FuriosaAI has officially switched on its RNGD AI accelerators in a European datacenter, marking the company’s FuriosaAI RNGD Europe launch. The news arrived at the RAISE Summit in Paris, where the chips—pronounced “renegade”—were revealed as power-efficient challengers to Nvidia‘s high-power lineup.
For you, this means a fresh alternative in the AI accelerator market. Instead of relying solely on Nvidia’s hardware, you now have an option that emphasizes efficiency without compromising on capability. As AI workloads expand, having a choice like this can help you balance performance and energy costs.
RNGD on Paper: 512 TFLOPS at 180 Watts
So, what does that efficiency advantage look like in concrete numbers? A direct comparison of raw specifications paints a clear picture. Each FuriosaAI RNGD accelerator delivers 512 teraFLOPS of FP8 compute while holding to a 180-watt thermal envelope. That is a strong showing in itself, but the contrast with Nvidia’s workstation-class GPU makes it even more telling.
Nvidia’s RTX Pro 6000 offers double the memory and compute performance, but it draws more than three times the power to reach those benchmarks. This gap in power consumption becomes critical when you scale up inference workloads across multiple servers. Energy costs quickly become a significant factor in total cost of ownership, and every watt you save directly improves your bottom line.
How RNGD Stacks Up Against Nvidia’s Data-Center GPUs
For inference tasks — where the model is trained and you run predictions — raw compute is only part of the equation. The efficiency per watt often dictates how many queries you can handle within a given energy budget. With 512 FP8 teraFLOPS packed into a 180-watt envelope, the RNGD positions itself as an inference accelerator built for throughput per watt. This makes it a compelling option for data centers that want to cut energy use without sacrificing performance.
The FuriosaAI RNGD Europe rollout brings this power-sipping hardware to a market increasingly focused on sustainable AI operations. When you compare the Nvidia RTX Pro 6000 vs RNGD, you see that higher raw compute isn’t always the best choice for your inference accelerator spec decisions. Instead, AI chip power efficiency can offer a smarter path forward for many real-world deployments.
The NXT RNGD Server: Air-Cooled and Rack-Compatible
After comparing raw compute figures, you might wonder how FuriosaAI’s approach translates to real hardware you can actually plug into a datacenter. The answer comes in the shape of the NXT RNGD Server, a machine built around eight of those RNGD accelerators. This isn’t a prototype locked away in a lab – it’s a 3kW system that is completely air-cooled and designed to fit into standard racks.
This simplicity is a big advantage for operators who want to add AI inference capacity without ripping out their existing infrastructure. The server’s air-cooled design means you don’t need expensive liquid cooling loops or specialized rack layouts. Instead, you can treat it much like a standard compute server, sliding it into a free slot and connecting power and networking. The lower power draw of 3kW for eight accelerators also keeps heat density manageable, reducing the strain on your facility’s cooling system.
FuriosaAI is already proving this works in practice. The RNGD servers are installed at Equinix’s LS2 datacentre in Lisbon, marking a concrete step for the Korean company’s expansion into Europe. This deployment at a major colocation provider shows that you can integrate the hardware into a production environment without custom modifications.
Retrofit Feasibility for Existing Datacenters
For companies running older facilities, retrofitting often means tearing out old racks or upgrading power distribution. With the NXT RNGD Server’s standard rack integration and air cooling, you may be able to skip those expensive upgrades. The hardware’s lightweight infrastructure requirements make it a practical option if you are looking to add inference capacity in space-constrained or power-limited racks.
- Standard racks: No need for proprietary chassis or modified rails.
- Air cooling: Works with typical room-level or row-level cooling systems.
- Moderate power: 3kW per server is manageable for most datacenter power distribution units.
This combination of compatibility and efficiency gives you a more flexible path toward deploying FuriosaAI RNGD Europe hardware, especially compared to high-power alternatives that demand a full facility redesign.
Software Ecosystem: FuriosaAI’s Compiler-First Approach
That flexibility doesn’t stop at the hardware level. To truly make FuriosaAI RNGD Europe hardware practical for developers, the company has invested heavily in a software stack built around a compiler-first philosophy. Instead of treating software as an afterthought, Furiosa runs a dedicated compiler-focused R&D lab in Lisbon, alongside a new flagship office. This setup signals a serious commitment to making model deployment feel straightforward, not like a puzzle.
The core idea here is that a smart compiler can handle the heavy lifting of optimization. When you write a model in a familiar framework, the FuriosaAI compiler stack automatically maps those operations onto the hardware for efficient execution. This means you don’t need to manually tune kernels or rewrite code to get good performance. The runtime then manages how your model runs on the actual chip, handling memory and scheduling so you can focus on your application.
