Chip startup FuriosaAI has launched its NXT RNGD server, designed for on-premise and private cloud AI workloads.

The company says the server has been engineered to meet the “rapidly growing demand” for AI inference and addresses the challenge of scaling workloads in the face of high energy consumption, costly infrastructure upgrades to support GPUs’ power, and cooling requirements.

According to Furiosa, a typical NXT RNGD server deployment uses 3kW of power, compared to an Nvidia DGX H100 server, which has a maximum power consumption of 10.2kW.

As a result, a standard 15kW data center rack can accommodate up to five NXT RNGD servers, compared to just one Nvidia DGX H100 server, the company said. Furthermore, this means NXT RNGD servers can be deployed easily into existing infrastructure, with Furiosa noting that currently, more than 80 percent of data centers are air-cooled and operate at 8kW per rack, or less.

Powered by eight RNGD cards, the servers deliver four petaflops of FP8 compute in a single standard 4U unit. The appliance is equipped with a total of 384GB of HBM3 memory operating at 12Tbps memory bandwidth.

The RNGD AI inference chips are built using the company’s Tensor Contraction Processor chip architecture, delivering 512 terraflops of FP8 performance with a thermal design power (TDP) of 150W. Each RNGD card is fabricated using TSMC’s 5nm process, with 48GB of HBM3 memory and a TDP of 180W.

In July, LG AI Research said that when running its EXAONE LLM on RNGD-powered infrastructure, it was able to achieve 2.25x better LLM inference performance per watt compared to an unnamed “GPU-based solution,” with an RNGD-powered rack able to generate 3.75x more tokens compared to a GPU rack operating with the same power constraints.

Founded in 2017 and headquartered in South Korea's capital, Seoul, FuriosaAI has raised approximately $115 million across four funding rounds to support the development of its RNGD chip.

Earlier this year, the startup turned down an $800m acquisition offer from Meta, with the startup reportedly wanting to grow its business as an independent company instead.

The NXT RNGD server is currently in sampling with global customers, but is expected to be available for order in early 2026.