AI Server Manufacturing: Robots and ABB

AI server manufacturing has turned into one of the fastest growing factory jobs in the world. Each rack for a modern GPU cluster holds dozens of accelerators, liquid cooling and a power system that ABB, in its new Infinitus portfolio, expects to reach 1 MW per rack. Contract manufacturers are hiring and building plants in Wisconsin, Texas and Poland, and they are starting to bring robots onto lines that are still mostly hand work. This post looks at how servers were built before, what an AI server line looks like in 2026, where ABB robots could fit, and what has to change before a rack can be built mostly by machines.

An ASUS AI pod built on Nvidia's GB200 NVL72 rack at Computex 2024 in Taipei. Photo: Geekerwan / Wikimedia Commons (CC BY 3.0)

From hand-wired mainframes to corkboard servers

Early computers were built one cabinet at a time. IBM announced its System/360 family in 1964, and its plants turned out machines whose card cages, backplanes and cables were wired and tested by trained technicians. A single mainframe took up a room, and the factory work looked more like building an instrument than a product line. Each customer got a machine configured for its own needs. Very little of that work was automated, because volumes were small and every machine differed. Skill sat in the hands of the people who did the wiring.

An IBM System/360 Model 40 with its gates open, showing the wired card cages. Photo: Myra Siason / Wikimedia Commons (CC BY-SA 2.0)

When internet companies needed thousands of cheap machines, the craft moved into the data center itself. In 1999 Google's early engineers built servers from off-the-shelf PC boards and hard drives laid on trays lined with cork, packing as many as they could into each rack. One of those corkboard racks is now at the Computer History Museum in Mountain View. They were assembled by hand, with cables tied in bundles, and the design assumed that parts would fail and be swapped instead of repaired.

Google's 1999 corkboard servers at the Computer History Museum. Photo: Marcin Wichary / Wikimedia Commons (CC BY 2.0)

By the 2000s, server and PC assembly had moved to large contract manufacturers in Taiwan and mainland China. Surface mount machines placed components on the boards automatically, but final assembly stayed manual: rows of workers in cleanroom suits fitted memory, drives and cables and screwed the chassis shut. Labor was cheap and products changed every year, so a robot cell that needed long programming made little sense for a board the next model would replace.

Workers in cleanroom suits on a hand assembly line at an electronics factory in Shenzhen. Photo: Steve Jurvetson / Wikimedia Commons (CC BY 2.0)

Inside the AI server boom, and where ABB robots fit

AI changed the size of the product. A rack such as Nvidia's GB200 NVL72 joins 72 GPUs with liquid cooling and its own power shelves, and contract makers build and test it as one unit. Supermicro said on September 23, 2026 that it is shipping Nvidia Vera Rubin NVL72 racks. Foxconn plans to add 700 AI server jobs at its Mount Pleasant plant in Wisconsin by mid-2027, and a Foxconn-led consortium is planning a $20 billion hub in Poland, with AI servers made at Miękinia from 2029.

Foxconn's glass sphere building at its Mount Pleasant, Wisconsin campus. Photo: Chad Davis / Wikimedia Commons (CC BY 2.0)

Robots are arriving slowly on these lines. Foxconn planned to put fewer than 10 wheeled humanoid robots on a GB300 server line in Houston from mid-April 2026, doing part transfer, screw fastening and materials handling, and its engineer told Instrumental there is no timeline for harder tasks. The robots carry swappable end effectors and are trained partly in Nvidia's Isaac Sim before they are tuned on the line. Fixed industrial robots already do this kind of work elsewhere. At ABB Smart Power's plant in Vaasa, Finland, a line built by JOT Automation uses ten IRB 1200 robots for scanning and assembly, an IRB 1600 and IRB 2600 for loading and an IRB 910SC SCARA, and builds eight switchgear models without changing tools.

An ABB robot assembling switchgear parts on the JOT Automation line in Vaasa, Finland. Photo: ABB

The same tools could fit server work, though ABB has not announced a server assembly project. An IRB 910SC or IRB 1200 can drive screws and press memory modules with force control, a YuMi or GoFa can plug connectors next to a human builder, and Flexley Mover AMRs can carry trays between stations. ABB robots already work at the other end of a server's life: since 2024 ABB Robotics has worked with start-up Molg on microfactories that take apart retired data center servers. Our post on data center maintenance robots covers the racks once they are installed.

ABB robots taking apart server hardware in a Molg microfactory. Photo: ABB

What a robot-built AI rack still needs

Power is the first thing changing. ABB launched Infinitus in September 2026 as a source-to-rack 800 V direct current portfolio, and it says next-generation AI chips will push racks from today's roughly 200 kW toward 1 MW. Higher voltage DC means fewer conversion steps and less copper for the same power. For assembly, fewer cables help too. Busbars and blind-mate connectors, which mate as a tray slides in, are easier for a robot than routing a loose cable by hand.

A dense Supermicro blade server packed with processors and memory. Photo: Dmitry Nosachev / Wikimedia Commons (CC BY-SA 4.0)

Rack makers are already designing for robots. In September 2025 SoftBank showed a server rack with no cables at the back, using busbars for power, blind-mate liquid cooling connections and optical connectors, built so robots could install and swap servers. If designs like that spread from data centers back into factories, more of the final assembly can be done by standard arms. Engineers usually lay out and test such cells first in RobotStudio, where cycle times and reach can be checked before any hardware is ordered.

Numbered robot cells on ABB Smart Power's automated assembly line in Vaasa. Photo: ABB

The open problems are mostly about change and cost. GPU platforms change every year or two, so a robot cell has to be reprogrammed quickly, which is why high-mix lines like the Vaasa one matter. Testing a finished rack still needs skilled people, and a liquid-cooled rack weighing more than a tonne is moved with lifts and vehicles, not by a robot arm. For now, the most likely picture is mixed lines: robots for screws, memory and trays, people for cables and checks, and AMRs carrying the work between them.

Conveyors and ABB robot cells on the high mix line in Vaasa. Photo: ABB

Innovation

AI solutions for effortless ABB robot control.

Automation

Robotics

ceojohntran@chatwithrobot.net

+84905311611

© 2025. All rights reserved.

qtran1215@gmail.com