FuriosaAI has secured a planned deployment of more than 7,000 RNGD inference accelerators at a new Stockholm data centre, giving the South Korean chip designer a significant European test for hardware intended to compete with conventional GPU-heavy infrastructure.

The project brings together FuriosaAI, computing-platform integrator I/ONX HPC, and data-centre developer Velox. Its first 2MW of AI capacity is expected to begin operating in early 2027, followed by another 8MW later that year as the site progresses towards a total planned capacity of 15MW.

I/ONX intends to install 1,800 RNGD accelerators in the opening phase and more than 7,000 across subsequent expansions. The chips will operate alongside GPUs within the company’s Symphony SixtyFour platform, allowing different workloads to be directed towards hardware selected for their particular performance, power, and software requirements.

Velox is developing and operating the physical facility, which will sell capacity through an infrastructure-as-a-service model, while I/ONX will act as system integrator and platform provider. FuriosaAI is responsible for the accelerators, drivers, software stack, and architectural support needed to run models on its hardware.

Inference challenges the GPU default

Much of the current AI infrastructure market has been shaped by training large models, where high-end GPUs and tightly coupled clusters dominate spending. Commercial use creates a different workload, because inference involves generating responses, recommendations, images, or other outputs repeatedly after a model has been trained.

Purpose-built inference accelerators are intended to reduce the cost and electricity required for that recurring work, although they also face a substantial software and adoption problem. Developers have spent years optimising applications for established GPU ecosystems, and any alternative supplier must show that models can be moved, maintained, and updated without losing reliability or forcing customers into extensive engineering work.

FuriosaAI says RNGD operates at a thermal design power of 180 watts and will improve performance per watt when used for suitable inference workloads. The Stockholm project could provide evidence for that claim at facility level, but the companies have not disclosed expected model performance, customer pricing, utilisation, or independently verified comparisons with current GPU systems.

The announcement consequently represents an infrastructure commitment rather than a completed operating benchmark. Technical validation, software integration, and supply planning will continue over the coming months, while the larger deployment depends on the first phase operating successfully and demand arriving for the service.

More suppliers, not European sovereignty

European policymakers have made access to AI computing capacity part of their industrial strategy, and the EU is developing a wider plan for expanding sovereign compute infrastructure. A Swedish facility using a South Korean processor does not create a European semiconductor supply chain, but it could reduce dependence on a narrow group of dominant chip and cloud suppliers.

Infrastructure resilience does not require every component to originate within the region, although it does require operators to have credible alternatives, portable software, and enough purchasing power to avoid becoming captive to one hardware roadmap.

Stockholm offers an established data-centre market, reliable power infrastructure, and connectivity to customers across northern Europe. Yet a project of this size will still be judged on electricity access, construction delivery, cooling, customer demand, and whether accelerator supply can keep pace with the proposed expansion.

The heterogeneous design may prove more consequential than any individual processor. As AI workloads diversify, operators are likely to combine CPUs, GPUs, inference accelerators, storage systems, and networking hardware rather than build every service around the most expensive general-purpose chip available.

By early 2027, the first 2MW phase should begin to show whether that approach produces a usable service rather than an attractive architecture diagram. FuriosaAI has gained a valuable European reference project, but the deployment will earn its weight only when customers are running production models and the promised efficiency survives measurement outside the supplier’s own specifications.