Two Korean semiconductor startups are positioning themselves as critical partners for Europe's push toward AI independence by deploying neural processing units (NPUs) that consume far less power than traditional graphics processors. Furiosa AI and Rebellion are aggressively expanding their European presence through new offices, partnerships with local AI companies, and ambitious sales targets, capitalizing on a geopolitical shift toward sovereign AI development .

What Are Neural Processing Units and Why Does Europe Care?

Neural processing units are specialized chips designed specifically for running artificial intelligence inference, the process of using a trained AI model to generate outputs like chatbot responses or image classifications. Unlike graphics processing units (GPUs), which were originally built for rendering video game graphics and later adapted for AI training, NPUs are purpose-built for efficiency on edge devices and data centers where power consumption directly impacts operating costs and environmental impact .

Europe's interest in NPUs reflects a broader strategic concern: reducing dependence on US-dominated AI infrastructure. As geopolitical tensions rise and countries seek to build sovereign AI capabilities, NPUs offer a compelling alternative because they deliver comparable performance to GPUs while consuming significantly less electricity. This efficiency advantage is particularly valuable for data centers running inference workloads at massive scale, where power bills can exceed hardware costs .

How Are Korean Startups Capturing Europe's AI Market?

• Geographic Expansion: Furiosa AI established a new entity in Portugal last month, strengthening its European network alongside existing offices in the United States and Germany, positioning itself as a local partner rather than a distant supplier .
• Strategic Partnerships: Rebellion CEO Park Sung-hyun personally met with executives at French AI company Mistral AI to discuss supplying the company's second-generation chip, called "Rebel 100," to European data centers .
• Aggressive Sales Targets: Furiosa AI aims to sell 20,000 units of its second-generation chip, "Renegade," by year-end, demonstrating confidence in European demand for locally-sourced AI hardware .

Both companies are pursuing what industry analysts call a "sovereign AI" strategy, leveraging NPUs to position themselves as core infrastructure partners for European governments and enterprises seeking to reduce reliance on US technology. This approach mirrors broader trends in the semiconductor industry, where custom silicon designed for specific tasks is increasingly replacing general-purpose processors .

Why NPUs Matter More Than You Might Think

The global AI chip market was valued at approximately $67 billion in 2024 and is projected to exceed $300 billion by 2030, according to market research firm Grand View Research . Within this expanding market, NPUs represent a growing segment because they address a critical pain point: the enormous power consumption required to run AI models at scale.

To understand the significance, consider that training large language models like GPT-4 can consume millions of kilowatt-hours of electricity, raising major sustainability concerns for data center operators and governments alike . Inference, which happens billions of times per day across chatbots, recommendation systems, and autonomous vehicles, compounds this problem. NPUs are designed to reduce power consumption for inference workloads compared to GPUs through optimizations in memory bandwidth and precision handling, making them particularly attractive for cost-sensitive and environmentally conscious deployments .

For Europe, this efficiency advantage translates directly into lower operating costs, reduced carbon emissions, and reduced dependence on Taiwan's TSMC, which manufactures the most advanced chips globally and creates a geopolitical vulnerability for the entire AI industry .

What Does This Mean for the Broader AI Hardware Landscape?

The success of Furiosa AI and Rebellion in Europe signals a fundamental shift in how AI infrastructure is being built. For years, NVIDIA dominated the AI accelerator market with roughly 70 to 80 percent market share, primarily through its GPU offerings . However, major technology companies including Google, Amazon, Microsoft, and Meta have all developed custom silicon tailored to their specific workloads, recognizing that one-size-fits-all hardware cannot optimize for both cost and performance .

Google's Tensor Processing Unit (TPU), designed specifically for neural network math, delivers 459 teraFLOPS per chip at BF16 precision, a measurement of computational speed at lower numerical precision . Amazon offers the Trainium2 for training and Inferentia2 for inference. Microsoft developed the Maia 100 for training large language models inside Azure. Meta built the MTIA (Meta Training and Inference Accelerator) for inference at its massive scale of content recommendation .

Korean startups entering this competitive landscape with NPUs optimized for inference represents a new phase in AI hardware fragmentation, where regional players with deep expertise in power-efficient design can capture meaningful market share by addressing specific geographic and regulatory needs.

"AI workloads fundamentally have memory constraints as the issue," explained Jung Hee-jin, Head of Venture Investment at SK hynix America, when discussing how semiconductor architecture must be designed from the ground up to reflect these constraints.

Jung Hee-jin, Head of Venture Investment at SK hynix America

This comment reflects a broader industry recognition that the future of AI hardware lies not in raw processing power alone, but in architectural innovations that address the real bottlenecks: memory bandwidth, power efficiency, and latency. NPUs, by focusing specifically on inference rather than trying to handle both training and inference, can optimize for these constraints in ways general-purpose GPUs cannot .

The Geopolitical Dimension of AI Hardware

Europe's push toward sovereign AI development is not purely technical; it reflects genuine concerns about supply chain vulnerability and technological autonomy. The entire global AI industry depends on TSMC in Taiwan for manufacturing the most advanced chips, creating a single point of failure for the world's AI infrastructure . By developing and deploying NPUs from Korean suppliers, European companies and governments can diversify their supply chains while supporting the development of alternatives to US-dominated AI platforms.

Furiosa AI and Rebellion are well-positioned to capitalize on this shift because they bring both technical expertise in power-efficient chip design and the advantage of being non-US suppliers. Their expansion into Europe signals that the AI hardware market is becoming increasingly multipolar, with regional players competing for market share based on efficiency, cost, and geopolitical alignment rather than raw performance alone .