A $110B demand side: the first deduplicated count of what AI buyers actually pay

The deduplication rule is what makes the figure usable, and also what keeps it small. The model counts the dollar an end customer pays and stops there: a dollar spent with Anthropic on Claude, of which fifty cents goes to Amazon to serve it, is reported as one dollar rather than a dollar fifty ^[7]. Both legs are tracked internally, but only the end-customer dollar reaches the headline ^[7]. Anyone who has sized this market by adding up vendor revenues has been counting the same money more than once.

The arithmetic behind the two headline numbers is worth doing yourself. A $175 billion run rate implies a most recent month of roughly $14.6 billion ^[15]. Set that against the trailing twelve months and the run rate is about 1.59 times the full prior year of sales, meaning the latest month annualised sits roughly 59 percent above the trailing-year average month ^[16]. That single ratio is the growth claim. It is also the number to handle carefully, because a run rate is one month with a multiplier attached.

What the count leaves out matters more than what it includes if you are using it as a benchmark. Professional services and systems integration are excluded, on the reasoning that when a Fortune 500 company commits to AI, only part of that spend reaches an AI company ^[8]. Internal AI uplift, such as better recommendation systems lifting ad revenue at Meta or Google, is excluded, as are internal efficiency savings ^[9]. China is modelled but not in this version ^[10]. So a finance line that bundles integrators and change management is not measuring the same thing, and any share-of-market calculation against this denominator is a non-China calculation.

The report sets out to answer how far these revenues go towards covering the investment expense ^[14], and the supply side of that equation is the part already visible: chips, memory, power transformers and cooling sit largely in public companies whose disclosures and forward order books show what is being spent ^[13]. The coverage ratio is the division between the two sides. Now that the denominator exists, the absence of that quotient is the most conspicuous gap in the picture.

Confidence should be set by how the number was built. Much of the revenue sits in private companies including OpenAI, Anthropic, Cursor and ElevenLabs, none of which is obliged to disclose anything ^[11], and the rest flows to hyperscalers that are public but do not consistently break out AI segment revenue ^[12]. The model works from public statements by hyperscalers, neoclouds, suppliers and customers, plus well-reported leaks and self-reports carrying confidence scores ^[2], assembled into item-by-item P&L, balance sheet and cash flow reconstructions that the authors say are auditable back to the contributing data point and its weighting ^[3]. That is a stronger method than a survey and still not a filing.

The growth comparison, roughly three times faster than the mobile or internet waves ^[17], and the reported intent of senior executives across industrials, insurance, finance and pharma to invest more ^[4], both come from the same publisher. The first is model output. The second is Azeem Azhar's conversations, which is directional evidence, not a booked order.