there first with duct mains that don't bend.
There's also a structural quirk in how the work happens. On most projects, the bulk of electrical modeling falls on the installing contractor rather than the design engineer. So if that scope gets treated as filler inside a general MEP effort, the data quality drops exactly where the risk is highest. Teams that assign electrical modeling to a dedicated group tend to get cleaner routing logic across revisions, and coordination decisions that survive the next set of drawings.
The dashboards look different too. Instead of counting clashes generically, you're tracking feeder route stability, room-level gear layouts, and whether the busway centerline has moved since the last vendor submittal.
Repeatable halls mean repeatable assemblies. That's the opening for prefab, and data centers are close to an ideal case — the same rack row, the same overhead configuration, forty times.
Shops that run this well measure a handful of things:
shop hours versus field hours for the same scope
spool rework rate — how many fabricated assemblies came back wrong
kit delivery hit rate against the install sequence
install duration per row, tracked week over week
The uncomfortable truth is that prefabrication multiplies whatever your model quality already was. Accurate model, and prefab becomes a genuine schedule weapon. Drifting model, and you're paying a shop to build scrap with excellent efficiency.
Linking model objects to schedule activities gives you a simulation of the build. Run it, and sequencing problems show up as visual nonsense — crews stacked in one room, a wall closing before the gear behind it is set.
Cost loading takes it further. Earned value pulled from model status is closer to reality than a percentage estimated in a trailer on a Friday afternoon. Not perfect. Closer.
The version that works in practice is unglamorous: install status colored by zone, updated weekly, projected in the coordination meeting. Green means installed, yellow means in progress, red means blocked and here's why. Nobody needs a rendering. They need to know which rooms are stuck.
Open clashes by trade and zone. A zone that stays red for a month is a decision nobody is making.
RFI aging. The median matters less than the tail. Anything over 30 days is quietly eating your schedule.
Model status versus planned install curve. Divergence here is your earliest warning of a slip.
Prefab kit throughput. Falling throughput usually means the shop is waiting on coordination.
Long-lead risk. Which equipment, which vendor, what's the current promise date, how many times has it moved.
As-built completeness. Tracked continuously, or it becomes a panicked scramble at closeout.
Data hygiene is the whole game and it's thankless.
Models drift. Hangers and insulation thickness go unmodelled, so the corridor that looked fine is 4 inches tighter than the file says. Field crews stop updating status because they're busy installing, and within three weeks your dashboard is confidently wrong. Scope changes outrun the model, then somebody makes a decision from stale geometry.
Garbage in, expensive rework out. Same as any analytics pipeline, except the errors get poured in concrete.
The teams that get value here have someone whose job is guarding the dataset — reconciling scans against the model, chasing status updates, killing broken families before they spread. It's maintenance work. It just happens to protect the largest cost line on the project.
The interesting part starts after the handover. A coordinated as-built model becomes an asset register: every panel, every piece of gear, tagged, located, with clearances and access paths recorded. Facility managers are getting sharper about demanding this, and they're right to.
Capacity planning for phase two gets easier when you can see exactly what's installed and where the spare pathway space sits. Maintenance planning stops depending on whoever has been there longest. And when the next density jump arrives — because it will — the team retrofitting the hall isn't starting from a folder of PDFs and a guess.
