Anthropic leads OpenAI by 2.9 points overall, 7.4 among the heaviest AI spenders

Anthropic's penetration of the model-serving cohort is 93.2%, OpenAI's 85.8% ^[9][8]. That is a 7.4 point spread ^[1] inside the most AI-intensive firms Ramp can see, against 2.9 points across businesses generally ^[3], so the lead is roughly two and a half times wider at the sophisticated end than in the market as a whole ^[2]. For anyone modelling where the enterprise developer relationship settles, that asymmetry carries more information than the headline crossover.

Ramp's proxy is card spend on model-serving platforms, which resell access to hundreds of models, and it says plainly the proxy is imperfect ^[11]. Take it at face value and these users are the opposite of a budget cohort. Annualise June and the median model-serving user is running about $2,976 of AI spend per employee per year, against roughly $127 for the median AI-spending business ^[4]. The cohort's median was also up 21% month over month ^[7]; compounded across a year that is 9.8x ^[5], which nothing sustains, but the sign is the point. Cheaper tokens are arriving as more work attempted rather than smaller invoices, which is Ramp's own reading of it ^[12].

The pricing inference is where the money sits. Ramp argues the absence of broad price cuts from the American labs is itself evidence they have not lost meaningful business share ^[13]. That only holds if the labs would cut in response to share loss rather than to their own margin arithmetic, and it leaves the burden of proof on the cheap-model thesis: five months of movement added 1.3 points of AI-spending businesses, a 29% relative gain off a small base ^[6], and only 3.6% of the firms doing it have stopped buying from OpenAI or Anthropic altogether ^[3].

Ramp's stated outlook is bearish on Chinese models winning broad American adoption, citing adoption that stays low and concentrated among AI-intensive firms plus the incumbents' distribution, product and trust advantages ^[14], while conceding that sensitivity to frontier pricing is real and cheaper alternatives are attractive for specific tasks ^[16]. The labs named in that discourse are DeepSeek, Alibaba and Moonshot AI ^[15]. One limit worth holding in view: because the index is built from spend ^[1], a firm counts as a model-serving user when it has a vendor relationship, not when any measurable share of its tokens has actually moved. Vendor counts move earlier and more easily than workloads do, which means the 5.8% figure is the loosest possible reading of migration, and it is still small ^[5].