Korean artificial intelligence chip companies are forging strategic partnerships to embed neural processing units (NPUs) into critical infrastructure, from data centers to financial institutions. These collaborations reflect a broader industry shift toward specialized AI hardware that can handle inference workloads more efficiently than general-purpose processors. The developments suggest that the next wave of AI deployment will depend less on cloud-based computing and more on purpose-built silicon designed for specific use cases.

What Are Neural Processing Units and Why Do Companies Need Them?

A neural processing unit is a semiconductor chip optimized to rapidly process the mathematical operations required when AI models perform real-world tasks. Unlike graphics processing units (GPUs), which were originally designed for rendering video, NPUs are built from the ground up to handle the specific computational patterns that AI inference demands. This specialization allows them to deliver better performance per watt of power consumed, a critical metric for large-scale deployments where energy costs can quickly become prohibitive.

Financial institutions face particular constraints that make NPUs attractive. Network separation regulations in the financial sector require banks to run AI services on internal networks rather than external cloud platforms. This regulatory requirement has created demand for on-premises AI infrastructure that can operate independently while maintaining security and compliance standards.

How Are Companies Positioning NPUs in Next-Generation Infrastructure?

FuriosaAI, a Seoul-based chipmaker, announced a strategic partnership with Broadcom, a global semiconductor giant, to develop integrated AI inference platforms. The collaboration goes beyond simple chip co-development to create what the companies describe as a "next-generation AI infrastructure platform" that combines computing, networking, and software capabilities.

FuriosaAI's third-generation AI accelerator, currently in development, will feature a Tensor Contraction Processor (TCP) architecture paired with HBM4 and HBM4E memory, advanced memory technologies that enable faster data movement between the processor and storage. Broadcom's advanced packaging technology will also be integrated into the design. Sampling of these new chips is scheduled to begin in the first half of 2028.

The partnership targets what the companies call the "token factory" era, referring to the massive-scale AI inference environments where large language models (LLMs) process enormous volumes of text and generate responses. FuriosaAI's second-generation accelerator, called Renegade (RNGD), has already been validated by Samsung SDS and the LG AI Research Institute, demonstrating real-world performance in enterprise and cloud provider environments.

"AI inference performance is no longer determined by simple computing performance alone, and data reuse and communication efficiency between servers and racks have become key competitive factors. By combining the technologies of FuriosaAI and Broadcom, we will build a platform that solves the key bottlenecks in large-scale agentic AI environments," said Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group.

Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group

What Role Are NPUs Playing in Edge and Consumer Infrastructure?

Beyond data center deployments, NPUs are being integrated into consumer-facing infrastructure. Broadcom has introduced the BCM68850, a 50G ITU-PON home gateway system-on-chip (SoC) that includes an integrated neural processing unit and native Wi-Fi 8 compatibility. This device represents a shift toward pushing AI inference closer to the broadband edge, where home gateways increasingly need to handle higher-speed access links while running local services.

The BCM68850 is designed to support symmetric 50-gigabit performance, meaning it can handle equally fast upload and download speeds. It includes features such as "intelligent self-healing" capabilities for anomaly detection and predictive bandwidth optimization, along with security features including Post-Quantum Cryptography (PQC), which protects against future threats from quantum computers. Broadcom is currently sampling the BCM68850 to early access customers and partners.

Steps to Understanding NPU Deployment in Your Infrastructure

• Identify Your Workload Type: Determine whether your organization needs to run AI inference on-premises due to regulatory requirements, latency constraints, or data sensitivity. Financial institutions, healthcare providers, and government agencies often fall into this category.
• Evaluate Power and Cooling Requirements: NPU-accelerated systems consume less power than traditional GPU-based approaches for inference tasks, but you should still assess whether your data center or facility can accommodate the thermal and electrical demands of new hardware.
• Plan for Integration with Existing Networks: New NPU platforms like those from FuriosaAI and Broadcom are designed to integrate with existing networking infrastructure, but you should verify compatibility with your current switches, cabling, and management systems before deployment.
• Consider Long-Term Roadmaps: Sampling timelines for next-generation chips extend into 2028, so organizations should align their infrastructure planning with these availability windows and plan for multi-year technology transitions.

How Are Financial Institutions Adopting NPU Technology?

KB Financial Group, one of South Korea's largest financial services companies, has deepened its partnership with Rebellions, another Korean AI chipmaker, to source NPUs and related technologies for advancing AI services across the organization. KB Financial and Rebellions first began collaborating in 2022 when KB Investment participated in early-stage funding. The relationship expanded in 2023 when Rebellions was selected for "KB Starters," KB Financial's startup incubation program, and KB Securities joined follow-on investment rounds.

Under the new strategic business partnership agreement signed in May, Rebellions will provide KB Financial with NPUs and other technologies needed to run AI services in practice. KB Financial will support Rebellions through business financing, management assistance, and financial services for executives and employees. The two companies also agreed to identify medium to long-term cooperation projects to advance Korea's AI ecosystem.

"Building on our partnership with Rebellions, we will work with various AI and tech partners to further strengthen KB Financial's AI competitiveness and secure global-level AI finance leadership," a KB Financial official said.

KB Financial Group official

What Do These Partnerships Signal About the Future of AI Hardware?

The convergence of partnerships between Korean chipmakers and global technology leaders suggests that specialized AI inference hardware is becoming a critical component of enterprise infrastructure. Rather than relying on cloud providers and general-purpose GPUs, organizations are increasingly investing in purpose-built silicon that can deliver better efficiency, lower latency, and compliance with regulatory requirements.

FuriosaAI's collaboration with Broadcom is particularly significant because it combines chip design expertise with networking and infrastructure capabilities. The partnership acknowledges that modern AI inference at scale is not just about raw computing power; it also depends on efficient data movement between processors, memory, and network infrastructure. This holistic approach to platform design reflects lessons learned from deploying large language models and agentic AI systems in production environments.

The timeline for these developments extends into 2028, indicating that organizations planning AI infrastructure investments today should begin evaluating NPU-based solutions and understanding how they fit into their long-term technology roadmaps. The combination of regulatory drivers in finance, efficiency gains in data centers, and edge computing requirements in consumer broadband suggests that NPUs will become as foundational to AI deployment as GPUs have been to AI training over the past decade.