South Korea: Daum's AI research summary runs on a 100% domestic stack

First "full-stack sovereign" commercialization: the Solar model from Upstage on FuriosaAI's RNGD NPUs, without Nvidia.

In plain terms

AZX, the operator of the Daum portal, would run its AI search summary function ("AI Overview") on a fully domestic stack: the "Solar" large language model from Upstage executed on the "RNGD" (Renegade) NPUs (inference chips) from South Korean FuriosaAI. It is presented as the first commercialization of a "full-stack sovereign" AI stack in Korea - local model and hardware, without Nvidia GPUs.

Our take

This is a concrete case, not a lab announcement. A high-traffic public service - Daum's "AI Overview" - switches a real inference load to local silicon. The interest is not technological patriotism, it's economics: inference, not training, is the cost that bleeds in production when a service scales. A dedicated inference NPU like the RNGD, if it delivers on its promises of cost per watt, hits exactly where it hurts, and reduces dependence on a strained global GPU queue. FuriosaAI has been playing the efficient inference card for years; securing a production deployment alongside Upstage and AZX is the customer validation that most Nvidia challengers lack. The scope remains to be measured: summarizing a search result is a modest load, far from a frontier model served at scale.

To watch

The published cost/performance figures - if any -, the extension to other services, and whether other Korean players follow. Silicon sovereignty is judged by volume, not by press release.