9 min

The century in the warehouse

No, your chatbot prompt doesn't cost a bottle of water. The better question is what all those datacenters are actually buying us. I think the answer is: a lot.

aiagidatacentersenergyopinion
SharePost

Every shortage I care about is really the same shortage: not enough trained minds. We have the raw materials for clean energy, cheap housing, and cures for disease. What we don't have is enough researchers to figure out how to put those materials together. That sounds too simple. But test it against the big problems one at a time and it holds up disturbingly well. It's also why I think the two most common takes on AI right now, that "artificial general intelligence" is marketing fluff and that the datacenters behind it are a menace, both miss the biggest thing happening in the world.

1. The real bottleneck

Think about what scientific progress is actually made of. A new drug takes ten to fifteen years and a few billion dollars to approve. Most of that time isn't spent doing anything dramatic. It's spent being wrong, slowly. Chemists make candidate molecules that fail. Trials run for years and come up empty. And everyone waits for one of a few thousand specialists, somewhere on Earth, to have the right idea in the right decade.

If that sounds abstract, make it concrete. Ask why we don't have a cure for Alzheimer's yet. It's not that the universe forbids one. It's that the number of people qualified to hunt for it is tiny, their careers are short, and most of their attempts fail. The cure exists somewhere in the space of things we haven't tried. We just don't have enough hands to search it.

Materials science moves at the speed of individual careers. One professor, one lab, one lifetime. Fusion, batteries, catalysts, antibiotics. All of it waits in the same line, behind the same scarce resource: expert attention. And that's the one resource we've never been able to make more of on demand. Population growth added minds slowly. Education added them slowly. Nothing has ever added them fast.

That's the frame I use for the phrase "artificial general intelligence." Strip away the branding and it's a claim about that bottleneck. The claim: trained thinking is about to become something you can manufacture. Build it in bulk. Point it at a problem. Scale it up like any other industrial product. Minds, on demand. Maybe the phrase makes you cringe. That doesn't matter. The claim is either true or false, and it's enormous either way.

2. The machines started doing science

For most of my life, that claim was easy to laugh off. The examples were parlor tricks. A computer that plays chess doesn't cure cancer.

That changed. And the sequence is worth spelling out, because each step was, in its day, the official example of what machines would never do.

Go was supposed to be safe for a generation. It fell in 2016. Protein folding was a fifty-year grand challenge; scientists spent entire careers working out single structures. AlphaFold turned it into a lookup table with about 200 million entries, and won the 2024 Nobel Prize in Chemistry. To be clear about what that means: a task that used to consume a PhD now takes a search box. Research math was supposed to be the real test. In July 2025, two separate AI models hit the gold-medal line at the International Math Olympiad, on the same problems as the students, under the same time limit, writing in plain language. And this year, specialized systems have started closing open problems. Questions no human had ever answered.

The quieter stuff below the headlines matters even more. Drug companies are building billion-dollar GPU clusters, Eli Lilly among the first, to do for drug testing what AlphaFold did for protein shapes: predict which molecules bind, which are toxic, which targets are worth chasing. "Co-scientist" systems now come up with hypotheses, design the experiment that would prove them wrong, and iterate on the results.

None of this tells you the year general intelligence arrives. I don't trust anyone who claims to know. But the direction is no longer ambiguous. The interesting question isn't whether machines can do science. It's what happens when they do a lot of it.

3. The compressed century

The best answer I've seen comes from Dario Amodei, who runs Anthropic. I'll say up front that I'm not a big fan of the man, still. But a good argument doesn't care who makes it, and this one holds up. He calls it the compressed century. Take all the discoveries you'd expect biology to produce between now and 2100, on the old schedule of scarce specialists being wrong slowly. Now ask what changes if the researcher is no longer scarce. His guess: the road to cures for most cancers, most infectious disease, and much of the biology of aging could be traveled in five to ten years instead of seventy-five.

You don't have to buy the strong version. Take the weak one. Suppose manufactured thinking merely doubles the pace of medical research. That alone would be the biggest humanitarian event of the century. Measure it in the people who get to live because a cure came in 2045 instead of 2080. That's somebody's grandmother. That might be you.

And biology is just one field. The same math applies everywhere attention was the real constraint. Better batteries. Carbon capture. The long engineering grind between fusion physics and actual fusion power plants. Or education: a tutor who is patient, expert, and free could sit with every child on Earth, not just the children of the rich.

