What data-centre demand changes on a busway shop floor

Plenty has been written about why data centres are moving to overhead busway. Almost none of it is written for the people who have to build the stuff. What follows stays inside the factory: what changes in your fabrication process when the busway you are quoting is going into an AI hall rather than a warehouse.

The demand context, briefly

Two numbers frame everything that follows. The legacy design baseline for a server rack sat at 5 to 15 kW, and that number held for roughly two decades because it was set by what perforated-tile air cooling could remove from a 600 × 1200 mm cabinet. AI-optimised racks have broken it. Current GPU-dense cabinets draw 60, 80 and 100+ kW, and rack-scale platforms are being specified above that.

Two consequences reach the shop floor. Overhead busway is displacing cable tray in these halls, because tap-off units can be relocated without re-terminating branch circuits and because a busway run occupies far less volume than the equivalent parallel cable sets. And tap-off and joint ratings that used to be exceptional are now routine: 400 A continuous and above is a common specification where 100 A used to be typical.

Everything below follows from that second number. A joint carrying 400 A instead of 100 A dissipates sixteen times the power for the same contact resistance, inside the same enclosure, next to the same insulator.

Joint-plane tolerance stops being generous

The fabrication consequence of higher joint current is that the tolerance band you used to work inside was not a quality target, it was thermal margin. At 100 A you could afford a joint that made contact over 70% of its nominal area. At 400 A you cannot, because the temperature rise scales with I² while your contact area has not moved.

Three dimensions carry that load in a sandwich busway joint.

The conductor offset. Each phase bar is joggled from its run plane to the plane the joint block expects. If one bar sits half a millimetre high, its fish plates enter skewed and the Belleville stack loads one edge of the contact patch harder than the other. The metal that carries current at a bolted interface is only the asperity tips that actually touch, and how many of them touch depends on pressure distribution, so uneven pressure means less conducting area for the same bolt torque.

The fish-plate hole position. Every plate in the block shares one bolt. An off-position hole cannot centre on that bolt, which produces the same eccentric loading from a different cause. This is why a fish-plate machine is specified on hole pitch accuracy rather than on tonnage, and why holding better than ±0.1 mm/m matters more than the 200 kN it can put through the punch.

Flatness. A fish plate with 0.2 mm of bow across its length does not flatten under 18 kN of bolt load in the way people assume, because the plate is stiff in that direction and the Belleville washers are designed to accommodate thermal movement, not to press bowed parts straight.

None of these are new requirements. What has changed is that the margin that used to hide them is gone.

There is a fourth dimension that gets blamed on the forming machine and usually is not its fault: incoming stock condition. A conductor delivered with 1 mm of twist over its length, or with edge camber from a poorly set slitting line, arrives at the joggling station with a datum that is already wrong. The machine will faithfully put the offset where the program says relative to the clamp, and the bar will still not sit flat in the stack. If joint-plane complaints are appearing at a rate that does not correlate with machine maintenance, work through the incoming inspection record before anyone touches the servo tuning.

More joints per run, and each one is a serial dependency

Busway is built in standard section lengths, typically 3 m, so a run's joint count is roughly its length divided by three. A 60 m run has about twenty joints. Every joint is in series: every joint's resistance adds to the run's voltage drop and every joint's temperature rise is a candidate hot spot in an infrared survey.

Higher density does not just raise the current in each joint. It raises the number of runs, because a hall that used to be fed by four runs now needs twelve, and it shortens the distance between distribution points. The result on the shop floor is a higher joint count per metre of shipped product than the same factory was producing five years ago.

That changes where the bottleneck sits. If joint-end forming and fish-plate production were previously side operations feeding a line whose pace was set by housing assembly, they are now on the critical path. It also changes the arithmetic on error accumulation. A systematic pitch error of 0.1 mm/m contributes 0.3 mm per 3 m section; across twenty sections pushing the same way that is 6 mm, which joint clearance absorbs. At 0.5 mm/m the same run drifts 30 mm and the final section will not meet a fixed building interface. Random error would go as the square root of the count, but machine error is rarely random. It comes from a lead screw or an encoder, so it accumulates linearly.

Tap-off windows are a forming operation

Plug-in busway carries tap-off openings along its length, and the density of those openings is one of the things data-centre clients push on, because the whole argument for busway is that a rack can be moved and re-fed without electrical alteration. Published plug-in ranges have historically offered around five tap-off points per 3 m length. Where cabinets are being re-provisioned frequently, that pitch tightens.

Each opening is more than a slot in the housing. The conductor behind it has to present a contact face to the tap-off unit's jaws, and in several published designs that face is a tag pressed out of the conductor itself rather than a separate riveted contact. Forming a tag out of a 6 mm conductor is an embossing operation with the same considerations as the joggle: local work hardening, springback that shows up as a height error rather than an angle error, and a tolerance that is set by the mating jaw rather than by anything on your drawing.

