It’s 2:14 a.m. in a data hall somewhere, and a drive has just failed. The air is a steady 78 degrees and loud enough that you’d raise your voice to be heard. Somewhere down a cold aisle lined with identical racks, a light has turned amber. A human technician would walk the aisle, scan a label, pop the sled, swap the drive, log the ticket, and move on. It takes a few minutes and roughly zero creativity. It’s also the kind of thing that happens thousands of times a year across a fleet of buildings.
That amber light is the job Meta is reportedly trying to hand to a machine. WIRED and TechRepublic both have pieces out on the company testing robot technicians inside its data centers, and the number that’s traveling fastest comes from an International Business Times writeup: machines could take over as much as 80 percent of some workers’ tasks.
I review tools for a living, which mostly means I spend my days watching demos and then finding out what breaks in week three. So let me tell you what I actually think about that 80 percent.
Why data centers are the easy mode of robotics
If you were designing a physical environment purpose-built for a robot to succeed in, you would end up with something very close to a hyperscale data center. Flat floors. Wide, straight aisles. Consistent lighting. No customers. No pets. No toddlers leaving Lego on the ground. Hardware that arrives in standardized form factors, mounted at predictable heights, labeled in machine-readable ways, with a software system that already knows exactly which component failed and where it lives.
Compare that to a warehouse, a hospital, or anyone’s kitchen. Home robotics keeps stalling because homes are chaos generators. A data hall is closer to a factory floor that someone already optimized for automation without meaning to.
So the surprising part isn’t that Meta is trying this. It’s that the industry took this long to point robots at the one indoor space where nearly every hard problem in robotics has already been engineered away.
The 80 percent that isn’t the hard part
Here is where I get skeptical, and it’s the same skepticism I bring to any tool that ships with a percentage in its pitch.
“80 percent of some workers’ tasks” is doing a lot of quiet work in that sentence. Two hedges in one phrase: some workers, and tasks rather than jobs. Task automation and job automation are different animals, and the gap between them is where most enterprise automation projects go to die.
The routine 80 percent of a data center tech’s day is legible, repeatable, and documented. The remaining 20 percent is the cable that was routed wrong three years ago by someone who no longer works there. It’s the rack that’s warmer than the sensor claims. It’s the failure that looks like a drive problem and is actually a backplane problem. It’s noticing that the thing you were sent to fix is not, in fact, the thing that’s broken.
That 20 percent is not 20 percent of the difficulty. In my experience with automation tooling of any kind, the last slice is usually most of the engineering budget, and it’s the slice that determines whether you can actually reduce headcount or just give your existing team a faster helper.
Meta has been here before
The context that makes this story interesting isn’t robotics. It’s that Reuters published a piece on how Mark Zuckerberg’s plan to replace Meta staff with AI imploded. Same company, same underlying bet, recent enough that the bruise should still be tender.
That’s worth holding onto, because it suggests a pattern I see constantly in tool adoption. Leadership sees a capable demo and extrapolates a headcount line. The demo is real. The extrapolation is a guess. When the guess doesn’t land, the tool gets blamed, even though the tool did exactly what it said it would.
Robots swapping drives could be genuinely useful and still not deliver the org chart Meta is imagining. Those are separate questions, and press coverage tends to collapse them into one.
What I’d actually want to know
If someone handed me this system to evaluate, the demo video would be the least interesting artifact in the room. I’d ask:
- What’s the escalation rate — how often does the robot call a human, and is that trending down or flat?
- What’s the mean time to repair compared to a human tech on the same task class?
- How much retrofitting did the buildings need, and does that cost transfer to older facilities?
- What happens on a genuinely bad night, when dozens of failures hit at once and the queue backs up?
- Who maintains the robots, and does that team cost less than the team it replaced?
None of those answers are public yet. What’s public is a set of headlines and a percentage with two hedges in it.
My honest read: this is a sensible place to put robots, probably the most sensible indoor place anyone has picked in years, and the technical work is likely real. I’d just hold the labor conclusions loosely. A machine that swaps 80 percent of the drives is a good tool. It isn’t the same thing as a data center that runs itself, and Meta’s own recent history is the best argument for keeping those two ideas apart.
🕒 Published: