FuriosaAI RNGD AI Inference Accelerator and NXT RNGD Server

FuriosaAI is a Korean AI chip company founded in 2017, with its address in Gangnam-gu, Seoul, and offices in Silicon Valley and Lisbon. It designs data-centre inference accelerators that aim to deliver large language model performance with far lower power than GPUs, so enterprises can run AI in standard air-cooled facilities.

RNGD is built on a Tensor Contraction Processor architecture in TSMC 5nm, offering 512 TFLOPS FP8, 48GB of HBM3 at 1.5 TB/s and 256MB of on-chip SRAM within a 180W TDP on a PCIe Gen5 card. The NXT RNGD Server combines up to eight cards in a 3kW-class appliance for agentic AI. The Furiosa SDK takes PyTorch models through quantization to serving, integrates with Kubernetes and offers an OpenAI-compatible endpoint. Users include LG AI Research for EXAONE models, Upstage and Samsung SDS.

• Tensor Contraction Processor: 5nm architecture delivering 512 TFLOPS FP8 at 180W TDP.
• High-bandwidth memory: 48GB HBM3 at 1.5 TB/s plus 256MB on-chip SRAM.
• NXT RNGD Server: up to eight RNGD cards in a 3kW-class appliance for agentic AI workloads.
• Furiosa SDK: workflow from PyTorch models through quantization to serving and deployment.
• Cloud-native integration: Kubernetes support and OpenAI-compatible API endpoint.
• Validated models: EXAONE models up to 236B parameters on four cards and Qwen3-32B.

A Vietnamese cloud or data-centre operator can offer an energy-efficient LLM inference service on NXT RNGD servers in air-cooled halls. A large enterprise or bank can host Vietnamese-language LLMs on-premises for internal assistants. AI startups can serve open models through an OpenAI-compatible endpoint at lower power cost.

The company reports up to 2x higher token output per kilowatt than an NVIDIA RTX PRO 6000 on Qwen3-32B in FP8, measured on its own setup. LG AI Research reports 2.3x better power efficiency than A100 for EXAONE models. The 180W, air-cooled design lets operators add inference capacity without liquid cooling. Kubernetes integration and an OpenAI-compatible API simplify migration of existing serving pipelines.

Inference now dominates AI compute demand, and power availability in data centres has become a key constraint. Operators are looking for GPU alternatives that cut energy per token and fit existing air-cooled racks. Southeast Asian data-centre growth, including in Vietnam, and sovereign AI initiatives create demand for efficient inference hardware; the company lists a Singapore hub for APAC expansion.

Founded in 2017; based in Gangnam-gu, Seoul, with offices in Silicon Valley and Lisbon and a Singapore hub for APAC. Raised more than USD 250 million through a Series C Bridge round. Partners include TSMC, SK hynix, ASUS, GUC, Broadcom and Equinix.