FuriosaAI said on August 4 that it will supply 1,800 RNGD inference accelerators for the first phase of a Stockholm AI data center being developed with I/ONX HPC and Velox. The first 2 MW of compute is expected online in early 2027, with another 8 MW later that year and more than 7,000 additional RNGD accelerators planned as the facility scales toward 15 MW.
FuriosaAI said on August 4 that it will supply its RNGD inference accelerators for a new AI data center in Stockholm being developed with I/ONX HPC and Velox. The partners expect the facility to scale to 15 MW, combining FuriosaAI's specialized inference hardware with GPU capacity.
Velox is developing and operating the facility, while I/ONX HPC is serving as the system integrator. FuriosaAI's role covers RNGD silicon, drivers, its software stack, reference architectures, and technical support for deployment.
The first phase targets early 2027
The first 2 MW of AI compute capacity is expected online in early 2027. The partners plan to add another 8 MW later that year, with subsequent phases taking the site toward its 15 MW target.
FuriosaAI says I/ONX plans to deploy 1,800 RNGD accelerators in phase one, followed by more than 7,000 additional RNGD accelerators in later phases. That wording implies a cumulative deployment above 8,800 units. Yonhap reports a total of 8,800 units, while Seoul Economic Daily and ChosunBiz describe the cumulative plan as 8,800 or more than 8,800. These accounts are broadly consistent, although the partners have not published an exact final ceiling for the later phases.
A mixed accelerator design
FuriosaAI says its chips will operate alongside GPUs on I/ONX HPC's Symphony SixtyFour platform. ChosunBiz reports that RNGD is intended for large language model and agentic-AI inference, while GPUs would handle training and other compute tasks. The division reflects a heterogeneous design in which different workloads are assigned to hardware chosen for the job rather than using GPUs alone.
The company describes RNGD as a 180-watt inference accelerator. That specification helps explain the project's emphasis on power-efficient serving, but it does not establish the economics of the deployment by itself. The partners have not yet published workload-level results for this facility, such as model throughput, latency distributions, utilization, availability, or cost per token.
What infrastructure teams should watch
For ML platform teams, the meaningful proof will come after deployment. Useful comparisons will need to hold model, precision, batch profile, service-level targets, and software maturity constant across hardware. Integration with serving frameworks, scheduling across mixed accelerators, observability, and model coverage will matter as much as chip-level performance. The project is therefore a significant planned European deployment for non-GPU inference silicon, not yet evidence that RNGD outperforms GPU alternatives in production. The first operational milestone is the planned 2 MW phase in early 2027.
Key Points #
- 1FuriosaAI says phase one will use 1,800 RNGD accelerators, with more than 7,000 additional units planned across later expansion phases.
- 2The first 2 MW is expected online in early 2027, another 8 MW later that year, and the facility is designed to scale toward 15 MW.
- 3The heterogeneous design pairs RNGD inference hardware with GPUs, but production throughput, latency, utilization, and cost metrics have not yet been published.
Scoring Rationale #
This is a notable planned European AI infrastructure deployment involving specialized inference silicon at meaningful data-center scale. It is relevant to practitioners evaluating alternatives to GPU-only inference, although operational performance and the exact final accelerator count remain unverified.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.