FuriosaAI Brings Its Power-Sipping AI Chips to European Datacentres
South Korean AI chip startup FuriosaAI has quietly made its European debut, deploying its RNGD accelerators inside Equinix's LS2 facility in Lisbon, Portugal.
The company, founded in 2017 by June Paik and Hanjoon Kim, has spent most of its existence focused on the South Korean market, picking up customers like LG Electronics along the way. Now it's eyeing Europe, where the buzz around sovereign AI compute gives smaller chip vendors a rare opening that Nvidia's dominance elsewhere tends to slam shut.
The RNGD cards themselves are built on TSMC's 5nm process and sit well below the performance ceiling of Nvidia or AMD's top-tier silicon. Each PCIe card packs 48GB of HBM3 memory, 1.5TB/s of bandwidth, and delivers around 512 teraFLOPS of dense FP8 compute. Not record-breaking numbers, but that's not really the pitch.
The pitch is power consumption. Each card draws just 180 watts. Compare that to Nvidia's RTX Pro 6000, which offers roughly double the memory and compute but pulls over three times the power. For datacentre operators watching their electricity bills climb, that gap matters.
Eight RNGD cards form Furiosa's NXT RNGD Server, a 3kW air-cooled system with up to 384GB of HBM total. That's enough headroom to run reasonably serious enterprise models like OpenAI's gpt-oss 120B, LG's Exaone 236B, or Qwen 3-30B-A3B at decent context lengths and concurrency. The air-cooling angle is worth noting too, since it means these servers can drop into existing rack infrastructure without liquid cooling retrofits.
The European push probably isn't just about moving inventory. Furiosa is already working on a third-generation accelerator with Broadcom, adapting its Tensor Contraction Processor architecture into a multi-die system-on-package design using HBM4 or HBM4e memory. Broadcom's Ethernet and PCIe switching technology will also feature, enabling larger scale-up clusters beyond the current eight-card ceiling.
Furiosa isn't alone in this approach. Meta recently showed off its latest MTIA chips developed with Broadcom's involvement. OpenAI and Google have also announced chip collaborations with the networking giant, though both have been cagey about specifics.
The third-gen parts should be considerably more capable, but HBM4 is only just arriving on the market and HBM4e is further out still. Realistically, those chips won't be landing in production datacentres anytime soon.
For now, Europe gets the renegade. Whether that name turns out to be aspirational or accurate probably depends on how badly operators want an alternative to the usual suspects.