The Invisible Trillion · How Much Do We Actually Need?

The Ceiling the Models Leave Out

Every argument so far has been about misallocation — compute pointed at tasks that don’t require it. There is a simpler argument underneath all of them, and it sets a hard ceiling the demand projections rarely include: the people the intelligence is for.

Intelligence can be made essentially infinite and essentially free. The capacity to absorb it cannot. A person can metabolize only so much real content in a day — on the order of twenty to thirty thousand words of genuinely new, substantive material, the figure the chapter on attention develops at length — and that number is set by human biology, not by how many GPUs are running. You can double the supply of generated intelligence every year for a decade; you cannot double how much of it a human being takes in. Past a point, additional output isn’t consumed. It is produced and skimmed, or produced and discarded, or produced and never opened at all.

The honest counterargument is that not all of it is for humans. A growing share of AI output feeds other software — agents calling agents, pipelines processing documents, systems talking to systems — and that machinery has no attention limit. True, and it matters. But a large fraction of what the projections count is still human-facing, and that fraction runs straight into the wall. A demand model that assumes consumption scales with capacity is modeling a consumer who does not exist. And what gets produced past the ceiling does not simply vanish. It becomes the subject of the next section.

The Slop Economy

There is a word that has entered the technical and public vocabulary in the past two years for the category of AI-generated content I just described. The word is slop. It is used as a noun. The standard definition, which has stabilized across recent industry coverage, is low-quality content generated at scale by AI systems for the primary purpose of capturing algorithmic distribution and monetization, with minimal regard for whether the content is useful, true, or wanted by anyone.

Slop has economic logic. The marginal cost of producing AI content approaches zero; the marginal revenue from ads or affiliate links on a piece that gets even minimal engagement is small but positive. When cost is near zero and revenue per piece is small but positive, the rational play is volume: produce as much as possible, accept low engagement on each, and aggregate the thin positive margin across an enormous number of pieces. That is the strategy, and it is being run at vast scale by individuals, agencies, and platforms across the global content economy.

The compute consumed by slop production is substantial. Each AI-generated article is a few thousand tokens of inference. Each AI-generated image is a model invocation. Each AI-generated video is a much larger model invocation. Each is, individually, small. Across billions of pieces of content per day, the compute load is enormous. The hyperscale infrastructure being built is, in significant part, supporting this load.

The output of this compute, however, is consumed at miserable engagement rates. A Bynder study found that 62% of consumers are less likely to engage with content they believe is AI-generated. A SmythOS finding showed that 52% reduce engagement with content they identify as AI-generated. The Capgemini and Baringa surveys found that 77% want to know when content is AI-generated, and that a substantial fraction of those informed users then disengage.

The readership metrics for AI-generated SEO articles, where they’ve been studied, show short dwell times, high bounce rates, and minimal social sharing. The content is technically consumed — briefly loaded and scrolled past — but not engaged with in any meaningful sense.

The economic model of slop only works because of an arbitrage between the content’s actual value (near zero) and the algorithmic systems that distribute and monetize it. Search engines, recommendation algorithms, and advertising networks pay out based on technical metrics — clicks, impressions, time-on-page — that do not measure whether the user got anything valuable from the interaction. AI-generated content can score well on these technical metrics through volume and optimization, even when the actual value to the human users is negative.

Google’s March 2026 Core Update was an attempt to penalize this pattern. Google explicitly stated that the update would “reward genuinely helpful content” and would target sites publishing AI content “primarily to manipulate rankings.” The update has had some effect, with several high-profile content farms losing substantial traffic. But the underlying economics still favor slop production. The penalty for being caught is occasional ranking demotion; the reward for not being caught is steady ad revenue. The producers run the cost-benefit calculation and continue producing, with marginal adjustments to evade detection.

The compute footprint of slop production is the most morally indefensible component of the AI infrastructure buildout. It is compute being expended to produce content that, by every available measure of engagement, the human population does not want, consumed by recommendation algorithms that distribute it primarily because the production is cheaper than the alternatives, supporting an advertising economy that monetizes attention without measuring whether attention was actually paid.

When the residents of Boxtown breathe formaldehyde from gas turbines, some of the electricity those turbines produce is being used to train models that will be used to produce slop that nobody will read, monetized through advertising networks that pay out based on metrics that do not measure value. This is the moral situation the chapter title is reaching toward. The harm in the neighborhood chapters is being imposed in service of an output that, in significant part, is itself harmful to the people consuming it, by polluting the information environment, by drowning out genuine human-created content, and by training the algorithmic systems we use to find information to surface slop ahead of more valuable alternatives.

I do not know what fraction of the AI infrastructure buildout is supporting slop production specifically. The accounting is not public, and the providers have no incentive to break it out. A reasonable estimate, based on the engagement data and the visible content economics, is that something like a third of consumer-facing AI inference is supporting content production that, by reasonable measures of human value, is net negative. Another third is supporting content that has modest positive value. The final third is supporting content and applications that have substantial positive value — the genuine productivity gains, the real assistive applications, the educational and informational uses that work.

These fractions are necessarily approximate. The exact numbers depend on how you weight different kinds of value, and reasonable people will disagree. The point is not the precise number. The point is that, on any reasonable accounting, a substantial fraction of the AI compute being built is being used to produce things that the human population does not net value. The buildout is not, in the simple narrative, producing useful capability that justifies its cost. It is producing a mix of useful capability, useless capability, and actively harmful capability, with the harmful component large enough that it deserves to be named.

This is the version of the inflation hypothesis I most want you to take seriously. It is not that the demand for AI is fake. There is real demand for some AI applications, and that demand is growing. The inflation is in the gap between how much compute is being built and how much useful AI that compute will actually produce. The buildout is being justified by demand projections that conflate useful inference with slop inference. The neighborhoods in those earlier chapters are being asked to host data centers on the assumption that the inference those data centers will run is all valuable. A meaningful fraction of it will not be.

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