Summary
- FuriosaAI is installing RNGD AI inference servers at Equinix’s Lisbon LS2 data centre.
- The RNGD accelerator uses a 5nm Tensor Contraction Processor architecture, with each accelerator operating at a 180W thermal design profile.
- The deployment gives FuriosaAI a European colocation environment for customer evaluation, performance testing, and production-readiness work.
FuriosaAI is installing RNGD AI inference servers at Equinix’s Lisbon LS2 data centre, giving the South Korean chip company a live European infrastructure base for customer evaluation and deployment work.
The installation is tied to FuriosaAI’s recently opened Lisbon office and is intended to make its RNGD accelerator available to European enterprises. The company said the on-site environments will allow customers to test performance and usability on advanced models.
The RNGD accelerator uses a 5nm Tensor Contraction Processor architecture and is rated at 512 TFLOPS FP8 with a 180W thermal design profile per accelerator. FuriosaAI integrates up to eight accelerators into its NXT RNGD Server, creating a 3kW-class inference system for production AI workloads.
Those figures put the story beyond a routine hardware availability update. Inference hardware has to prove itself inside real facility constraints: rack power, thermal behaviour, network proximity, maintenance access, remote management, and customer support.
By using Equinix infrastructure in Lisbon, FuriosaAI is putting its accelerator stack into a global colocation environment rather than keeping it in a lab or vendor-controlled test setting. That gives the deployment a more practical route to customer trials across Europe.
Inference enters the facility floor
Europe’s AI infrastructure debate has been dominated by training clusters, GPU shortages, and sovereign cloud proposals. Inference brings a different pressure pattern. Once AI applications move into regular use, the cost, latency, power draw, and operational repeatability of serving models become decisive.
FuriosaAI’s pitch rests on useful output per watt and per rack, not just theoretical accelerator performance. A 3kW inference server is not an extreme load by AI standards, but customers will judge the platform on how much dependable throughput it can deliver within ordinary data centre operating limits.
Lisbon gives the company a sensible European landing point. Iberia is drawing more attention because of renewable energy potential, international connectivity, transatlantic routes, and expansion opportunities outside the most congested data centre metros. A deployment at LS2 also places the hardware inside a carrier-connected environment that can serve enterprise testing without requiring each customer to build its own platform.
The agentic AI market that FuriosaAI is targeting can create uneven and high-concurrency inference demand. That pushes the burden onto software scheduling, observability, model support, and network performance as much as raw silicon. Hardware that performs well in isolation still has to run cleanly inside customer workflows.
Alternative silicon meets European constraints
Europe’s search for AI capacity has created openings for alternatives to the dominant GPU supply chain. Those openings are not guaranteed. Buyers still have to weigh tooling, ecosystem maturity, model compatibility, procurement risk, and service support against any gains in efficiency or cost.
A live deployment in Equinix gives FuriosaAI a way to reduce adoption friction. Customers can test systems closer to their operations while seeing how the hardware behaves under standard facility controls, including power, cooling, access, remote monitoring, and support processes.
The next test is repeatability. A first installation can validate performance and create reference customers, but the market is moving towards larger clusters and consumption models that feel closer to cloud capacity than specialist equipment resale.
For Lisbon, the deployment adds another AI workload to Iberia’s infrastructure map. For FuriosaAI, it turns a hardware claim into an operating environment where power density, thermal behaviour, and customer usability will decide whether RNGD can become a real European compute option.
