The Invisible Trillion · The Other Ledger

And the Buildout Itself

Everything on the ledger so far is about chips and computing in general — the long, accumulated good of fifty years of getting better at building and using them. But the rest of this book is about something more specific and more recent: the enormous, expensive, controversial buildout of AI data centers happening right now. It would be easy to read the cost chapters and conclude that that buildout, specifically, is all cost and no benefit. It isn’t, and honesty requires putting its real upside on the ledger too — even though, as the chapters ahead will argue at length, the benefit is almost certainly smaller than the scale of what is being built.

Some of the good is already here. The same AI models that produce the slop the later chapters will complain about also do genuinely useful work. They have measurably sped up software development — and not at the margin. A large enterprise software system that used to take hundreds of engineers most of a decade can now be specified, generated, and tested with a quantity of compute that costs, in raw terms, almost nothing. The expensive parts of building software — knowing what to build, integrating it, verifying it works — remain stubbornly human, but the part that used to be slow and costly, the writing of the code itself, has collapsed in price by a factor that is genuinely hard to overstate. Whatever else the buildout has produced, it has produced that, and it is a real economic good that will compound for decades.

Some of the good is in the assistive applications that the “Access” section above only began to describe. A blind person can now point a phone at a room and hear it described. Two people with no language in common can hold a real conversation through a translation app that, five years ago, would have produced garble. A student anywhere can ask an AI tutor to explain a hard idea at exactly her level, as many times as she needs, without embarrassment or judgment. These are not the headline uses of AI, and they do not generate the revenue that justifies the buildout to its investors, but they are real, and for the specific people who use them they are not small.

And some of the good is in the use case the chapters ahead will hold up as the clearest model of what compute used well actually looks like: self-driving. After decades of effort, autonomous vehicles have reached the point where, in the cities where they operate, they crash less often than human drivers. The compute behind that — heavy training in the data center, fast inference on the car itself — is genuinely saving lives, and will save more as it spreads. It is the clearest existing proof that the infrastructure this book criticizes can, pointed at the right problem, do something unambiguously worth doing.

I want to be careful not to let this section do more than it should, because the rest of the book is going to spend a great deal of effort arguing that the buildout is substantially larger than these genuine goods require. Both things are true at once, and holding them together is the entire point. The buildout is producing real value — in software, in science, in assistance, in safety. It is also being built well ahead of that value, on a bet that the value will grow to fill the capacity. The benefit is real. The question the rest of the book asks is not whether the benefit exists, but whether it is large enough, and well-enough distributed, to justify what is being spent and who is being asked to pay for it.

All 19 chapters of The Invisible Trillion · Library · All courses