Three deals in weeks pull the open-weight distribution layer inside vendor stacks

The person who has to answer for this on Friday keeps a hub URL in a deploy script and a router key in a secrets manager, and neither line changes this week even though the ownership behind both does.

What the money buys is distribution rather than technology. Nvidia already ships its own family of open-weight models, Nemotron, and TechCrunch reports that uptake has not been large ^[6]. Buying the largest US developer space for open models gives the company a mass of users it can point toward its chips and standards ^[5]. The pressure behind that, per TechCrunch, runs the other way: Nvidia wants less dependence on its deals with hyperscalers and frontier labs at a moment when OpenAI and Google are building inference chips of their own, including OpenAI's Jalapeño, whose capabilities were announced the same week ^[4].

Add the reported prices and the three transactions come to at least $26bn, of which Nvidia's two account for $19bn ^[17][18]. The layer being bought is used by 6% of companies in Ramp's survey of spending data, and by 2% of software engineers in Jellyfish's ^[9][10].

Teams tell themselves open weights are the hedge against frontier pricing. What the people running them are actually doing, according to Jellyfish AI product lead Nik Albarran, is reaching for control and configurability, not savings, and mostly on repeated inference workloads such as customer service chat, where one tuned model answers the same question cheaply at volume ^[13][11]. For coding and agentic work, frontier models still win, helped by easier access and in some cases a token subsidy ^[12]. Albarran's own condition for that changing is frontier prices continuing to rise, and workflows maturing enough to justify self-hosting ^[13].

Stripe put the routing purchase in token terms, with Patrick Collison saying tokens are the central currency for companies building with AI and that real-world economic potential depends on making good use of scarce compute ^[8]. The obvious independent substitute has the same profile as the thing that just sold: Fireworks, whose CEO Lin Qiao says it processes 40 trillion tokens a day, more than either Gemini's or OpenAI's APIs, and which is regularly discussed as an acquisition target itself ^[14][15].

A way to sort this by Monday. First axis: does the dependency run at build time, where a bad week breaks CI, or at request time, where it breaks the product. Second axis: one path or two, meaning whether you hold your own copy of the weights and a funded second route for the same traffic. Request time with one path is the box to fix first, and the fix costs real money: mirrored artifacts, version drift to manage, a second bill for capacity you hope not to use.

Download counts and hub popularity do not show whether you could leave. What shows that is time-to-swap, measured as the hours between a terms change and the same requests being served another way. For the companies outside Ramp's 6%, the cheap version of this work is knowing which registry their vendors pull from, and who owns it now ^[9][1].