Summary
Samsung SDS
NPUaaS instance
NXT RNGD Server
Collaboration start to commercial launch
10 mo
Commercial NPUaaS from a Korean CSP
1st
RNGD cards per on-demand instance
1 – 8
Samsung SDS operates Samsung Cloud Platform (SCP), serving major Korean enterprises across industries and the public sector. In 2025, demand for AI compute grew as enterprises accelerated AI adoption and the government advanced its national AI program. Samsung SDS needed to expand its inference capacity with an option that could deliver competitive efficiency and pricing without relying solely on mainstream GPUs.
Furiosa’s RNGD and NXT RNGD Server offered that alternative. On July 20, 2026 Samsung SDS launched Furiosa’s RNGD series on SCP as a NPU-as-a-Service, announcing it as the first commercial NPUaaS offering from a Korean cloud provider.
Challenge
Delivering Sovereign AI Inference Beyond GPUs
Enterprises and public-sector organizations using Samsung Cloud Platform needed greater AI capacity, often within domestic cloud environments that support their sovereignty requirements. GPU supply constraints and pricing made it challenging to deliver AI compute at a price and scale the service sustainably.
Adding an alternative meant more than introducing a new processor. The offering had to meet five stringent requirements: performance aligned with the service levels of Samsung SDS’s initial compute offerings, a lower total cost of ownership (TCO), compatibility with SCP's OpenStack-based infrastructure, a production-ready software stack, and responsive engineering support. It also had to operate within existing air-cooled data centers.
Solution
Standard Cloud Operations with RNGD
The NXT RNGD Server and its software stack were designed for enterprise and cloud data centers, allowing Samsung SDS to integrate Furiosa RNGD into SCP's OpenStack environment without redesigning its virtualization layer.
Through device passthrough, SCP attaches dedicated RNGD cards to virtual machines, giving customers direct access to their allocated cards. Customers can select 1, 2, 4 or 8 cards through the SCP console. The preconfigured Furiosa Runtime & Driver (FRD) 2026.2.0 image simplifies setup, allowing customers to begin running inference when their instance boots.
With eight passively cooled 180W RNGD cards drawing approximately 3 kW at the server level, the NXT RNGD Server can operate in existing air-cooled halls without requiring a liquid-cooling retrofit.
Target Workload
On-Demand, Multi-Tenant, Cloud-Integrated
Economics
Access Compute Without the Hardware Overhead
Customers can access RNGD inference capacity through a subscription, selecting from one to eight cards without purchasing and deploying their own servers. This model shifts the initial investment and infrastructure from the customer to the cloud service, making dedicated NPU capacity available through an established procurement model.
Development
Expanding Capacity and Workloads
Following the launch, Samsung SDS is expanding Furiosa RNGD capacity in stages and developing a wider pipeline of enterprise and public-sector workloads. A standard path for ISV proofs of concept is also intended to help promising applications progress from evaluation to production on SCP.
VISION
A New Compute Option in a Familiar Cloud Model
Samsung SDS described the launch as a way for customers to use high-performance AI more flexibly and more cost-effectively. For Furiosa, it demonstrates that an NPU can be consumed like a GPU instance, within the virtualization, billing, and operations layers customers already know and trust. As a step toward making domestic AI compute a practical, competitive option for organizations building and scaling AI in Korea, this partnership points toward the future of heterogenous compute, where organizations can choose the right compute for their AI workloads.
