FuriosaAI brings low-power AI chip to Europe as sovereign compute demand grows

FuriosaAI Launches RNGD AI Accelerators in Europe

FuriosaAI has officially entered the European market by deploying its RNGD AI accelerators at Equinix's LS2 datacenter in Lisbon, Portugal1

1

2

2

. The South Korean AI chip startup, founded by June Paik and Hanjoon Kim in 2017, announced the deployment on Tuesday, marking a significant shift from its previous focus on the domestic South Korean market where it secured wins with companies like LG Electronics1

1

. The European expansion comes as the region pursues sovereign AI compute capabilities and seeks alternatives to American silicon suppliers2

2

.

Low-Power AI Chips Target Energy-Conscious Market

The RNGD accelerators—pronounced "renegade"—distinguish themselves through energy efficiency rather than raw performance. Built using TSMC's 5nm process technology, each PCIe card features 48 GB of HBM3 memory, 1.5 TB/s of memory bandwidth, and delivers 512 teraFLOPS of dense FP8 performance while maintaining a thermal design power of just 180 watts1

1

2

2

. For comparison, Nvidia's RTX Pro 6000 offers twice the memory capacity and compute with comparable memory bandwidth, but consumes 3.33 times the power1

1

. This efficiency gap positions FuriosaAI as a compelling Nvidia alternative for enterprises where energy costs and sustainability matter.

Tensor Contraction Processor Architecture Powers Enterprise Inference

Eight RNGD accelerators combine to form Furiosa's NXT RNGD Server, a 3 kW system featuring up to 384 GB of HBM memory1

1

. The Tensor Contraction Processor architecture enables the system to run large enterprise models including OpenAI's gpt-oss 120B, LG's Exaone 236B, and Qwen 3-30B-A3B at substantial context sizes and concurrency levels1

1

. Because the systems rely on air-cooling rather than liquid cooling, they integrate seamlessly into existing datacenter racks without requiring costly retrofits1

1

2

2

. "We unlock the ability for enterprises to run inference sustainably and reliably," said co-founder and chief executive June Paik2

2

.

Strategic Timing Aligns with European AI Infrastructure Push

The Lisbon deployment builds on FuriosaAI's existing presence in the city, where the company operates a compiler-focused R&D lab and a new flagship office2

2

. The announcement coincides with the RAISE Summit in Paris and arrives as European enterprises actively seek efficient AI compute they can source domestically2

2

. Europe's push to build independent AI infrastructure and reduce reliance on American technology creates opportunities for suppliers offering power-efficient alternatives at a time when energy bills continue climbing2

2

. The move appears focused on building brand recognition and software familiarity in the European market rather than simply offloading excess inventory1

1

.

Broadcom Partnership Signals Next-Generation Ambitions

FuriosaAI is already developing third-generation accelerators in collaboration with Broadcom, adapting its Tensor Contraction Processor technology into a multi-die system-on-package design1

1

. The next-generation chip will utilize faster HBM4 or HBM4e memory and incorporate Broadcom's Ethernet and PCIe switching technology to support larger scale-up clusters beyond the current eight-way systems1

1

. The chip targets frontier models with a trillion or more parameters for hyperscale inference applications2

2

. FuriosaAI joins other companies licensing technologies from Broadcom for next-generation accelerators, including Meta's latest MTIA chips, with OpenAI and Google also disclosing collaborations1

1

. However, reliance on HBM4 memory—which is only reaching the market this year—means these advanced accelerators likely won't appear in production environments soon1

1

. RNGD itself is already in mass production, manufactured using TSMC's process with SK hynix memory, backed by more than $250 million in funding2

2

.