Co-Packaged Optics Race: Strategic Approaches from NVIDIA and Broadcom

Over the past decade, switch bandwidth has scaled alongside hyperscale cloud and distributed workloads, with AI now dramatically accelerating this trend. This growth has been enabled by high-speed pluggable optical transceivers. However, scaling within existing QSFP and OSFP form factors is becoming increasingly difficult, as connector density, front-panel bandwidth, and especially power consumption are emerging as major constraints.

 

Looking ahead to 1.6T optics and beyond, switch I/O must move from 112G-class to 224G-class electrical SerDes. At these data rates, the electrical channel across the package, PCB traces, and connectors becomes extremely lossy and difficult to equalise, causing a sharp rise in SerDes power.

 

 

Co-packaged optics addresses this limitation by placing the optical engines in the same package as the switch ASIC, drastically shortening the electrical path between the switch die and the optics. This reduces equalisation requirements, improves signal integrity, and relaxes front-panel density and thermal constraints. As a result, CPO is gaining significant attention as the next architecture for next-generation switching.

 

The shift toward co-packaged optics is also reshaping competitive positioning among semiconductor vendors. This article examines the differing approaches of NVIDIA and Broadcom, with further detail available in IDTechEx's latest report, "Co-Packaged Optics (CPO) 2026-2036: Technologies, Market, and Forecasts".

 

 

 

 

NVIDIA's role in AI infrastructure originates from high-performance computing. The company is a global leader in GPUs and has extended this position into AI infrastructure by combining accelerators, networking, and software platforms for machine learning and simulation. Its platform relies on tight integration between hardware and software, particularly CUDA and NVLink, which together enable communication within accelerator systems.

 

Broadcom occupies a different position in the ecosystem. The company is a market leader in networking ASICs and supplies custom ASICs to hyperscale cloud providers such as Meta, Google, and Amazon. Its products form the switching backbone of large-scale data-center networks.

 

 

In the scale-up domain, NVIDIA's leadership is associated with NVLink, which tightly connects accelerators within a system. NVLink is a short-reach, copper-based electrical interconnect that provides high-bandwidth, low-latency GPU-to-GPU communication and enables memory sharing across multiple GPUs. Together with CUDA, the interconnect, GPU memory hierarchy, and software stack are co-optimised, allowing multiple GPUs to function as a single logical compute resource.

 

Broadcom approaches this domain through Ethernet expansion. Its Tomahawk Ultra platform, part of the Scale-Up Ethernet (SUE) initiative, is designed to extend Ethernet connectivity into areas traditionally served by Cu-based interconnects across accelerators. Fabricated on TSMC 5 nm technology, Broadcom states the platform can support substantially more directly connected chips than current NVLink switch configurations, positioning Ethernet as a chip-to-chip connectivity alternative.

 

 

In scale-out infrastructure, NVIDIA is introducing optical networking products aimed at cluster connectivity. The company has announced the Quantum-X Photonics InfiniBand switch and the Spectrum-X Photonics Ethernet switch, both targeting distributed AI clusters.

 

Broadcom's primary focus remains Ethernet switching. Its Tomahawk 6 and Jericho 4 platforms are designed for large distributed clusters and long-haul data transmission. The Jericho4 device functions as an Ethernet fabric router/switch specifically intended for distributed AI infrastructures and supports lossless RDMA over Converged Ethernet (RoCE) over long distances.

 

 

Beyond individual clusters, both companies also target inter-site connectivity. NVIDIA's Spectrum-XGS extends the Spectrum-X Ethernet suite by integrating switch silicon, SuperNICs, and photonic front-end to support large AI clusters across sites. The company reports performance improvements in cross-data-center communication using NCCL (NVIDIA Collective Communications Library) relative to typical approaches.

 

Broadcom positions Jericho4 similarly as a fabric solution capable of operating beyond a single data-center facility, supporting distributed AI deployments over extended distances using Ethernet fabrics.

 

 

Both companies incorporate co-packaged optics using TSMC's semiconductor packaging approaches, with the optical engine integrated using the COUPE (Compact Universal Photonic Engine) platform together with SoIC-X 3D packaging technology. COUPE is designed to co-integrate electronic and photonic integrated circuits through a dedicated electrical interface that minimises coupling loss between the electronic IC (EIC) and the photonic IC (PIC). It also supports fibre attachment using both grating couplers and edge couplers with low insertion loss, providing improved optical coupling performance compared with traditional discrete photonic engine implementations. The 3D hybrid bonding technology enables high interconnect density while maintaining efficient system power operation.

 

 

NVIDIA positions optical interconnects as an integrated component of its system architecture for its customers, where networking, accelerators, and software operate together as a unified platform. The objective is platform-level integration and system efficiency.

 

On the other hand, Broadcom pursues a solution-oriented strategy centered on modularity and supply-chain scale. Instead of pursuing full vertical integration, the company offers complete modular platforms to hyperscale customers, who integrate these components into their own system architectures. Taken together, the comparison shows that co-packaged optics is not only a component-level technology shift but also part of differing product strategies across companies.

 

IDTechEx's report, "Co-Packaged Optics (CPO) 2026-2036: Technologies, Market, and Forecasts", offers an extensive exploration into the latest advancements within co-packaged optics technology. The report examines key technical innovations and packaging trends, providing a comprehensive analysis of the full value chain, from photonic components to system integrators. It evaluates the activities of major industry players and presents detailed market forecasts describing how CPO adoption may influence future data-center architecture.

 

Central to the report is the recognition that advanced semiconductor packaging is the cornerstone of co-packaged optics technology. Particular emphasis is placed on heterogeneous integration and photonic-electronic co-integration, which enable optical I/O to be incorporated directly within switch packages and support future deployment of CPO systems.

 

For more information on this report, including downloadable sample pages, please visit www.IDTechEx.com/CPO, or for the full portfolio of related research available from IDTechEx, see www.IDTechEx.com/Research/Computing.