Summary
- The Stockholm project starts with 2MW and 1,800 FuriosaAI accelerators, before expanding towards 15MW.
- More than 7,000 RNGD accelerators are planned across later phases alongside other GPU and accelerator technologies.
- Technical validation, software integration, cooling design, and supply planning remain under way ahead of the first deployment.
FuriosaAI, I/ONX HPC, and Velox DCI Europe plan to deploy more than 7,000 AI inference accelerators at a new 15MW data centre in Stockholm, with the first 2MW of compute capacity due online in early 2027.
The initial phase is expected to contain 1,800 of FuriosaAI’s RNGD accelerators. A further 8MW is planned later in 2027, taking the project to 10MW before a subsequent expansion towards the full 15MW facility capacity.
Velox is developing, operating, and managing the physical data centre under an infrastructure-as-a-service model. I/ONX will act as system integrator and platform partner, while FuriosaAI will supply accelerator silicon, drivers, software, and architectural support.
The partners said construction has begun. Technical validation, software integration, and supply planning are continuing ahead of the first phase, leaving the programme dependent on both building delivery and the readiness of the compute platform.
A heterogeneous inference platform
The planned system is not limited to a single accelerator supplier. I/ONX’s Symphony SixtyFour platform is intended to support heterogeneous configurations, with FuriosaAI silicon operating alongside GPUs and other accelerators.
That approach is designed to match different inference workloads with different computing resources rather than treating every model as a job for the same class of GPU. The commercial proposition rests on obtaining more useful output from each megawatt while preserving the flexibility to run models with different latency, memory, and software requirements.
FuriosaAI rates the RNGD accelerator at a thermal design power of 180 watts and positions it for high-throughput, low-latency inference. The company argues that purpose-built inference silicon can reduce energy use and heat output compared with relying entirely on higher-powered general-purpose GPU platforms.
Those claims will require validation at system level. Accelerator power ratings do not capture the full energy demand of servers, memory, networking, storage, cooling, power conversion, and redundancy. Performance per watt will also depend on the models being served, utilisation, batching, latency requirements, and the maturity of the software stack.
The Stockholm deployment therefore combines a hardware trial with a facility-delivery programme. More than 7,000 accelerators will require substantial network fabric, electrical distribution, maintenance access, and heat-removal capacity even where individual devices consume less power than competing products.
Two megawatts before fifteen
The phased timetable reduces the amount of equipment that must be installed and commissioned at once. The first 2MW can establish whether the platform, software, and facility operate as intended before the project commits to the additional 8MW planned later in 2027.
It also leaves room to change the technology mix. AI hardware cycles are shorter than data centre building lives, and a project expected to expand over several phases may face new generations of accelerators, different rack architectures, or revised customer demand before reaching 15MW.
Sweden offers an established data centre market, comparatively low-carbon electricity, and a climate that can support efficient heat rejection for much of the year. Those advantages do not eliminate constraints around available grid capacity, construction lead times, or the local use of recovered heat.
The project’s focus on inference is notable. Training clusters attract the largest power headlines, but deployed AI services create continuous demand for token generation and model execution. Inference capacity can require a different balance of latency, network proximity, utilisation, and hardware economics from large training runs.
Velox and I/ONX will need to turn that proposition into an operational service rather than a hardware count. Customers will require defined availability, performance, security, data-handling, and support commitments, while the operator must maintain a mixed silicon estate across successive deployments.
The first 2MW will provide the evidence. Until that phase is commissioned, the 7,000-accelerator figure remains a planned expansion target attached to a facility that is still being built and integrated.
