Jev reaches 35% of Vercel gateway teams while taking a small share of spend

Jev reached roughly 35% of paid teams in its first two weeks on Vercel AI Gateway, the company says in its October Production Index. Vercel calls that its fastest model adoption on record. The same report puts Jev at 0.27% of gateway spending during that two-week window. Both figures describe this platform, not the entire AI market. Vercel estimates spending at published list prices; actual bills can differ.

Vercel’s table dates the gateway launch to September 17. It separately reports that 55.6% of teams that tried Jev in week one returned in week two. Its full-September figures give Jev about 1% of tokens and 0.15% of spending. Those monthly shares use a different period from the initial two-week spending figure and should not be substituted for it.

TechCrunch’s October 16 coverage describes Jev, from TypeSafe AI, as a decision model for choosing between predetermined outputs. It reports a limited release on September 15, before the gateway date. The article also cautions that gateway usage offers a limited view, particularly for providers that sell through their own APIs. Its coverage summarizes Vercel’s data rather than providing an independent measurement of adoption.

Vercel’s product page describes AI Gateway as a unified interface to models from multiple providers. That context helps explain the boundary of the index: it observes traffic through a particular routing service. It does not claim to capture every direct provider relationship or every deployment outside that service.

Our analysis: the striking result is the distance between broad team reach and a small spending share. That is not a contradiction. Reach answers whether a team used the model; spending share answers how much of estimated spending went to it. Neither number, alone, tells a reader how important the model was to a particular production system.

Consider a purely illustrative gateway with 100 paid teams. If 35 use a model at least once, it can reach 35% of those teams even if their requests are small. A different model could serve only a few teams but receive expensive, long-running requests. That example explains why the metrics can diverge; it is not a reconstruction of Vercel’s customer counts, prices or workload distribution.

Token share adds another question. A request count describes invocations, while token volume describes the amount of text-like processing counted by the service. A compact classification request and a lengthy generation task could make very different contributions to those totals. We would therefore avoid treating popularity by requests, usage by tokens and revenue by spending as interchangeable rankings.

The return figure also needs its stated window. Returning in the second week is evidence of repeated use under that definition. It does not establish months of retention, how often each team returned, or whether the model became that team’s default. A buyer interested in durable adoption would need a longer observation period and a definition of meaningful use before drawing that conclusion.

For someone evaluating Jev, the reported decision-model design suggests a specific comparison: a task with a defined set of acceptable labels or outcomes. We would test the actual categories used by the application, record incorrect choices and ambiguous cases, and compare cost per acceptable result. This is a proposed evaluation method; we have not benchmarked Jev and do not infer that it can replace an unrestricted assistant.

That evaluation should preserve the job being measured. A lower invoice for a different workload would not demonstrate that the same application became cheaper to operate. Similarly, trial use by many teams would not establish accuracy on a new classification problem. The index is useful for identifying a model to investigate, while deployment decisions still need evidence from the intended task.

Vercel’s figures thus support a bounded conclusion: Jev spread quickly among paid teams on its gateway while taking a small share of measured spending. The broader question of global adoption remains outside that dataset. Keeping the platform, period and metric visible makes the report more useful than turning one adoption ranking into a universal market-share claim.

If Jev fits a classification task you already run, compare accuracy and cost per acceptable result on that task, and keep team reach, token share and spending share separate when interpreting Vercel’s index.