The RNGD Chips Are Live in Lisbon
FuriosaAI just went live at Equinix's LS2 datacenter in Lisbon, Portugal. Their RNGD — pronounced "renegade" by the company's own engineers, apparently as a joke that stuck — accelerators are now running production inference workloads on European soil. This is the startup's first deployment outside South Korea, and honestly? It's about time.
I've been tracking Furiosa since they first showed up on my radar a couple years back, and their trajectory has been quietly impressive. Founded in 2017 by June Paik and Hanjoon Kim — before LLMs were even a blip on anyone's radar, back when people still thought AI chips were a niche play — the company spent years building something most observers didn't notice: a tensor contraction processor (TCP) architecture that actually makes sense for inference workloads. Not training. Inference. The part of the AI stack where most datacenter budgets actually bleed.
Now they're in Lisbon. And if you've ever wrestled with power costs at scale, this matters more than the headline suggests.
What the RNGD Chip Actually Does
Let's cut through the spec sheet noise. The RNGD is built on TSMC's 5nm process, ships with 48GB of HBM3 per PCIe card, and delivers 1.5TB/s of memory bandwidth. That's solid. But the number that actually makes me sit up is 180 watts.
One hundred and eighty. Watts. Per card.
For context, Nvidia's RTX Pro 6000 — their closest competitor in this space — offers roughly twice the memory and compute while consuming 3.33x the power. That's not a marginal difference. That's a fundamental architectural choice that says "we optimized for inference, not training." And in a market where electricity bills are the second-largest cost after hardware itself, that's the kind of thing that makes procurement teams pay attention.
The RNGD delivers 512 teraFLOPS of dense FP8 performance. That's not training-grade throughput, and Furiosa isn't pretending it is. But for inference — serving large language models at scale, handling concurrent requests with tight latency budgets — it's genuinely competitive. I've seen clusters like this in production, and the power savings alone can justify the deployment even before you factor in performance.
The NXT Server: Eight Cards, One Rack Unit
Furiosa isn't just selling chips. They're offering the NXT RNGD Server — a complete system that packs eight RNGD accelerators into a single 3kW air-cooled chassis. That's 384GB of HBM memory in one box, enough to run some genuinely large models at production scale.
OpenAI's gpt-oss 120B? It runs. LG's Exaone 236B? Also fits. Qwen 3-30B-A3B at large context sizes with real concurrency? Yes. And here's the part that matters for anyone who's actually deployed AI infrastructure: it's air-cooled. No exotic liquid cooling. No custom datacenter modifications. It plugs into existing racks and runs.
I've seen too many AI chip startups promise the world and deliver a system that requires a complete infrastructure overhaul. Furiosa's approach is refreshingly pragmatic. They're not asking you to rebuild your datacenter. They're asking you to plug in and run.
Why Europe? Why Now?
Here's where it gets interesting. Furiosa could've stayed in South Korea. They've got LG Electronics as a customer, they've got domestic demand, and they've got a working product. But Europe represents something different — a market waking up to the strategic importance of sovereign AI compute.
Data privacy regulations are tightening. Geopolitical tensions are rising. Enterprises and governments across the continent are asking: what happens if we can't access American cloud providers? What happens when export controls tighten further?
The Lisbon deployment at Equinix's LS2 facility puts Furiosa right in the middle of this conversation. It's not about offloading excess inventory — though I'll admit, that's always a factor for any hardware company expanding internationally. It's about brand recognition. Software familiarity. Getting European operators to trust a non-American chip before they need it.
That's a long game. But Furiosa seems willing to play it.
The Broadcom Partnership: What Comes Next
Furiosa isn't resting on the RNGD. They're working with Broadcom on a third-generation AI accelerator that combines their tensor contraction processor technology with Broadcom's networking expertise. This isn't a side project — it's the company's next big bet.
The new chip will use a multi-die system-on-package design with HBM4 or HBM4e memory. That's the cutting edge of high-bandwidth memory, and it's only just hitting the market now. Expect significantly higher performance and scalability compared to the current eight-way systems.
Broadcom's Ethernet and PCIe switching tech will also support larger scale-up clusters than what Furiosa is building today. Think beyond single servers to multi-rack deployments without the networking headaches that usually come with that.
This is the kind of partnership that could give Furiosa real legs. Broadcom brings distribution, networking know-how, and credibility in enterprise silicon. Furiosa brings a proven inference architecture that's already running in production. Together, they're building something neither could do alone.
But here's the reality check: HBM4 and HBM4e are still early. We probably won't see these chips in the wild anytime soon. The RNGD deployment in Lisbon is about now. The Broadcom partnership is about later.
The Bigger Picture: Can Furiosa Actually Win?
Let's be honest. The AI chip market is brutal. Nvidia dominates. AMD is pushing hard. And there's a whole ecosystem of startups — Tensordyne, Tenstorrent, Cerebras — all claiming to have the next breakthrough.
Furiosa's advantage is narrow but real: power efficiency. They're not trying to beat Nvidia at training. They're not trying to build the fastest general-purpose AI accelerator. They're building the most power-efficient inference chip in its class, and that's a defensible position.
European datacenters are hungry for options. Not because they don't like Nvidia — though some of them have opinions about that — but because relying on a single supplier for critical infrastructure is risky. Furiosa offers an alternative that's already proven in production, already running at scale, and already saving power.
Will they become a household name? Probably not. Will they carve out a meaningful slice of the European inference market? I think so. Especially if the Broadcom partnership delivers on its promise.
The Lisbon deployment is just the beginning. But it's a solid one.
For more insights on AI chip competitors, check out our articles on Etched’s $1 Billion Inference Bet and DeepSeek’s plans to develop its own AI chips.