Krosoft | Measuring the AI Economy

Executive Summary

The clearest AI discourse shift today was away from model mystique and toward market accounting. On an otherwise thin signal day, the most consequential item was Exponential View's attempt to put a bottom-up, de-duplicated number on the generative AI economy: $110 billion in sales over the past 12 months and a revenue run rate above $175 billion. The important part is not that these numbers are now settled fact. It is that a serious slice of AI conversation is trying to move from vibes, rankings, and product theater to auditable claims about demand, infrastructure payback, and pricing power.

What Happened

In "The state of the AI economy", Azeem Azhar and collaborators introduced a first pass at measuring AI demand from the customer side rather than inferring everything from chip orders and hyperscaler capex. Their core methodological claim is that they are de-duplicating revenue across the stack: if an end customer pays an application provider, and that provider pays a model company or cloud platform, the same dollar should not be counted multiple times.

That produces the headline figure: $110 billion in AI sales over the last year, with a current annualized pace above $175 billion. Just as notable are the surrounding claims. The writeup argues that current AI-linked hyperscaler revenue roughly covers depreciation on AI infrastructure under its useful-life assumptions, and that price cuts are still expanding the market because usage rises faster than price falls. Their estimate is that every 10% decline in token prices drives roughly 12-18% more token consumption.

This is a v1 market model, not a final answer. But it is a more useful kind of AI discourse than the daily habit of treating every new model, funding round, or benchmark as if it automatically proves durable economic transformation.

Why It Matters

The best way to read this is not "AI has been definitively priced." It is "AI discourse is finally being forced to pick a denominator." For the last two years, much of the public conversation has bounced between two shallow poles: spectacular capability demos on one side and blanket bubble talk on the other. What has been missing is an attempt to say, in plain economic terms, how much real spend exists, where it sits in the stack, and whether that spend is enough to justify the infrastructure buildout.

That makes today's signal important even if the exact numbers change later. A debatable model is still progress if it gives people something concrete to attack, refine, or corroborate. The developing canon around AI has increasingly said that the interesting question is no longer whether people can get striking outputs from models; it is whether usage is becoming routine, budgeted, and sticky. This report reinforces that shift. It treats token prices, depreciation schedules, and demand elasticity as central discourse objects, which is exactly where serious AI discussion was always headed.

The Bigger Story

There is also a subtler narrative here: AI conversation is becoming less centered on frontier symbolism and more centered on economic plumbing. The argument that lower prices can still increase total spending matters because it complicates the lazy assumption that commoditization automatically destroys the category. If falling unit prices unlock much larger usage, then the question becomes who captures value as intelligence gets cheaper to distribute: model labs, application companies, cloud platforms, or entirely new workflow owners.

That is a healthier frame for the next phase of AI discourse. It does not ask whether AI is "real." It asks where the real money is, what assumptions support it, and how much of today's spend is exploratory versus embedded in recurring work.

Workflow Implications

For builders and operators, the practical takeaway is to stop using model quality alone as the shorthand for market durability. When evaluating tools or vendors, add three checks: whether usage is expanding as prices fall, whether the product owns a budget line rather than a curiosity line, and whether the value accrues at the model layer or somewhere closer to the workflow.

On the implementation side, the most useful supporting artifact today was Simon Willison's brief post on building a SQLite version of browser-compat data. It is not the day's main story, but it does illustrate the same broader shift in miniature: AI tools are increasingly interesting not as demo objects, but as accelerants for mundane infrastructure work that makes data more usable and workflows more queryable.