Furiosa AI Plans to Double AI Chip Production as 2nm Stork Takes Aim at NVIDIA

Furiosa AI plans to more than double production of its second-generation RNGD inference chip as the company prepares its next-generation 2nm Stork accelerator. The South Korean startup wants to produce between 40,000 and 50,000 RNGD units next year, compared with around 20,000 units expected this year.

ETNews reports that rising demand for AI inference has pushed Furiosa AI to expand production faster than planned. The company expects businesses to require more dedicated inference hardware as agentic AI services and large language models handle larger workloads.

Furiosa AI has already started mass production of RNGD, which focuses on large language model inference. Samsung SDS and LG AI Research are among the customers using or testing the company’s hardware.

Furiosa AI prepares its 2nm Stork chip

Stork will use advanced 2nm compute dies with HBM4 or HBM4E memory, which should provide the bandwidth needed for large AI data centres. Furiosa AI is working with Broadcom on packaging, Ethernet and PCIe technology for rack-scale systems.

The design includes two compute chiplets, multiple memory stacks and dedicated input-output controllers. Furiosa AI says this architecture will improve performance per watt and increase token throughput compared with traditional GPU-based systems.

The company also plans to make deployment easier through its software platform. Its SDK can map PyTorch code directly to the chip, while its Virtual ISA gives developers more control over hardware performance.

Furiosa AI expects to begin sampling Stork during the first half of 2028. The company hopes the chip will help it compete with NVIDIA in the fast-growing AI inference market.