A South Korean artificial intelligence chip company has secured a major contract to power a new data center in Sweden, signaling growing international demand for specialized AI processors beyond traditional graphics processing units. FuriosaAI announced it will supply 8,800 neural processing units (NPUs) to an AI data center being built in Stockholm by U.S. infrastructure company I/ONX HPC and data center operator Velox.

What Are Neural Processing Units and Why Do Data Centers Need Them?

Neural processing units are specialized computer chips designed specifically to handle artificial intelligence workloads, particularly the inference phase where trained AI models process real-world data and generate responses. Unlike graphics processing units (GPUs), which excel at training AI models, NPUs are optimized for running already-trained models efficiently and with lower power consumption. This distinction matters enormously for data center operators trying to balance performance with energy costs.

The Stockholm facility represents a shift in how large-scale AI infrastructure is being built. Rather than relying exclusively on GPUs, the data center will use a mixed approach where FuriosaAI's NPUs handle large language model inference and agentic AI tasks, while GPUs handle model training and other computational work. This heterogeneous computing environment allows the facility to optimize both overall performance and power efficiency simultaneously.

How Will the Stockholm Data Center Deployment Unfold?

• Phase One Timeline: A 2-megawatt initial facility will begin operating in early 2027, equipped with 1,800 units of FuriosaAI's second-generation Renegade (RNGD) NPU chips manufactured by Taiwan Semiconductor Manufacturing Company (TSMC)
• Expansion Plans: An additional 8-megawatt facility will be constructed by the second half of 2027, with progressive introduction of NPUs bringing the total to 8,800 units across all phases
• Infrastructure Scale: The complete project will reach 15 megawatts of total capacity, making it a substantial AI computing hub for European operations

The Renegade NPU will be mounted on I/ONX HPC's "Symphony SixtyFour" high-density AI server platform, a specialized hardware configuration designed to maximize computational density while managing thermal and power constraints.

What Does This Deal Mean for FuriosaAI and the Broader AI Chip Market?

This contract represents a significant validation of FuriosaAI's technology in the competitive global AI chip market. The company, which is based in Seoul, has been developing NPU technology to compete with established players in AI acceleration hardware. By securing a large-scale European deployment, FuriosaAI demonstrates that its inference-focused approach resonates with data center operators seeking alternatives to GPU-centric architectures.

"This project is an important milestone in applying FuriosaAI's AI inference resources to a large-scale European data center. We will continue to accelerate the expansion of next-generation AI inference infrastructure together with our global partners," said Baek June-ho, CEO of FuriosaAI.

Baek June-ho, CEO of FuriosaAI

The deal also reflects broader industry trends in AI infrastructure. As large language models and other generative AI applications move from research labs into production environments, the bottleneck has shifted from training capacity to inference efficiency. Data center operators increasingly recognize that inference workloads require different hardware optimization than training, and NPUs are purpose-built for exactly this use case.

Why Is Europe Investing Heavily in AI Data Center Infrastructure?

Europe has been accelerating its AI infrastructure investments to reduce dependence on U.S.-based cloud providers and build sovereign computing capacity. The Stockholm facility fits into this broader European strategy to develop homegrown AI computing resources. By partnering with a Korean chip manufacturer rather than relying exclusively on U.S. or Chinese suppliers, the project also diversifies the geographic and corporate sources of critical AI infrastructure.

The heterogeneous computing approach used in Stockholm also reflects a maturation of the AI infrastructure market. Early data centers often treated all AI workloads the same, deploying uniform GPU clusters. Modern facilities recognize that different AI tasks have different hardware requirements. Inference workloads, which dominate production AI systems, benefit from specialized processors like NPUs that deliver better performance per watt than general-purpose accelerators.

FuriosaAI's entry into the European market with this substantial order signals that the NPU category is moving beyond niche applications into mainstream data center deployments. As AI inference becomes a core business function for enterprises across Europe, the demand for efficient, specialized inference hardware will likely continue growing, potentially opening doors for other NPU manufacturers to secure similar large-scale contracts.