Furiosa

Market position

FuriosaAI positions RNGD as an inference alternative to GPU-based systems, citing a benchmark in which RNGD outperforms the RTX Pro 6000 with its latest SDK, and emphasizing total cost of ownership, per-rack throughput and power efficiency. Through Samsung SDS it is associated with what is described as Korea's first domestic NPU-as-a-service offering.

Its principal technical differentiator is the Tensor Contraction Processor architecture, which departs from the fixed-size matmul primitives used in most commercial accelerators, combined with a low-power, air-cooled appliance design (3 kW per 8-card server) intended to fit standard 15 kW enterprise racks and a compiler-centric software stack for model flexibility.

Technology

RNGD accelerators implement the Tensor Contraction Processor architecture, described in an ISCA 2024 paper, which treats tensor contraction — a higher-dimensional generalization of matrix multiplication — as the hardware primitive instead of the fixed-size matmul instructions used by most commercial deep learning accelerators. The company frames the resulting benefits as high-throughput, low-latency execution, lower energy draw permitting standard air cooling and fewer racks, and compiler-level programmability for evolving models. The software stack includes the Furiosa SDK (release 2026.3 introduced a new kernel framework), PyTorch 2.x integration, and cloud-native deployment components such as containerization, SR-IOV and Kubernetes.

Go-to-market

The company runs the Furiosa Access Program, a structured online and offline evaluation path through which customers and partners can evaluate, integrate, qualify and deploy Furiosa accelerators, with access locations in Seoul (Korea), the Bay Area (USA), Lisbon (Portugal) and Johor Bahru (Malaysia). It also goes to market through partnerships with Samsung SDS (domestic NPU-as-a-service in Korea), Broadcom, Equinix, and I/ONX and Velox for data center capacity, and publishes benchmarks, white papers and conference research (ICML 2026, ISCA 2024) to support technical evaluation.

Enterprises deploying LLM and agentic AI inference in on-premises, managed or colocation data centers, including customers constrained by rack power density, plus cloud and infrastructure providers offering NPU-based services.

Geography

Headquartered in Seoul, South Korea (Gangnam-gu, Dosan-daero 145). Furiosa Access evaluation locations are listed in Seoul, the Bay Area, Lisbon and Johor Bahru, and the program is described as available worldwide. European deployments include RNGD at an Equinix data center in Lisbon and a planned 15 MW AI data center in Sweden with I/ONX and Velox.

Risks & controversies

The company posted a notice warning that individuals impersonating its recruiting team have solicited personal information — identification documents and bank account copies — from candidates under the pretext of setting up equipment for a data-crawling freelance role, and stated it does not request such information during hiring.