The Invisible Trillion · How Much Do We Actually Need?
The Inflation Hypothesis, Stated Plainly
Let me state the inflation hypothesis in its strongest form, because it is the hypothesis I most want this chapter to put on the table.
The hypothesis is this. A substantial fraction of the current AI data center buildout is being constructed on the basis of demand projections that assume:
All AI inference will continue to route through frontier-class models, when in fact the trajectory of model capability improvements at the small-model end means that most inference will migrate to edge devices over the next three to five years. Each unit of inference that migrates is a unit of demand that hyperscale infrastructure no longer needs to serve.
Consumer engagement with AI-generated content will continue to support the slop economy, when in fact engagement metrics for AI-generated content are already declining, regulatory pressure is increasing, and the underlying advertising economy is shifting away from impression-based monetization toward engagement-based monetization that slop will not survive. The compute being built to support slop production will increasingly produce output that has no economic outlet.
Enterprise AI features will continue to be priced on usage-based billing that supports current infrastructure ROI, when in fact the unit economics of frontier API calls in production enterprise software are already breaking down, and enterprises are migrating to cheaper alternatives (smaller models, local inference, batch processing) that consume dramatically less hyperscale capacity per dollar of value delivered.
Training demand will continue scaling with model capability improvements, when in fact the major labs are increasingly relying on architectural improvements, data quality improvements, and reinforcement learning techniques that do not require ever-larger training runs. The marginal training compute required to deliver the next generation of frontier capability may be smaller than the previous generation, not larger.
Specialized applications (drug discovery, scientific computing, robotics) will scale fast enough to absorb infrastructure overhang from consumer/enterprise demand softening, when in fact the specialized applications are growing from a small base and will not, on any realistic timeline, scale to fill the capacity being built.
If these assumptions are wrong — and the available evidence suggests that several of them are at least partially wrong — then the buildout is, in a literal financial sense, overbuilding. The capacity will exist; the demand will not arrive to fill it; the assets will be written down; some of the operators will fail; the residents of the affected neighborhoods will be left hosting infrastructure that was built for a demand pattern that never materialized.
This is the hypothesis. I want to be clear about how confident I am in it, because the chapter would be dishonest if it overstated.
My confidence level: I think the inflation hypothesis is probably mostly right, but with substantial uncertainty. The available evidence — the utilization numbers, the engagement metrics, the strategic shifts among major labs, the migration toward edge inference, the breakdown of frontier API unit economics in enterprise applications — all points in the direction of the buildout being substantially ahead of sustainable demand. The counter-evidence — the genuine demand from scientific applications, the real productivity gains from some enterprise tools, the continuing growth in consumer engagement with AI products — is real but appears insufficient to fill the capacity being built.
The honest summary is something like this: the buildout will probably turn out to have been roughly 50–70% justified by real durable demand, and 30–50% built on speculative assumptions that will not be validated. The 30–50% overhang will be absorbed in some combination of asset write-downs, capacity repurposing, and unit-economic deterioration. The residents of the neighborhoods hosting the overbuilt portion of the buildout will not be compensated for their share of the cost.
These percentages are imprecise. They are my best honest estimate, based on the available data. Industry analysts who are betting their careers on this question disagree with each other by margins much wider than the precision of any individual estimate. The point is not that the inflation is precisely 40% or precisely 30%. The point is that meaningful inflation almost certainly exists, and that the standard industry narrative — that the buildout is fully justified by demand — does not survive contact with the utilization data.
All 19 chapters of The Invisible Trillion · Library · All courses