South Korea just got its first taste of NPU-as-a-Service, and it did not come from the usual suspects. Samsung SDS launched what it’s calling Korea’s inaugural subscription-based NPU cloud offering on July 16, 2026, running on FuriosaAI’s RNGD inference accelerator inside the Samsung Cloud Platform. The target market is telling: government agencies and public-sector clients, not just the hyperscaler crowd chasing benchmark scores.
What the RNGD actually does, and why the numbers matter
FuriosaAI’s RNGD chip, internally called Renegade, is a 180W TDP PCIe card built on TSMC’s 5nm process and paired with SK Hynix HBM3 memory. It entered mass production in January 2026, which means this service launch is running on hardware that is barely six months into commercial life.
The performance figures presented at FuriosaAI’s inaugural RENEGADE 2026 Summit are worth sitting with. The RNGD handles up to 7.4 times more concurrent users than the NVIDIA RTX Pro 6000. On the Qwen3-32B FP8 model workload, the RNGD supports 1.8 to 2 times more users per kilowatt compared to the competition.
Samsung SDS is offering flexible configurations ranging from 1 to 8 accelerator cards, covering training, inference, and serverless AI tasks. The service sits within two existing Samsung platforms, FabriX and Brity Works, which means enterprise customers do not have to rebuild their operational workflows around new tooling.
