The Invisible Trillion · How Much Do We Actually Need?

Where This Leaves Us, And What Comes Next

I have laid out, in this chapter, the empirical case that the AI buildout is being constructed substantially ahead of any durable demand signal that justifies it. The utilization data points to capital deployment running ahead of usage. The industry-by-industry tour shows that the categories most often invoked to justify hyperscale construction — robotics, science, healthcare, education, finance, defense — are either smaller than the rhetoric implies, served better by other architectures, or growing from bases small enough that they cannot absorb the capacity being built. The slop economy reveals that a meaningful share of the inference traffic currently flowing through hyperscale infrastructure is producing output that, by every available measure of engagement, people do not actually want.

Beneath all of it sit two limits the demand models tend to skip: a human attention span that caps how much output can ever be consumed, and a material ledger — every machine mined into being and discarded out of it — that the most expansive justifications never total. The inflation hypothesis follows directly from these observations: a substantial fraction of what is being built will not, on any realistic timeline, be filled by useful demand.

If the chapter ended here, it would leave the reader with the diagnosis but not the prescription, and with the diagnosis but not the engagement with the strongest counter-arguments. Both deserve their own treatment, and both will get it.

The next chapter — The Architecture That Works — addresses the practical question that follows naturally from this one. If the buildout is substantially overbuilt, what would the right buildout look like? That question has a surprisingly concrete answer, and the answer is being demonstrated in one specific application area where physical constraints forced the industry to get it right. The next chapter walks through what honest sizing would look like, resolves the apparent contradiction between idle enterprise capacity and user-experience scarcity, examines the specific case of xAI’s Colossus facility as an illustration of how the inflation works in practice, names the structure of the financial system around the buildout as an anticipation economy, and uses self-driving as the case study for what compute used well can actually accomplish.

The chapter after that — The Bill — engages with the two arguments most often deployed by the buildout’s defenders. The financial markets are voting on the buildout, and the votes are coming in at trillion-dollar valuations. The United States and its allies face a real geopolitical contest with authoritarian competitors, and the Ukraine war has demonstrated that AI on the battlefield has decisive consequences. Both arguments are serious. Both deserve direct engagement rather than dismissal. The third chapter takes them on, examines what they imply and what they don’t, and ends on the question that runs underneath the entire arc: at what cost, and to whom?

For now, hold the diagnosis. The buildout is largely ahead of demand. The demand that does exist could be served by architecture that the next chapter will describe. The costs being borne now are being borne by people who have not been adequately represented in the decisions. And the question of how this resolves — whether the bet pays off for the bettors, whether the buildout finds its eventual demand, whether the host communities recover — is one that the next two chapters can illuminate but not resolve.

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