The building is unremarkable from the outside: a shed on an industrial estate outside Sheffield with a roller door and no signage worth photographing. Inside is where the interesting part happens, because this is one of a small number of British sites that assemble and test server racks rather than importing them finished. I spent a day on the floor.
What actually gets built here
A rack is not a product so much as an assembled system: cabinets, power distribution, cooling, cabling and the servers themselves, tested together before they ship. The work is part logistics, part electrical engineering and part thermal engineering, and the failure modes are physical rather than software.
On the day I visited, the floor was building 42U racks for two customers: a regional managed service provider and a university research group. Each rack takes between two and four days to build and test, depending on how much custom cabling the customer specified.
The bit nobody thinks about: power
A single rack of current-generation AI servers can draw more than 40kW. That is not a figure most buildings were designed for, and it changes everything upstream of the servers.
- The electrical supply has to be sized for the peak, not the average
- Busbars replace cables, because cables at that current are thick and inflexible
- Rear-door heat exchangers and direct-to-chip liquid cooling appear where air alone cannot cope
- Floor loading becomes a structural question, not a facilities one
"The servers are the easy part," the site manager told me. "The building is the hard part." He was not exaggerating: much of the recent investment here has gone into power distribution and cooling rather than anything with a processor in it.
Testing is the product
Every rack is burn-tested before it leaves. That means running it at load for a sustained period, monitoring inlet and outlet temperatures, checking for hot spots across the front face, and verifying that the redundant power supplies fail over cleanly.
The tests catch the things that would otherwise become a customer outage: a fan pointing the wrong way, a cable run that blocks airflow, a firmware mismatch between nodes. It is unglamorous work and it is the reason the racks ship with a warranty rather than an apology.
The skills problem
Everyone I spoke to raised the same issue. The people who can build and test this equipment are not being trained fast enough, and the shortage is in the practical end of the trade rather than the graduate end.
"We can hire engineers who can design a rack," one of the leads said. "Finding someone who can spot a badly seated connector from across the room takes years." The site now runs its own training programme, which is a quiet admission that the market cannot supply what it needs.
Why build here rather than import
Three reasons came up repeatedly, and only one of them is cost.
- Lead times. Imported racks arrive in weeks. A UK build can turn around in days when a customer's requirement changes.
- Customisation. British customers increasingly want specific configurations for power and cooling, which is awkward to specify across a supply chain.
- Service. When something fails, the person who built it is a two-hour drive away.
The cost argument is weaker than people assume. Labour is more expensive here, and the components are imported either way.
The AI question
The order book changed shape about eighteen months ago. Where racks used to be filled with general-purpose compute, the recent demand is dense GPU capacity, and that has pushed every downstream decision: more power, more cooling, more weight, tighter tolerances.
It has also made the site's work less repeatable. Standard racks can be built from a template. A 40kW liquid-cooled rack cannot, and each one is closer to a bespoke engineering job than a manufacturing line.
If you want to understand why the chip inside matters so much to the building around it, our explainer on neural processing units covers what these accelerators actually do.
What it means for buyers
For anyone specifying infrastructure in the UK, three practical points emerged from the day.
- Ask for the burn-in report, not just the specification. The test data tells you more than the datasheet.
- Size the power and cooling for what you will run in three years, because retrofitting a building is far more expensive than over-specifying a rack.
- Check the warranty terms for on-site response, which is where a domestic builder genuinely differs from an importer.
The bottom line
Britain does not manufacture many of the components inside a server rack, and it is not going to. What it does have is a small industry that assembles, tests and supports them, and that industry is growing because lead times and service matter more than they used to. The building outside Sheffield is not glamorous. It is, however, the reason a rack can be replaced in a day rather than a quarter.
Further reading: Uptime Institute on data centre tier standards and TechUK on the UK data centre sector.




