With the explosive growth in demand for generative AI computing power, efficiently connecting tens of thousands, or even millions, of XPUs (accelerated processing units) has become a fiercely contested battleground for semiconductor companies. At OFC 2026, the American Fiber Optics Communications Exhibition held in Los Angeles this year, Broadcom showcased its strong technological portfolio, not only expanding its end-to-end product portfolio for gigawatt-scale AI clusters, but also revealing the future development of networks moving towards 200T transmission scale.
In an earlier communication with Taiwanese media, Natarajan Ramachandran, Director of Product Marketing for Broadcom's Physical Layer Products (PLP) division, and Rajiv Pancholy, Director of Optical Systems (OSD) division, shared Broadcom's in-depth observations on the AI Fabric architecture.
From "Vertical Scaling" to "Horizontal Scaling": Broadcom's "All-Round" Connectivity Strategy
The core of Broadcom's demonstration this time is to address the power and bandwidth challenges encountered by AI clusters in "vertical scaling" (Scale-up) and "horizontal scaling" (Scale-out).
"Vertical Expansion":Focusing on high-speed connections between GPUs within a rack, Broadcom has launched a PCIe Gen6 switch and a 3.5D XDSiP platform, which achieves 3D-IC integration through Face-to-Face (F2F) technology. This is crucial for custom accelerators that pursue ultimate performance.
"Horizontal expansion":The focus was on rack-to-rack networking, with Broadcom showcasing the industry's only mass-produced 102.4T Tomahawk 6 Ethernet switch and the Thor Ultra 800G NIC designed for hyperscale clusters.
Natarajan Ramachandran emphasized: "Optical digital signal processors have become the cornerstone of AI data centers. Without these high-performance optical digital signal processors (such as Broadcom's 400G/lane Taurus solution), AI clusters simply cannot continue to expand."
A direct confrontation with NVIDIA? How does Broadcom view the "paradigm shift" in co-package optics?
The market is currently paying close attention to NVIDIA.Shift in the product roadmap—In particular, starting with the Vera Rubin architecture, the common package optics (CPO) design was introduced in Spectrum switches, and it is expected that when the Feynman architecture generation is introduced in 2028, the NVLink connector will also adopt the common package optics design.
Rajiv Pancholy pointed out that Broadcom has been deeply involved in co-package optical technology for many years, and NVIDIA's subsequent shift to co-package optical design just confirms that Broadcom's long-term investment direction is correct. He further emphasized that as transmission rates move from 200G and 400G to higher transmission rates, the transmission distance limit of traditional copper wires has reached a bottleneck. Therefore, he believes that opticalization is an inevitable trend, and Broadcom has been investing in the development of this technology for a long time.
In response to NVIDIA's closed ecosystem, which tends to sell "whole rack systems," Broadcom emphasized that "open standards" can drive more diverse development opportunities. It also explained that Broadcom has formed an alliance with companies such as AMD, Meta, NVIDIA, Microsoft, and OpenAI to promote open specifications for "optical computing interconnects," thereby promoting the development of larger-scale AI cluster architectures.
Furthermore, Broadcom believes that through open standards, hyperscalers can flexibly match the optical technologies and XPUs of different vendors. For example, they can be paired with NVIDIA GPUs or choose to use GPU products provided by AMD, without being restricted by a single system vendor.
New technologies and supply chain challenges
In response to industry proposalsReplacing laser light sources with microLEDsNatarajan Ramachandran holds reservations about this cost-saving idea. He analyzes that any new technology, in order to gain a foothold in the 400G or even higher-specification market, must first prove its stability in the low-speed market. Currently, microLEDs face severe bandwidth limitations and the difficulty of fiber optic cabling (requiring a huge number of fibers), making it difficult to shake the position of EML or Silicon Photonics (SiPh) in the short term.
Regarding the supply chain, Natarajan Ramachandran pointed out that the lead time for small PCBs used inside optical transceivers has surged from 6 weeks to 6 months. This is mainly because the 200G generation specifications require complex mSAP processes, and only a few Taiwanese and Chinese suppliers (such as TCM and Zhen Ding, related process suppliers) can provide stable production capacity. Broadcom is currently ensuring stable production capacity by signing long-term contracts of 3-4 years.
Conclusion: The "Plug and Play" Vision of the 200T Era
Broadcom outlines its vision for the future of AI infrastructure: more energy-efficient, more open, and more fully optical. As network architecture shifts from traditional cable signal extension to vertically integrated optical extension, Broadcom, with its vertical integration capabilities from switch chips to optical digital signal processors, touts its ability to steadily lead the next evolution of AI network architecture.
In contrast to NVIDIA's product strategy, Broadcom emphasizes that it has always maintained an open attitude. Therefore, although there is market competition between Broadcom and NVIDIA, they are also in a cooperative relationship. However, Broadcom expects to build greater development flexibility through an open market strategy.
