Furiosa AI Stork: 2nm Chip With HBM4 Memory Targets Lowest Cost Per Token

Furiosa AI is reshaping the AI hardware market by committing to a massive production expansion and revealing details about its next-generation Stork chip. Buyers and data center operators should care because this move signals a serious challenge to NVIDIA's dominance in the inference market. The company plans to double its current output, which means more hardware will be available to enterprises seeking alternatives to traditional GPUs.

Furiosa AI ramps production to 50,000 units for its next-gen accelerator

The Stork chip represents Furiosa's third-generation effort, built on Broadcom's advanced 2nm process node. It pairs this compute technology with HBM4 and HBM4E memory standards to maximize bandwidth efficiency. The architecture relies on a Tensor Contraction Processor design that aims to deliver higher performance-per-watt than even the most efficient existing GPUs.

Stork Specifications

• Process Node: 2nm
• Memory Standard: HBM4/4E
• Memory Capacity: 432 GB
• Architecture: Tensor Contraction Processor
• Production Volume (Next Year): 40,000 – 50,000 units

Production plans show a clear ramp-up trajectory for the company's hardware. Furiosa intends to increase output from 20,000 units this year to between 40,000 and 50,000 units by next year. The company is already shipping its second-generation Renegade chips to major customers like Samsung SDS and LG AI Research.

Stork is expected to begin sampling in the first half of 2028, giving the industry time to prepare for its capabilities. The chip targets the lowest cost per token, positioning itself as a direct competitor to NVIDIA in efficiency metrics. June Paik, Furiosa's co-founder and CEO, stated that combining Broadcom's infrastructure with their software stack creates a comprehensive solution for the token factory era.

This expansion confirms that Furiosa is moving beyond prototype stages into significant manufacturing scale. The combination of 2nm fabrication, HBM4 memory, and high-volume production plans establishes Stork as a viable alternative for large-scale AI inference workloads.