The Robot Math Problem Nobody Can Agree On

Someone posted a genuinely interesting question online recently. The gist: if minimum wage in parts of the US is still $7.25 an hour, how is any of this robot and AI infrastructure supposed to be cost-effective? How do you justify a billion-dollar data centre to replace people you’re already paying almost nothing?

The thread that followed was one of those rare internet discussions where the argument actually moved somewhere. I’ve been chewing on it since.

The short answer, from people with the patience to do the maths, is that the equation looks very different depending on which job you’re talking about. A robot welding cell at $150k runs nearly 24 hours a day, doesn’t take sick leave, doesn’t get bored, and one technician can oversee several of them at once. Against a $100k welder plus benefits plus safety compliance plus the constant risk they’ll just leave, the numbers start to shift. That’s not a hypothetical. That’s already happening on factory floors. “Dark factories” is the phrase, which sounds dystopian and is.

But then someone else in the thread pointed out that the power consumption estimates people were throwing around were wildly off. One calculation assumed a humanoid robot would draw about 100 watts, which is roughly a light bulb. Someone pushed back: a Boston Dynamics Atlas unit uses around 2 kilowatts every 30 minutes. A decent AI GPU alone draws 1000 watts and costs ten to twelve thousand dollars. If you’re running a general-purpose robot that needs to make decisions in a complex, unpredictable environment, you’re not doing that on a light bulb’s worth of power and a cheap local chip. Not yet.

And that “not yet” is doing a lot of work in this conversation.

The more honest framing seems to be that there are actually several different things being discussed under the same label. Narrow automation, the kind that sorts packages or welds the same seam ten thousand times, has been cost-effective for years and is already displacing people. That’s not new. General-purpose humanoid robots that can do whatever a human can do in any environment, that’s a much harder and more expensive problem, and anyone claiming it’s basically solved is selling something. Then there’s white-collar AI, which is genuinely already cheaper than a junior engineer for a lot of specific tasks, and the companies betting their futures on it getting better fast enough to replace senior engineers too.

I work in IT. I’ve watched AI coding tools get genuinely useful over the past couple of years in ways I didn’t expect. They’re not replacing senior engineers yet. They are absolutely changing what a small team can do, and they’re going to keep changing it. I hold that thought alongside the fact that the data centres powering all of this are enormous infrastructure investments with uncertain return timelines, and the companies building them are making bets that would make a normal CFO’s eye twitch.

One person in the thread put it simply: the ultra-wealthy are fine putting everyone’s eggs in one basket if there’s even a slim chance they come out as gods. That sounds flip but it might be the most accurate description of the current moment. The economics don’t always have to fully pencil out when you have enough capital to ride the wave regardless.

What I keep coming back to is the human cost sitting underneath all of this arithmetic. The welding job that got automated didn’t disappear cleanly into a better job somewhere else. The person who held it had to figure out what came next, in a labour market that wasn’t designed to help them do that. We’re reasonably good at modelling the costs of building a robot. We’re much worse at accounting for what happens to the people on the other side of that transaction.

The printing press comparison came up in the thread, as it always does. It’s a fair point. Transformative technologies always look expensive and unreliable at the start. The question isn’t whether the technology will eventually get cheaper and more capable. It will. The question is what we decide to do with the gap between now and then, and who bears the cost of that transition.

I don’t have a clean answer to that. I’m not sure anyone does.