Do the division before you do the narrative. If about 8,000 job cuts amounted to roughly 10% of the workforce [6], the base is somewhere near 80,000 people [1], and a capital budget of $125bn to $145bn [7] works out to about $1.69m of capex per remaining employee at the midpoint [2]. Nobody commits $135bn to recover the wages of the people who walk the aisles pulling cables. So the payroll story, which is the story the technicians are living, is probably not the story on the capital committee's slide.
What is on that slide is closer to what Meta's robotics lead Eric Xu described at a conference last year: faster incident response, environmental monitoring, preventative maintenance [9]. Those are uptime line items, and uptime in a building full of accelerators is priced per rack-hour, not per shift. A machine that reboots a box at three in the morning without waiting for someone to drive in is worth something even if it is slower than the human it displaces, which the worker who supplied the 80% figure was careful to say it is [5].
And slower matters, because the cheaper substitution is already described in the same reporting: software that lets Meta hire lower-skilled staff and move roles to cost-friendly hubs like Denver, staffed by what one employee called "smart hands" who can take AI instructions without breaking anything [12]. An arm that needs supervision and stumbles on obstacles, battery life, visual inspection and tightly packed cabling [10] competes badly against a lower-wage human on a two-year view; earlier industry trials managed to crush servers on simple tasks [11]. The robot wins on the night shift and on the hazardous job, or rather, on the job nobody wants to relocate for.
Two separate 80% figures are in circulation here and they should not be stacked. One is a Meta employee's estimate of how much of some people's workloads a working cable-swapper would absorb [4]; the other is Unitree founder Wang Xingxing's shorthand for a robot handling about 80% of work in an unfamiliar environment, which he puts two to three years out and as long as five to ten if progress stalls [14]. ACE Robotics chairman Wang Xiaogang wants it by the end of 2027 [13]. That spread is the whole investment question, and it is unresolved by people selling the hardware.
Evidentiary quality is thin and worth saying plainly: the account rests on current and former staff speaking anonymously to Wired, with Kinova and ABB declining to comment, Watney silent, and Meta declining to comment on the testing [1][8]. Note also that the guidance range itself is $20bn wide, about 16% of the low end [4], and that if spending lands at $125bn the prior year was under $62.5bn, an increase of at least $62.5bn [3]. A robotics program disappears inside that forecast error. What would prove this desk wrong is a falling technician count per megawatt at newly commissioned sites while the finger-on-a-button device stays a pilot [3], because that would mean the substitution is happening in hiring plans rather than in hardware.