When people ask what AI is actually for, this is the answer that moves me. It's a lever on every problem that was secretly an attention problem all along.

I want to be honest about the uncertainty. The technology may stall. The gains may flow to too few people. The transition will be genuinely hard for a lot of workers, and saying so out loud is part of taking it seriously. But "it might not work" has never been a good reason to stop asking what it would mean if it did.

4. The buildings where it happens

Here's the part I find strangely beautiful. All of this, the compressed century and the manufactured minds, runs in plain warehouses off highways, in counties most people will never visit. A datacenter is racks, pipes, transformers, and a substation. It's the least romantic building of our time, and the most consequential one.

I think future readers will find our reaction to these buildings revealing. Because we've done this before. Victorians were warned that telegraph wires would change the weather. That train travel would scramble the human body. Early electricity was fought as a fire-breathing menace strung over every street, which, to be fair, it occasionally was. New infrastructure always shows up wearing a monster mask. And the mask always fits the era's anxieties better than it fits the building.

Our era's mask is resource panic. And the fit is just as loose. You've probably heard that one chatbot prompt costs a bottle of water. That number is off by a factor of 50 to 250. The honest figure is a few milliliters, somewhere between a few drops and a teaspoon. Even the researcher whose work seeded the bottle statistic has walked it back. Add up every American datacenter, every video stream, every credit card transaction, every AI model in the country, and together they draw about 0.2% of the nation's freshwater. Less than the golf courses of a few counties.

Energy is a bigger deal, but it's the same kind of story. Datacenters used roughly 1.5% of global electricity in 2024. Maybe 3% by 2030. About half a percent of energy emissions today. One prompt costs a few tenths of a watt-hour, a microwave running for one second. I have watched people type prompts by the light of a patio heater while worrying about this. The worry is sincere. But on the arithmetic, it's misplaced by orders of magnitude.

5. What the buildout actually buys

Here's a reading you don't hear much: the datacenter buildout is quietly buying things environmentalists spent decades asking for.

Datacenters are the strongest private demand for clean, around-the-clock power that has ever existed. Their operators need carbon-free electricity at every hour of the day, at a scale no utility would build on its own. So they're the ones signing the contracts that restart nuclear plants, fund next-generation geothermal, and pay for grid-scale batteries. Decades of activism asked for exactly this buyer. It showed up wearing a server rack.

The industry's efficiency record backs this up. Between 2010 and 2018, datacenter computation grew more than fivefold. Its energy use grew about 6%. Why? Because every wasted watt is a cost, and someone is literally employed to delete it. No other large electrical load in the economy comes close to that record.

None of this means there's nothing to fix. There is, and the list is real. Load concentrates: datacenters draw more than a fifth of Ireland's electricity, and the grid hookup queue in Northern Virginia is a genuine mess. Electricity pricing matters, because homeowners should not be subsidizing a tech giant's transmission upgrades. Construction discipline matters too. The ugliest datacenter story of the past year, fouled residential wells in Georgia, turned out to be sediment runoff from the construction site, before the facility ever switched on. A real failure, with a boring, enforceable fix.

Notice what that list actually is: utility regulation, cost allocation, permitting. Ordinary stuff. We know how to do it. It's the price we've paid for every grid-scale technology we've ever adopted. It's worth doing unusually well this time, given what's running inside.

6. Pick the interesting future

There's a version of this moment that will look obvious in hindsight, the way electrification looks obvious now. In that version, the weird warehouse boom of the 2020s was the visible edge of the largest expansion of scientific capacity in human history. The resource fears were the monster mask every new infrastructure wears for a decade. And the compressed century arrived, unevenly and imperfectly, but it arrived.

In the other version, we talked ourselves out of it. Over milliliters.

I'm not neutral between those futures, and I've stopped pretending to be. My optimism isn't faith. It's a reading of the score so far. The milestones keep landing. The costs keep coming in smaller than the folklore says. The benefits are stacked up behind the one bottleneck we've never before been able to widen.

The century is sitting in a warehouse off the highway. I say we go get it.

Sources

Disclaimer: No AI was used to shape or influence the opinions in this piece. The opinions are entirely my own. AI was used only for grammar checks and sentence structure optimization.