The Zuckerberg Complex

A machine that takes a person’s job has to be taught that job first, and it learns from data — records of how the work was done, and by whom. The race to automate is therefore also a race to own the information the machines learn from. No one is spending on it like Meta.

Mark Zuckerberg has committed to pouring “hundreds of billions of dollars” into building superintelligence¹. The company spent $72 billion on capital projects in 2025² and told investors to expect as much as $145 billion in 2026³, most of it on the data centers and computing that turn information into models. In June 2025 it paid $14.3 billion for a 49 percent stake in Scale AI⁴, a company whose business is labeling the raw data that trains artificial intelligence⁵, and folded Scale’s founder into a new “Superintelligence” lab.

The data itself is harvested. Meta trains its models on the public posts, comments and images of its own billions of users⁶; in Europe, where the law requires it, people were given until May 2025 to opt out⁷, and a privacy group moved to block it. In the United States, which has no such law, there is nothing to opt out of. A federal judge found that Meta had also downloaded books from a pirate library in 2022 to train on⁸, and ruled the use a “transformative” fair use — while noting the authors had simply argued their case badly.

The richest vein of all is the workplace, because work can be watched. Employee-monitoring software has grown into a market worth several billion dollars⁹, and by one industry estimate the share of large employers tracking their workers roughly doubled during the pandemic¹⁰. The tools record screens and keystrokes, clock the seconds between tasks, and read the tone of a voice on a call.

That watching is not incidental to automation; it is the input. The International Labour Organization notes plainly that the data generated by monitoring workers is used to “train” the machine-learning systems that manage — and can eventually replace — them¹¹. Amazon’s warehouses meter “time off task” to the minute¹²; its delivery vans carry four-lens cameras that record the entire route¹³ and flag the driver. Every flagged second is a labeled example of how the job is done, and how it might be done without a person.

Outside the workplace, the same logic has been bolted to the street. A company called Flock Safety runs the country’s largest network of automated cameras¹⁴ — more than 100,000 of them by mid-2026¹⁵, reading license plates across five thousand communities and, by outside estimate, scanning some 20 billion vehicles a month¹⁶. Investors valued it at $7.5 billion in March 2025¹⁷, and higher since.

What Flock built as a crime tool became a data pipeline with loose doors. In 2025, reporting and a state audit found that local police had run the system thousands of times on behalf of federal immigration agents¹⁸, and that officers had searched tens of thousands of cameras nationwide to track a Texas woman who had had an abortion¹⁹. Illinois found Flock had violated state law by letting federal agencies in²⁰; the company said it had not built the controls to stop them, and paused its federal pilots.

The value in all of this pools in the data layer and thins on the way down. Flock is worth around $8 billion; Scale AI, about $29 billion²¹; the surveillance and data-broker firms that feed the pipelines are valued in the tens and hundreds of billions. The people who do the actual labeling — teaching the models what a face or a road or a task is — work through gig platforms in the Global South for a few dollars an hour, sometimes cents a task²².

This answers a question the first chapter left open. The robots being built to cover the aging world’s empty jobs are not conjured; they are trained, on data taken from the workers they will replace and the public they will watch. The same firms that will own the machines already own the information that makes them possible.

And the spending only accelerates. Having bought half the largest data-labeling company and committed to the biggest build-out of computing power in its history, Meta raised its forecast again in April 2026, to as much as $145 billion for the year²³ — a sum larger than most governments’ entire budgets, aimed at teaching machines to do what people now do. Someone taught them. The next chapter follows the machines into the one workplace the aging world can least afford to automate: the bedside.