← Back to feed

FuriosaAI is installing its RNGD AI inference accelerator at Equinix’s Lisbon LS2 data centre, the company announced July 7. The deployment is the South Korean chip startup’s initial step into European infrastructure and the first opportunity for European enterprises to evaluate the hardware in a production-adjacent environment without shipping units from Asia.

Furiosa has simultaneously established a flagship office in Lisbon to support the deployment. The timing aligns with a broader European push toward non-US, non-Chinese AI infrastructure options as the EU Chips Act and digital sovereignty directives push governments and large enterprises to diversify their compute supply chains.

The RNGD Architecture

The RNGD — Reconfigurable Neural-network Dynamic accelerator — is built on a 5nm Tensor Contraction Processor architecture. That foundry node matches the Nvidia H100’s manufacturing process; the architectural departure is in how computation is organised. Rather than the SIMT (single instruction, multiple threads) model that underpins modern GPU compute, the TCP treats tensor operations as first-class primitives, eliminating overhead from the instruction pipeline that GPUs incur when handling the specific patterns of transformer inference workloads.

The performance spec is 512 TFLOPS at FP8 precision, within a 180W thermal design profile. The 180W constraint is deliberate: enterprise data centres are increasingly power-constrained, and a chip that delivers high throughput inside a strict power envelope has a density advantage per rack that raw TFLOPS figures obscure.

The NXT RNGD Server packages up to eight RNGD accelerators into a purpose-built inference chassis, running at 3kW-class total draw. The design targets maximum compute density per rack without requiring dedicated high-power infrastructure, a relevant distinction for European colocation facilities operating under grid access limitations that North American hyperscaler campuses do not face to the same degree.

The European Context

Furiosa’s Lisbon deployment arrives at a specific moment. The EU Formally Launched its Tech Sovereignty Package in late June, which includes a €200 billion data centre build target and EUR 2 billion for open-source infrastructure. Among the structural aims is reducing dependence on US hyperscaler infrastructure for sensitive workloads in health, finance, and government.

That creates a market opening for inference hardware that is neither Nvidia nor Huawei — the two dominant alternatives that also carry their own dependency and compliance complications. FuriosaAI is a private company backed by Samsung and other Korean investors; its supply chain runs through Korean and Taiwanese foundry partners rather than through any entity under current export control review.

AMD’s MI355X has made inroads with Chinese labs — GLM-5.2 was benchmarked running at 2,626 tokens per second at 2x lower cost than Blackwell on Instinct hardware. Furiosa occupies a different position: purpose-built for inference rather than retrofitted from a training-grade GPU, and physically present in Europe rather than requiring customs and export documentation.

What Comes Next

The Lisbon LS2 deployment is explicitly described as an evaluation and demonstration environment. European enterprises can test frontier models on RNGD hardware directly, comparing throughput, latency, and cost-per-token against their existing GPU-backed deployments. Equinix operates facilities across Frankfurt, Amsterdam, London, and Paris — the current Lisbon installation is likely a precursor to broader Equinix deployment if customer evaluation converts to demand.

Furiosa has not disclosed contract pricing for cloud access to RNGD capacity or the commercial terms of the Equinix arrangement. Revenue model, ASP, and gross margin structure for the inference-as-a-service path remain unspecified. What is clear is that a second non-Nvidia inference chip option — after Intel’s Gaudi, which has struggled to gain traction — is now physically present in European infrastructure for the first time.