FuriosaAI
FuriosaAI provides the RNGD data center inference accelerator, built on a proprietary Tensor Contraction Processor (TCP) architecture that delivers high performance-per-watt for LLM and agentic AI workloads. The 180W, air-cooled PCIe card achieves 512 TFLOPS (FP8) and is designed as a drop-in GPU replacement via an OpenAI-compatible API. RNGD is in mass production and deployed with customers including Samsung SDS and LG AI Research.
- Artificial Intelligence
- Semiconductor
- Hardware
- Enterprise Software
Funding
Raised to date
$808MRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
across 3 rounds
EIHAIB+12$620M · Series D
Announced Oct 2026
Disclosed cumulative: $808M
Founders
Product
Problem
Data center AI inference accelerators often consume excessive power, leading to high operational costs and limiting the widespread deployment of large language models (LLMs) and multimodal applications. Traditional GPU architectures, designed for graphics, struggle with the irregular memory patterns and high-frequency communication demands of modern inference workloads, creating an efficiency gap.
Solution
FuriosaAI provides the RNGD data center accelerator, built on a proprietary Tensor Contraction Processor (TCP) architecture that is optimized for the tensor operations central to AI inference. This clean-sheet design prioritizes high-bandwidth data movement and massive tensor operations, achieving superior performance-per-watt compared to GPUs. The RNGD operates within a 180W power envelope, enabling deployment in standard air-cooled data centers without liquid cooling or facility retrofits. A comprehensive software stack, including the furiosa-llm serving framework with an OpenAI-compatible API, allows for drop-in GPU replacement and seamless migration of existing client code. The accelerator is available as a PCIe card or within the multi-card NXT RNGD Server, and is also offered through cloud services like Samsung SDS's NPUaaS.
Target Audience
Primary customers are enterprises, cloud service providers, and public-sector organizations deploying large language models and agentic AI applications that require high-performance, energy-efficient inference infrastructure.
Features
- Tensor Contraction Processor (TCP) architecture designed for efficient tensor contraction operations and high-bandwidth data movement
- 512 TFLOPS (FP8) compute performance with support for BF16, FP8, INT8, and INT4 data types
- 48GB HBM3 memory with 1.5TB/s memory bandwidth and 256MB SRAM with 384TB/s on-chip bandwidth
- 180W power envelope enabling deployment in standard air-cooled data centers without facility retrofits
- Furiosa SDK including model compressor, serving framework, runtime, compiler, profiler, debugger, and APIs
- OpenAI-compatible API (furiosa-llm) for drop-in GPU replacement and easy migration of existing applications
- Integration with PyTorch 2.x and Hugging Face Hub for streamlined model deployment
- Support for containerization, SR-IOV, Kubernetes, and multiple-instance virtualization capabilities
- PCIe P2P support for LLM workloads and secure boot with model encryption