Ramp's index shows US companies using about 50% more AI tokens on a smaller bill

Keeping a token bill flat after a 41% price cut takes about 69% more usage, because 1 divided by 0.59 is roughly 1.69 ^[15]. Ramp counted about 50% ^[3]. If the two figures cover the same months, the bill ends at 0.59 times 1.5, or about 88.5% of where it began ^[16]. Put another way, buyers spent roughly 72% of the saving on extra tokens and kept the other 28% ^[18]. Ramp's data puts the effective price at about $0.68 per million tokens after the fall ^[8]. The article does not give the size of the spending drop or the period behind the 41%, so the 28% depends on the two figures sharing a window ^[3][8].

The populations differ too. Because the token sample leans toward large API buyers ^[6], the record usage and the lower bill may belong to different companies.

Kharazian said the competition is "making AI more accessible" ^[7]. Per token, it has also cut what the sellers collect by 41% ^[8]. Cryptopolitan's write-up says the dynamic could pressure provider margins while lifting demand for inference, cloud and chips ^[19]. Margin is price minus the cost of serving a token, and Ramp's transaction data measures the price ^[6].

The chip and cloud half of that claim is harder to get from these numbers. In the State of Tokenomics survey, 51% of companies said they rely heavily on frontier models, and only 24% expect to a year from now ^[10]. Frank Flight of Citadel Securities argues that adoption depends more and more on affordable compute, power and inference capacity ^[1]. If the volume growth sits in lighter models, a 50% rise in tokens ^[3] is a smaller order for that capacity than the headline figure.

Buyers have reasons to keep deploying. KPMG's Q3 AI Pulse found nearly six in ten leaders reporting measurable AI value ^[13]. "AI's value story is getting sharper," said Todd Lohr, KPMG's vice chair and head of client technology and innovation ^[14].

Two outcomes fit the September data. In one, agents make volume outrun price: the OECD notes agents can consume far more tokens per task ^[11], and Gartner forecasts inference costs per agentic workflow rising more than fivefold by 2028 ^[12]. In the other, routing keeps pulling the effective price down about as fast as usage rises, so token records and smaller bills keep arriving together ^[9]. I think the second fits September better. Buyers put about seven-tenths of the price cut into extra usage ^[18], short of the full offset that Citadel Securities says lower unit prices can unlock ^[2]. That view is wrong if Ramp's spending line climbs back above its July peak while the effective token price keeps falling ^[3].