FuriosaAI RNGD Inference Accelerator Lands at Equinix Lisbon: Power-Efficient AI for Europe

European AI teams need more inference capacity, but many grids, power envelopes, and legacy data centers cannot support megawatt‑scale GPU clusters without costly upgrades.[4] FuriosaAI, led by CEO June Paik, is deploying its RNGD inference accelerators at Equinix’s LS2 facility in Lisbon as an alternative—frontier‑grade inference inside a 3 kW power budget per server.[1][2][4]

For ML engineers, LS2 becomes a place to test and scale LLM and agent workloads on hardware tuned for tokens‑per‑watt, while staying within typical European rack and power limits.[4]

Key takeaway: Lisbon’s LS2 gives EU enterprises local, low‑latency access to power‑efficient inference hardware without waiting for AI‑only data centers.[1][2]

Why FuriosaAI Chose Equinix Lisbon for RNGD’s European Launch

FuriosaAI is rolling out its NXT RNGD servers at Equinix LS2 as the first node in a broader European footprint.[1][3][4] LS2 operates as a neutral, carrier‑dense colocation hub where customers bring their own networks and data and tap into specialized AI compute.

To anchor this, FuriosaAI opened a Lisbon office that combines:[1][4]

• Commercial operations
• Customer engineering
• R&D for compilers, chip design, and PCB design

Having engineering teams close to production infrastructure lets firmware, kernel, and compiler changes be validated quickly against real workloads.

Europe’s AI ambitions are constrained by:[4][5]

• Limited grid capacity and power availability
• Difficulty adding 30–40 kW GPU racks without major upgrades
• Cooling limits in legacy facilities

RNGD addresses this gap by delivering high inference density in 3 kW‑class servers that fit into standard air‑cooled racks.[1][4][5]

Equinix LS2 offers:[2]

• 2,050 sqm of colocation space over three floors
• Initial capacity for 625 racks
• An efficiency‑focused design aligned with FuriosaAI’s emphasis on energy‑efficient inference[1][2]

Real‑world angle: SaaS and fintech firms exploring LLM‑based support or analytics often find power and rack constraints—not model cost—block GPU cluster deployments; LS2 plus RNGD allows scale within existing colocation contracts.[4][5]

Inside the RNGD Architecture and NXT RNGD Server at LS2

RNGD is built on FuriosaAI’s proprietary Tensor Contraction Processor (TCP) architecture, fabricated on a 5 nm process.[1][2][4][5] Each accelerator delivers:

• 512 TFLOPS of FP8 performance
• 180 W thermal design power

This is tuned for production inference, especially LLMs where FP8 throughput and low TDP maximize tokens‑per‑watt within tight rack power budgets.[1][2][4][5]

At the system level, up to eight RNGD accelerators integrate into the NXT RNGD Server, yielding a 3 kW‑class system with:[1][2][4][5]

• Up to 384 GB of HBM across cards
• Capacity to run 100B‑plus parameter LLMs at high concurrency
• No need for exotic cooling or major facility upgrades[5]

Key point: NXT RNGD servers are air‑cooled and drop into standard racks, avoiding liquid‑cooling retrofits that can delay deployments for months.[4][5]

Compared to more power‑hungry accelerators, RNGD’s 180 W profile is modest; Nvidia’s RTX Pro 6000‑class devices can draw ~3.33× more power for similar workloads.[5] In a European rack capped at ~5–8 kW, this can be the difference between:

• One GPU‑heavy node, or
• A multi‑server inference cluster.[4][5]

The Lisbon deployment reflects a trend of placing AI compute where power is efficient and sustainable, similar to Hive Digital’s use of Paraguay’s renewable‑heavy grid for AI workloads.[7] AI hubs are now chosen for energy characteristics as much as latency.[1][7]

Within the inference ecosystem, RNGD sits alongside LLM‑optimized chips like OpenAI’s Jalapeño, which targets substantially better performance per watt than current state‑of‑the‑art devices.[8][9] Both approaches share a thesis:[4][8][9]

• Workload‑specific, full‑stack design beats raw peak FLOPS for serving large models.

Key takeaway: RNGD’s advantage is not just “512 FP8 TFLOPS” but delivering that throughput inside 180 W per card, enabling meaningful cluster sizes in power‑constrained European racks.[1][4][5]

Enterprise Benefits, Use Cases, and How to Engage in Lisbon

For EU enterprises, LS2 offers three primary benefits:[1][2][4]

• Local, low‑latency RNGD access for EU‑hosted data and users
• A controlled environment to benchmark LLM and agentic workloads on specialized inference hardware
• Predictable power and rack planning via 3 kW‑class servers

Typical engagement starts with a proof‑of‑concept supported by FuriosaAI’s Lisbon technical team, where customers port a few representative services—such as a retrieval‑augmented LLM API—to NXT RNGD servers.[1][4] They then benchmark:

• Tokens per second and latency at target concurrency
• Tokens per watt and per‑request energy cost
• Total cost of ownership vs. existing GPU setups over 1–3 years[1][4][5]

FuriosaAI positions RNGD for LLMs and agentic AI, with a software stack that reduces the need for hand‑tuned kernels and minimizes migration overhead.[4]

Illustrative use cases include:[1][4][5]

• Financial services: Risk models and multilingual chatbots that must stay within regulated data‑center power ceilings.[4]
• Customer support automation: High‑throughput ticket triage and conversational agents with fixed rack footprints.[1][4]
• Public‑sector and sovereign AI: EU‑based, power‑efficient infrastructure that aligns with data residency and sustainability mandates.[4][5]

Lisbon is a template for additional European rollouts, likely reusing the “local office + colocation + power‑efficient inference” pattern across more hubs.[1][3][4] PoCs proven in Lisbon are intended to be portable to future Furiosa‑enabled regions.

Key point: Understanding your workload on low‑TDP inference hardware early makes later scaling easier without re‑architecting around grid or cooling constraints.[1][4][5]

Conclusion: A Practical Path to Efficient European Inference

FuriosaAI’s RNGD deployment at Equinix LS2 combines a 5 nm, FP8‑optimized inference accelerator with a scalable, power‑aware colocation environment, offering European enterprises a practical way to grow modern AI workloads without waiting for entirely new data centers.[1][2][4][5]