Combination of Korean NPU with Broadcom's Ethernet and Packaging

Targeting the Agentic AI Era

Foreign Media: "Korea's FuriosaAI Joins the Broadcom AI Empire"

Shift in the AI Race: Beyond GPU Performance, the Battle for Data Movement and Power Efficiency Begins

"AI infrastructure is no longer about chip competition, but about a race to build Token Factories. FuriosaAI will work with Broadcom to build next-generation AI inference infrastructure."
Paek Joungho, CEO of Furiosa AI, is speaking at an event introducing the second-generation Renegade NPU on April 2nd. Photo by Paek Jongmin, Tech Specialist

Paek Joungho, CEO of Furiosa AI, is speaking at an event introducing the second-generation Renegade NPU on April 2nd. Photo by Paek Jongmin, Tech Specialist

View original image

Paek Joungho, CEO of FuriosaAI, has shown a strong commitment to ushering in the post-NVIDIA era by joining forces with global semiconductor company Broadcom. The international media also took notice as a domestic AI semiconductor startup and a world-leading chipmaker agreed to jointly build next-generation AI inference infrastructure.


FuriosaAI, an NPU startup specializing in AI semiconductors, announced on the evening of the 27th that it will collaborate with Broadcom to jointly develop a third-generation chip targeting the Agentic AI era. The two companies plan to build large-scale AI inference infrastructure by combining FuriosaAI's next-generation NPU with Broadcom's advanced packaging and Ethernet networking technologies. This marks the birth of an 'alliance' that leverages each company's strengths.


This collaboration goes beyond simply supplying FuriosaAI's NPUs; the companies will jointly design the entire AI infrastructure at the server and rack level. FuriosaAI will apply a TCP (Tensor Contraction Processor) architecture and HBM4/HBM4E memory to the third-generation Renegade chip, while Broadcom will provide ultra-high-speed Ethernet switches, PCIe, and advanced packaging technologies. Chip production will utilize TSMC's 2-nanometer process. FuriosaAI will also provide software support, not just the hardware.


Through this process, it is expected that the previous limitation of eight NPUs per server can be increased. This is the result of integrating NPU and network technologies. Whereas existing NPUs were intended for high-performance servers or low-density data centers, the addition of Broadcom enables the infrastructure to compete with GPUs in terms of performance. NVIDIA already supports GPU-to-GPU connections through NVLINK and Mellanox, making it possible to utilize large quantities of GPUs.

Eight NPUs from FuriosaAI are installed in the server. Photo by Paek Jongmin, Tech Specialist

Eight NPUs from FuriosaAI are installed in the server. Photo by Paek Jongmin, Tech Specialist

View original image

The company explained that sampling of the third-generation Renegade chip, developed in partnership with Broadcom, will begin in the first half of 2028.


Last year, FuriosaAI declined an acquisition proposal from Meta, but chose to collaborate with Broadcom. This decision reflects the view that the partnership could yield significant results and demonstrates FuriosaAI's elevated status over the past year. An industry insider commented, "Although CEO Paek's decision to reject Meta's offer during a severe crisis was surprising, the situation changed rapidly afterward."


Currently, Meta's market capitalization is approximately $1.6 trillion, which is still less than Broadcom's $2 trillion. While Meta's corporate value has stagnated due to falling behind in the AI race, Broadcom's growth has been remarkable. Broadcom is now ranked sixth among U.S. listed companies by market capitalization, while Meta is ranked ninth.


Broadcom, a global semiconductor leader, is rapidly expanding its custom AI semiconductor and networking business to major global tech companies such as Google and Meta, positioning itself as one of the main challengers to NVIDIA.


Industry experts interpret this collaboration as a reflection of the AI industry's shift from training to inference, where data transfer efficiency and power efficiency are becoming more important than raw GPU performance.


Charlie Kawwas, President of the Semiconductor Solutions Group at Broadcom, stated, "Inference performance is determined not by simple computation, but by data movement efficiency," adding, "The combination of FuriosaAI's TCP architecture and Broadcom's AI networking technology will enable next-generation AI infrastructure."



The global media also took note of this development. IT specialist outlet The Register reported, "Broadcom's custom ASIC business has added Korea's FuriosaAI," and highlighted the significance of combining Broadcom's AI networking technology with FuriosaAI's inference-optimized architecture. The analysis suggests that this collaboration is a solution to the data movement bottleneck between servers and racks that arises as AI data center scales increase.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing