The ratio is the number to hold on to. If $260m is 60 per cent of what the same customers spent querying OpenAI, Anthropic, Google, xAI and Perplexity directly [3], the direct bill came to roughly $433m [1], the two lines together to about $693m, and the hosting-and-serving layer accounts for 37.5 per cent of everything that panel spent on running or calling models in the quarter [2]. That is a real business. It is also a much smaller thing than the phrase "AI infrastructure" usually carries, and Ramp itself describes the ecosystem as a niche player not yet putting pressure on closed-source vendors [2].
The narrowness is sharper on the buyer side. About 1,900 Ramp customers spend anything at all on this layer [4], which averages roughly $137,000 a quarter each, or about $547,000 annualised [4], spread very unevenly, presumably, since a market that runs from a serverless router like OpenRouter at one end to customer-operated bare-metal clusters at the other has no meaningful average account [7]. Ninety-four per cent of those buyers also pay a closed-source provider [5], which leaves something like 114 companies in the panel that host and serve without renting a frontier model at all [3]. That is the entire observable population of pure self-hosters.
And the two lines move together: more closed-source spend comes with more infrastructure spend, which Ramp reads as complementarity rather than substitution [6]. The mechanism is unglamorous. A team that has decided to put a model into production has to host it somewhere and make it callable by other services [10], and once that plumbing exists it reaches whatever the vendor's catalogue reaches, closed models included [9]. The price pressure that a credible open-weight alternative is supposed to exert on closed-source list prices [9] is not visible in this quarter's dollars [2].
Two readings I cannot test from the material. Corporate spend data catches pay-as-you-go and misses whatever sits inside committed multi-year compute contracts, so the $260m could be a floor and the 60 per cent understated; and a panel of companies that use a spend platform is not the same population as an insurer running models on hardware it already owns. Neither possibility is in the excerpt, and I would want the vendor-level breakdown before believing either.
This is probably wrong, but the view I would take is that NVIDIA, in licensing Groq's LPU designs to get into model hosting and serving [1], is buying an option on the customer count rather than on the current accounts. The falsifier is clean enough. If the next quarter shows roughly the same 1,900 buyers each spending materially more, then this is a whale market whose revenue concentration cuts both ways, and the interesting risk is not that a buyer cannot find a second provider but that a provider cannot survive losing four accounts. If the count moves instead, from 1,900 toward five figures at flat per-account spend, the narrowness was a timing artefact and the layer is doing what plumbing does. On present evidence, 1,900 buyers and 114 pure self-hosters [4][3] is a market of specialists, and specialists do not set prices for anyone else.
