The most expensive thing in a hyperscale data center is not the GPU. It is the human walking the aisle to swap a cable, press a power button, or reseat a failing card. Meta is now testing robots to do exactly that work, and one trial could absorb up to 80 percent of some technicians' routine workload.
Data center robotics has spent years stuck in logistics: moving racks, reading barcodes, tracking inventory. The dexterous work stayed human. That line is blurring. Meta is testing robotic systems from Kinova, ABB, and Watney Robotics to perform cable swaps, server resets, and component reseating at facilities in Iowa and Ohio, according to the same WIRED reporting. The push signals that AI infrastructure operators are driving automation from software into the physical layer, where labor costs and worker shortages are becoming structural bottlenecks.
If robots absorb 80 percent of routine cabling work, the remaining 20 percent is where your hiring and costs will concentrate.
What exactly is Meta testing in its data center racks?
Meta is running multiple robot experiments across at least two campuses, each targeting a different slice of physical maintenance work.
At its Altoona, Iowa campus, which opened in 2014 and remains Meta's largest data center site, the company has been testing two dual-armed Watney robots for cabling work since June 2025. The robots are supervised by humans and cannot match human speed yet, but they show promise, according to a person familiar with the effort. Watney, a San Francisco startup that originally built a laundry-folding robot, pivoted to data center robotics in early 2025.
At the newer Prometheus campus in New Albany, Ohio, Meta is testing ABB four-wheel robots equipped with scissor-lift risers and six-axis arms to reseat components. The plan is for these machines to take on additional tasks with diminishing human oversight over time.
A third experiment involves a Kinova Gen3 robotic arm being evaluated for power cycling servers, cutting electricity to machines that need a hard reset. Kinova and ABB both declined to comment on the testing. Watney did not respond to requests.
Meta has also quietly deployed a much simpler device at some facilities: a finger-like mechanism that presses a power button on a Mac Mini or similar device when remotely prompted by a human operator. It is low-tech, but it eliminates a truck roll for a task that happens hundreds of times a day across a hyperscale campus.
The chart below shows the growth in confirmed robot systems at Meta data centers, from two deployed systems in 2023 to five in testing or deployment by 2026. The acceleration began in mid-2025 with the Watney cabling trials and continued into 2026 with the ABB and Kinova experiments.
Beyond the new tests, Meta already operates self-driving tugger robots that transport heavy server racks and wheeled, barcode-reading inventory machines at facilities in Iowa and Virginia. Those systems, first showcased in a 2023 YouTube video, are now also being used for failure inspections. But they have real limitations. The inventory robot's camera only detects grayscale, so it cannot differentiate between green and red indicator lights on equipment. It also struggles with corners and loose cables in its path. Because there are not enough of them, humans must move them between buildings using remote controls. Two sources describe the inventory bot as too slow. Meta previously explored using drones as a faster alternative but did not pursue the idea.
How much technician work can these robots actually take over?
The headline number comes from one Meta data center worker, who estimated that a successful cabling robot could replace up to 80 percent of some people's workloads. That figure is specific to cabling tasks, not all technician work, and it is an internal estimate, not a peer-reviewed benchmark.
Helen Oleynikova, CEO of Exclaim Robotics, a startup developing its own data center repair robots, offers a colder view. "There's been a lot of pilots and demos, but no provable working solution." Her company is building prototypes to replace broken parts and reset cables, but she sees the gap between demo and deployment as wide.
Workers at Meta's data centers report specific technical barriers. Battery-charging downtime is a recurring complaint, taking robots offline for significant stretches. Meta has acknowledged internally that robots are not equipped for the intense cabling work required to run Nvidia GB300 supercomputer systems, the next-generation AI training platforms that demand dense, high-bandwidth interconnects. If robots cannot handle the most complex cabling, they are limited to the routine end of the workload.
"Things have been designed for human hands forever to make everything a five-minute repair," says one former Meta employee. Redesigning server racks and cabling systems for robotic access would take years and require coordination across hardware vendors, facility operators, and robotics companies.
Paul Golding, who oversees physical intelligence work at chipmaker Analog Devices, says the demand is real and growing. "More and more of our customers are telling us 'we want to run their data centers using humanoid robots' because you can run them at higher temperatures, in the dark, and all these things." That vision extends beyond Meta. Microsoft has said it is looking into robots for "economically valuable tasks" like replacing faulty server racks. Amazon has discussed using robots to help recycle parts at its facilities. Google unveiled separate investments in data center robots in 2024.
What does this mean for data center costs and hiring?