Two things follow for production. The tag face is a contact face, so it goes through the same plating specification and the same handling discipline as the joint faces, and any rework that touches it destroys that finish. And the tag positions along the bar share a datum with the joint-end features at both ends of the same conductor, which means they want to be produced in the same clamping or from the same programmed origin. Producing tap-off features on a separate machine with its own fixture reintroduces the transfer stack-up that the rest of the process is trying to eliminate.

Plating moves from optional to specified

Bare copper oxidises. Cu₂O is semiconducting and CuO is an insulator, and a couple of micrometres of either at a contact interface is enough to be measurable at these currents. The failure mode is a loop: oxide raises contact resistance, resistance raises joint temperature under load, and higher temperature accelerates oxidation.

At 100 A that loop takes a long time to close, which is why bare-copper joints with a grease and a good torque figure served the industry for decades. At 400 A it closes faster, and in a data hall the consequence of it closing is not an outage on a lighting circuit. Specifications for high-density busway now routinely call for tin or silver plating on the contact faces, and increasingly on the fish plates as well as the conductors.

That has direct process consequences upstream of the plating tank.

Surface preparation becomes a controlled operation rather than a hand operation. Plating adhesion depends on the oxide and the rolling scale being removed uniformly, and a two-stage deburring and surface-conditioning machine does that to a repeatable finish across a 100 to 800 mm width in a way that a bench belt sander does not. The same pass removes the punch burr, which matters because a burr under plating is a locally thin coating over a locally high field.

Handling after plating becomes a defined step. A plated contact face that is dragged across a steel bench or stacked without interleaving arrives at the joint with the plating worn through at exactly the points that carry the load.

And the process order becomes fixed. Deburring before plating, plating before assembly, no rework of a plated face. Any shop that has previously formed, assembled, found a fit problem and reworked the part on the bench needs to design that loop out, because the rework destroys the finish the specification was written around.

Throughput and the cost of a changeover

Data-centre busway orders arrive as large quantities of a small number of configurations, delivered against a construction programme that does not move. That is a different problem from switchgear work, where the mix is wide and the quantities are small.

The lever is the fraction of the shift the machine spends cutting. Two things dominate it, and neither is spindle speed.

First, combining operations in one clamping. If a conductor is cut to length on one machine and joggled on another, the length datum and the offset datum are different fixtures, so every part carries a transfer stack-up on a dimension with no slack. A machine that saws and forms in the same clamping removes that error and removes a handling step. Busduct forming machines that do both are sized around it: 300 kN forming force against a 6,000 × 300 × 6 mm sheet capacity, with the X axis running at 48 m/min so that positioning between operations is not the cycle-time limit.

Second, program-to-part time. On repeat production this looks irrelevant, and it would be if a data-centre order were one configuration. It rarely is. A single project generates straight lengths, elbows, tees, reducers and offsets, each with its own conductor lengths and joint-end geometry, and every one of those is a setup. Nesting and job preparation software that takes the project as a whole rather than part by part is where the shift hours actually come from; whole-project nesting with an oddment library matters more on a job with fifty part numbers than on a job with five.

Traceability and batch records

Most fabricators moving into data-centre work budget for the machines and forget this.

Hyperscale and colocation clients audit their supply chain. They already hold the design's test certificate, because the busway OEM gave it to them. What they want from you is evidence that the units shipped to their site were made the same way as the unit that was verified. Both the North American and IEC frameworks push towards that from different directions. UL 857 listing carries follow-up factory inspection, which obliges you to keep building the construction that was tested. The IEC 61439 series offers verification by comparison with a verified reference design as an alternative to testing every variant, and that route is only defensible if your process is repeatable enough for the comparison to hold.

In practice that means a batch record that ties together, for each shipped section:

The practical obstacle is that most of this exists somewhere in a busway shop already, in a paper traveller, a spreadsheet and an operator's memory. Turning it into an auditable record usually means the CNC machines have to emit it rather than an inspector transcribing it. Machines with built-in job management and MES hooks make that a configuration exercise; machines without them make it a headcount.

One caution. A batch record is only as good as the measurement behind it. Recording that a part was made on a machine specified at ±0.1 mm/m says nothing about whether that particular part measured within tolerance, and an auditor who has seen a few factories knows the difference.

Where to start if you are quoting this work now

The busway OEM owns the type test and the rating. You own repeatability, and that is the obligation that got harder when rack density moved.

The cheapest place to begin is usually the tolerance audit rather than the capital purchase: measure the joint-plane offset and the fish-plate hole position on fifty consecutive parts from your current process and see what the spread actually is. If it sits inside the joint clearance with margin to spare, the equipment is not your constraint and the plating and records work is where the effort should go. If it does not, you have the number you need to justify the machine.

And if you are sizing conductor section for a new joint rating, the busbar ampacity calculator puts the I²R relationship in front of you before you commit to a plating specification.