Holding AI revenue flat now takes 69% more tokens than it did in March

An effective price per million tokens is a weighted average, so it drops when a lab cuts a list price and it drops again when customers buy a cheaper model at an unchanged one. Khazarian named both causes. On the basket alone, the frontier share of usage fell eight points in six weeks, about 15% below where it started ^[5]. Fortune's account of the data does not quantify the split between discounting and trade-down ^[22].

Volume has to cover the difference. At 68 cents against a March peak of $1.15, a lab needs 1.69 times the tokens to bank the same dollar, so throughput has to grow about 69% to stand still ^[1]. Huang calls that the "two exponentials" driving the price of AI compute, models growing more complex and more people and agents using them ^[8].

There is a reading of Ramp's own numbers that supports him. A firm that cut its AI dollars by nearly 10% while the price it paid per token fell 38% consumed roughly 45% more tokens ^[3]. Those two figures come from different cuts of the data over windows that overlap without matching, so 45% is approximate.

Against Huang's own benchmark the dollars look thin. He has said a $500,000 engineer should burn $250,000 a year in tokens ^[9]. Ramp's most AI-intensive firms spend about $7,200 per employee per month, which annualises to $86,400, or 35% of that target, and Ramp says the figure is tapering ^[10][2]. "Companies are increasingly starting to use Terra and Sonnet," Khazarian said, mid-tier models that are "highly performant and also cheaper" ^[11]. Amazon and Meta killed their internal token leaderboards in May, and Microsoft cancelled Claude Code subscriptions ^[15].

Cheap open weights do not explain it. Only 3.6% of businesses on Ramp's platform use open source or Chinese models ^[12]. Citadel Securities noted in June that a separate measure, Silicon Data's LLM Expenditure Index, had begun to fall, and put it down to a "bifurcation" between frontier AI concentrated among a few tech-heavy firms and the "everyday" AI the rest of the economy runs on ^[14]. Khazarian called the trend a "crack in the AI thesis" and said it is not a disaster or the bubble bursting ^[17]. "You have multiple metrics now starting to move in a negative direction," he said ^[16].

I would not read the falling price as weak demand, since unit deflation in inference is what the buildout was meant to produce. The dollars decide whether the debt gets serviced, and Morgan Stanley has flagged up to $300 billion of bonds financing neocloud buildouts as vulnerable if token prices do not keep up ^[7]. What would break the bearish case is per-employee spend among the top 1% turning up again while the per-token price keeps falling. In August it did the opposite ^[3].