Divide $7,400 by $11.95 and the top percentile of Ramp's panel is buying AI at 619 times the median American business, per employee per month [5][16], with another 11.4 times separating that percentile from the top decile beneath it [17]. Annualised, the leaders are paying $88,800 per employee per year [18], which sits at wage scale rather than software scale, and Giovanni Cattani's taxonomy says it should, because in his words "Demand for AI is ultimately demand for labor" [8][7].
That is also why the median line is unlikely to close the gap. Bounded work produces cost-saving returns and, Cattani argues, "converge[s] on the cheapest possible option" [9], so a smarter model does not lift the accountant's seat above $11.95 [5]; it lowers the clearing price. The revenue that pays for the frontier has to come from unbounded, long-horizon tasks, and he counts three of them, AI research, software engineering and trading, at roughly half of all frontier inference revenue [10], with the horizon axis borrowed from METR, where o3 runs about thirty minutes unattended and Mythos about three hours [12]. Each of the three is reflexive by construction: "Larger spend on tokens yields more revenue and a stronger balance sheet, which is spent on more tokens" [11].
Hold that against the $80 billion of compute Anthropic signed for in the last two months [6]. The newsletter maps Ed Zitron's "few hundred companies" onto Ramp's one percent [13][14], and one percent of 70,000 firms is 700 [19], so the commitment works out to about $114 million per firm at 700 and about $267 million at 300 [19]. The bases do not match, since Ramp's percentile spans two labs and the compute is one lab's, but the order of magnitude is the point, or rather the point is what it forecloses: capacity of that size is being underwritten against a customer census you could fit in a mid-sized conference.
Three things would break this read. The median could move, and a $11.95 line that doubles or triples on Ramp's next print would say bounded work is becoming a volume business rather than a rounding error [5]. The mix could sit off-panel, because the figure is business revenue observed on card and bill-pay rails, and nothing in the data says what share of total lab revenue those rails carry, or names a single buyer [1][21]. And Ara Kharazian's own wording is looser than the summary it produced: the top percentile "skew[s] heavily toward the tech sector and AI products and services" [3], which is not the same claim as the labs selling to themselves. What the evidence does carry is his benchmark, that this is concentration "unseen in any other software category we track" [2], against a norm where twenty percent from the top ten customers is unremarkable and forty percent from three is a yellow flag [15]. So the correlation question he asks, ahead of the OpenAI and Anthropic listings [4], is the one to price: the equity on offer carries the beta of its own customer list. The cheapest test is the median. If $11.95 has barely moved next month, the frontier is still being paid for by the frontier [5].