Meta's public position is that it needs more workers, not fewer. Company spokesperson Francis Brennan said in a statement that Meta is investing heavily in training and hiring, citing "a major shortage of skilled workers." This year, Meta launched a program to train thousands of people annually in electrical, mechanical, and plumbing work at no cost, with employment guarantees in Louisiana, Ohio, Indiana, and Texas for those who complete training. The company also partnered with building trade unions to fund apprenticeship programs.
But internally, the strategy appears more layered. Workers say Meta is developing AI software and other technologies that could allow it to hire lower-skilled, lower-paid workers and centralize some roles in financially advantageous areas like Denver, Colorado. "They no longer want people who can think independently, come up with creative solutions, or perform complex troubleshooting," one Meta data center worker says. "They want 'smart hands,' people who are just capable enough to follow AI instructions and not mess anything up."
For operators building or expanding data centers, the implications are concrete:
- Labor cost structure shifts. If robots handle 80 percent of cabling and reset tasks, the remaining 20 percent becomes the expensive part: complex troubleshooting, GB300-scale interconnects, and edge-case failures. Your hiring profile moves from "technician who can do everything" to "operator who supervises robots and handles exceptions."
- Facility design changes. Aisles, cable routing, and rack layouts designed for human hands may need rethinking. Robots that struggle with corners and loose cables today will work better in facilities built with their constraints in mind. New builds should factor in robotic access from the start.
- Tax incentive dynamics shift. Communities offer data centers property tax breaks in exchange for jobs. Meta's Altoona campus benefits from a tax break saving tens of millions of dollars annually, justified partly by employment. If robots reduce headcount, the political bargain weakens. Altoona Mayor Dean O'Connor says he has not given much thought to robots taking jobs, framing automation as inevitable: "We will deal with it as it comes."
- Training programs become a hedge. Meta's training program guarantees employment in select states, but the guarantee is for completion of training, not for long-term retention of specific roles. If automation accelerates, the program doubles as a pipeline for the "smart hands" tier.
This connects to a broader pattern: as AI infrastructure spending scales toward historic levels of capital deployment, the physical operations layer becomes the bottleneck. GPUs get faster, but someone still has to swap the cable.
Should you be planning for robotic data center operations?
The honest answer is: not yet, but soon enough that you should be watching the trajectory closely.
The robots in Meta's facilities today are supervised, slow, and limited. The inventory bot cannot tell red from green. The Watney cabling robots cannot match human speed. The ABB system is early in its task scope. Battery life is a real constraint. And the most demanding physical work, GB300-scale cabling, is beyond current capabilities.
But the trajectory matters more than the snapshot. Eric Xu, senior manager for robotics at Meta, said at a conference last year that the company's "long-term goals" include putting robots in data centers for incident response, environmental monitoring, and preventative maintenance. "We believe more collaboration and research will be needed, but we have to start, otherwise we don't have a chance to do this," he said.
What to watch:
- Watney's progress at Altoona. If the dual-armed cabling robots move from supervised testing to independent operation within 12 to 18 months, that is a signal the technology is crossing the deployment threshold.
- ABB's task expansion at Prometheus. The Ohio campus is Meta's newest build. If ABB robots take on tasks beyond reseating, it means the facility was designed with robotic access in mind, and new builds will follow that template.
- GB300 cabling solutions. The gap between routine cabling and GB300-scale interconnects is where robots currently fail. Watch for hardware vendors designing racks and cabling systems that are robot-friendly from the start.
- Competitor moves. Microsoft, Google, and Amazon are all investing in data center robotics. If one hyperscaler deploys a working solution at scale, the others will follow fast.
What not to do: do not assume robots will replace technicians wholesale within the next two years. The technology is not there, and the physical infrastructure was not built for it. Do not cut training programs or reduce hiring based on a pilot. The worker shortage Meta cites is real, and even with robots, facilities need humans for complex troubleshooting and exception handling.
Do start factoring robotic access into new facility designs. Do begin tracking which maintenance tasks in your operations are repetitive enough to automate. And do watch the gap between the 80 percent of tasks robots could handle and the 20 percent they cannot, because that 20 percent is where your highest-value human work will concentrate.
The physical layer is the last mile of data center automation
Software automation has already reshaped how data centers are provisioned and scaled. The physical layer has resisted because it was built for human hands and human judgment. Meta's tests are early, limited, and imperfect. But they mark the moment when hyperscalers stopped treating physical maintenance as a labor problem and started treating it as an engineering problem. That shift, more than any single robot, is what will reshape data center operations over the next decade. The question for anyone building infrastructure today is whether your facility will be ready for the machines that are coming, or whether it will be locked into a design that only humans can serve.
Sources
- WIRED - Inside Meta's Push to Put Robots to Work in Data Centers
- Ars Technica - Inside Meta's push to put robots to work in data centers
- Berliner Tageblatt - Meta tests robots for cabling, server resets and maintenance inside data centers