Supported Frameworks and Models
For any AI chip to gain traction, it needs to play well with the tools developers already use. FuriosaAI’s approach centers on strong support for industry-standard frameworks. You can import models built with PyTorch, ONNX, and other popular formats directly into the FuriosaAI environment. The compiler then optimizes the model for the specific hardware, aiming for low latency and high throughput. This PyTorch ONNX support is critical because it lowers the barrier to entry—you can take an existing model, run it through the optimization runtime, and deploy it on a FuriosaAI system without a complete rewrite. It’s a practical, step-by-step path from your existing workflow to production on new hardware.
Future Roadmap: Third-Generation Chip with Broadcom
While that deployment workflow makes adoption smoother, Furiosa is already pushing further. The company is working with Broadcom on a third-generation AI accelerator designed for frontier models with a trillion or more parameters. That’s a big jump — trillion-parameter models are the kind that power the most advanced language and reasoning systems, and they demand enormous compute and memory bandwidth. By partnering with Broadcom, Furiosa gains access to world-class chip design and manufacturing expertise, which is critical for scaling up performance while keeping power in check. This Broadcom AI partnership signals that Furiosa isn’t just competing on software compatibility; it’s investing in the hardware muscle needed to tackle the next wave of AI workloads.
On a similar note, FERC Orders Faster Grid Access for AI Data Centers explores this topic with concrete examples.
So when can you expect to see this next-generation chip? Furiosa hasn’t shared a specific timeline or performance targets yet. What is clear is that the company has raised more than $250 million to date, giving it the runway for long-term R&D. That kind of funding, combined with a partner like Broadcom, suggests a serious scaling ambition. For anyone keeping an eye on the AI hardware landscape, FuriosaAI RNGD Europe is a name to watch — the company is building a path from practical, deployable accelerators today to high-end silicon for tomorrow’s largest models.
When to Expect the Next Generation
Details on availability are still under wraps, but the direction is unmistakable. Furiosa is positioning itself to go head-to-head with established players in the trillion-parameter space. If you’re planning long-term AI infrastructure, the FuriosaAI funding and Broadcom partnership are strong signals that a third-generation accelerator is on the horizon. Keep an eye on announcements — the next milestone could come sooner than many expect.
Supply Chain and Market Positioning: A Non-US Alternative
One of the more practical advantages for European buyers is where FuriosaAI sits in the global supply chain. Because the company is headquartered in South Korea, its chips fall outside the scope of current AI chip export controls that primarily target US-designed hardware. For European datacenters and cloud operators, this could mean a more predictable supply route, with fewer regulatory hurdles to navigate.
This positioning also helps reduce geopolitical risk AI hardware often carries. If trade restrictions tighten between the US and certain regions, a South Korean semiconductor supplier offers a reliable alternative. It diversifies the European AI supply chain, giving buyers another option without relying entirely on American or Chinese sources.
Customer Adoption and Early European Deployments
Beyond the supply chain angle, the practical question is who will actually deploy these chips. Equinix is the first named customer, but broader adoption remains unclear. That said, the company’s focus on low-power, air-cooled designs lowers the barrier for entry. You don’t need specialized liquid cooling or massive power budgets to run these accelerators, which makes them easier to slot into existing server racks.
This design philosophy could appeal to mid-sized European operators who want to experiment with AI workloads without overhauling their entire infrastructure. If FuriosaAI can secure more partnerships across the continent, it may carve out a meaningful niche as a non-US alternative in the AI hardware market. The FuriosaAI RNGD Europe rollout is still in its early stages, but the supply chain benefits are already clear on paper.
Frequently Asked Questions
How can you deploy FuriosaAI’s RNGD accelerator in an existing data center setup?
You can integrate the RNGD into your current infrastructure through its standard PCIe form factor, which fits into most modern servers. Ensure your system has the necessary power delivery and cooling, as the card is designed for efficient operation. You then install Furiosa’s software stack and compiler to optimize your AI models for inference workloads.
Is the RNGD chip better suited for training or inference compared to Nvidia alternatives?
The RNGD is strictly an inference engine, not a training accelerator. It focuses on running already-trained models efficiently, offering a practical alternative for tasks like real-time AI inference. This makes it a cost-effective choice if your primary need is deploying models rather than building them.
Does Furiosa’s RNGD support popular AI software frameworks like PyTorch or TensorFlow?
Yes, Furiosa’s compiler and runtime support common frameworks such as PyTorch and TensorFlow for model deployment. You can convert your trained models into a format optimized for the RNGD hardware. The compatibility continues to expand, but you should check the latest supported model list for your specific use case.
