FuriosaAI Unveils RNGD Server to Rival Nvidia H100 With Lower Power Consumption

South Korean chip startup FuriosaAI, best known for rejecting Meta’s $800 million acquisition offer, has unveiled its latest product as the race for efficient AI infrastructure intensifies.

The company introduced the RNGD Server, an enterprise-grade AI appliance powered by its proprietary inference chips. Backed by LG AI Research, FuriosaAI is positioning itself as a direct challenger to Nvidia in high-performance AI computing.

The RNGD Server is designed to run large language models without requiring costly data center upgrades or excessive energy use. Each system delivers 4 petaFLOPS of FP8 compute and 384GB of HBM3 memory while consuming just 3kW of power. By comparison, Nvidia’s DGX H100 servers draw more than 10kW, making FuriosaAI’s design far more efficient at the rack level.

FuriosaAI says this efficiency is crucial because most data centers are limited to 8kW per rack or less. A standard 15kW rack can fit five RNGD Servers, whereas only a single DGX H100 could be supported in the same space. The company argues that this provides enterprises with a practical path to scaling AI workloads without incurring massive cooling and infrastructure investments.

The announcement was made in late September 2025, with the RNGD Server now sampling among global customers. FuriosaAI is expected to achieve general availability in early 2026, providing businesses with a timely alternative to Nvidia’s GPU-heavy systems. The company has already demonstrated real-time chatbot performance, running the open-weight gpt-oss 120B model on just two RNGD accelerators.

Financial momentum is also on FuriosaAI’s side. The startup recently raised $125 million in a Series C bridge round and deepened its partnership with LG, which uses RNGD hardware to power its EXAONE models. LG reports that the hardware provides more than double the inference performance per watt compared to GPU-based servers.

The RNGD Server will also benefit from FuriosaAI’s continuously updated SDK. Recent software improvements include inter-chip tensor parallelism, new compiler optimizations, and expanded quantization formats, which further enhance efficiency. These updates ensure that enterprises adopting the system can expect performance to improve over time.

With the RNGD Server, FuriosaAI is betting that efficiency will define the next stage of the AI hardware race. If adoption grows, the startup could become one of the few companies capable of challenging Nvidia’s dominance in data center AI computing.