Meta is testing robots that can handle physical work inside its data centres, according to several current and former workers familiar with the projects. The experiments reportedly range from a remotely controlled device that presses a power button to more ambitious machines intended to reconnect networking cables and reseat computer components.
The most consequential number in the WIRED report published on August 28 came from one unnamed Meta data-centre worker. That person estimated that, if a cable-swapping robot succeeds, it could replace up to 80 percent of some people’s workloads.
That is a conditional worker estimate, not a Meta forecast. The article did not report a measured 80 percent reduction in labour, an autonomous system operating at scale or a plan to cut 80 percent of technician positions. The distinction matters because a robot that completes one repeated motion can change a job without eliminating every diagnostic, safety and exception-handling task surrounding it.
Several different machines sit behind one headline
The reported programme is not one humanoid robot walking through every server hall. It is a collection of machines at different stages of development, each aimed at a narrow physical problem.
Meta has reportedly evaluated a Kinova Gen3 robotic arm for power cycling or cutting electricity to servers. A separate robot is being tested to swap networking cables. At some facilities, workers said, Meta already uses a much simpler remotely prompted device that resembles a mechanical finger and can press the power button on a Mac Mini or another machine.
Other systems cover mobility and inspection. A self-driving tugger transports heavy server racks, while an in-house wheeled robot scans barcodes to track inventory. Two dual-arm robots made by Watney Robotics have reportedly been tried for cabling work in Altoona, Iowa, since June 2025. At Meta’s Prometheus campus in New Albany, Ohio, a four-wheeled ABB platform with a lift and robotic arm has been used to reseat parts.
Moving a rack, reading a label, pressing a button and reconnecting a cable require different sensing and control. Reliable performance at one task does not establish that the same system can do the others.
Why swapping a cable is harder than it sounds
A network cable looks simple when held in a hand. Inside an operating data centre, it belongs to a controlled sequence. A technician must identify the right device and port, confirm the requested change, remove the correct connector, route the replacement without disturbing its neighbours and verify that service returned.
A robot must perceive and manipulate equipment designed around human bodies and human judgement. It encounters doors, corners, floor obstructions, changing rack layouts and rows of nearly identical ports. Cables overlap and flex. Connectors may use latches, and a misplaced lead can disrupt the wrong system. A low failure rate that appears impressive in a laboratory can still be unacceptable when thousands of interventions are performed on live infrastructure.
Software can restart many devices remotely, but physical intervention remains necessary when remote controls fail, power must be cycled locally or a connection is loose. That final gap between a command and the hardware is precisely where data-centre robotics could become valuable.
The 80 percent figure has narrow boundaries
The worker did not say that a robot had already automated 80 percent of a technician’s day. The estimate applied only if the cable-swapping machine proves successful, and it referred to up to 80 percent of some people’s workloads. It did not cover every role, every facility or Meta’s whole data-centre workforce.
Tasks also do not map directly onto jobs. Even if a machine performs most routine cable changes, people may still diagnose the original fault, approve the intervention, prepare equipment, manage unusual layouts, confirm network recovery, document the repair and step in when the robot cannot finish. Automation can reduce staffing needs, raise the number of machines each technician oversees or change the required skill mix. The report does not establish which outcome Meta expects.
Meta declined to discuss the specific tests. A company spokesperson said it was investing heavily in training and hiring data-centre workers and described a shortage of skilled labour. That response does not settle employees’ worries about future roles, but it means the 80 percent number should not be attributed to the company as an announced target.
Current pilots still rely heavily on people
The workers’ descriptions also show how far the machines remain from general autonomy. The Watney robots are reportedly supervised and cannot yet work as quickly as people. Meta’s inventory robot uses a grayscale camera that cannot distinguish red from green status lights, so some failure checks still need a person. It can struggle around corners and over cables, and a human must open doors and remotely guide it when it moves between buildings.
Battery charging creates downtime. Dense cabling in newer Nvidia GB300 systems reportedly remains beyond the robots’ present abilities. One robotics executive interviewed by WIRED said the industry has produced many pilots and demonstrations but no proven general solution for repairing data-centre equipment.
Those limitations do not make the trials meaningless. Narrow automation often begins with structured tasks and frequent supervision. The tugger and inventory machines reportedly operate in multiple Meta facilities, showing that a limited robot can be useful before it becomes versatile. But a supervised pilot is evidence of development, not proof of dependable autonomy or a workforce reduction.
Meta has a strong incentive to automate the physical layer
Meta’s own engineering account of maintaining large AI clusters describes a constant flow of hardware failures and maintenance work. A repair affects more than the part being replaced: capacity may need to be taken out of service safely, jobs drained from machines, dependencies checked and the hardware returned without destabilising a larger cluster.
The company is already applying software agents to detect and correct performance problems at hyperscale. Physical robots extend the same search for efficiency into the server hall, where software cannot turn a connector or press a button.
Scale strengthens the economic case. AI infrastructure is expanding rapidly, while data-centre sites may be built where skilled labour is limited. Repetitive inspections, heavy rack movements and routine resets are attractive candidates for machines. Meanwhile, the electricity, water and land demands described in earlier ScienceBlog coverage of AI data centres’ physical footprint underline that computation is never purely virtual. The facilities need continuous maintenance as well as power and cooling.
What would demonstrate a successful robotic system
A production system needs evidence beyond a compelling demonstration. Operators would need to know its error rate, average intervention time, charging and maintenance burden, and the share of cases completed without human rescue. Testing would have to cover different rack designs, cable densities, lighting conditions and unexpected obstructions.
Governance matters too. A robot needs reliable authorization before touching live infrastructure, a record of what it changed, an independent way to verify recovery and a safe rollback when the first action fails. Human oversight may move away from the rack, but it does not disappear merely because a mechanical arm performs the final motion.
Meta’s reported experiments are credible evidence that physical automation is moving deeper into data-centre operations. They may eventually alter technicians’ jobs substantially. They are not yet evidence that an autonomous robotic workforce can run a facility, nor that Meta intends to remove four out of every five technicians. For now, 80 percent is one worker’s estimate of what a successful cable-swapping machine might automate in some workloads, not a measured result or a corporate staffing forecast.