The Key Weapon Driving AI Data Center Upgrades: NVIDIA vs. Broadcom CPO Showdown
Updated: 16 hours ago
As AI data centers expand rapidly, optical interconnect technology is undergoing a major transformation. Co-Packaged Optics (CPO), a key next-generation technology for high bandwidth and low power, is steadily moving from the lab to volume production. This article compares the CPO strategies and technical approaches of the two current leaders, NVIDIA and Broadcom, and unpacks the system-architecture thinking and industry impact behind them.
Related Articles
Why Is CPO the Future?
As AI training scale and inference deployments keep growing, interconnect bandwidth demand between GPUs/XPUs is exploding. Traditional pluggable optics are constrained by the pluggable interface, power density and thermal bottlenecks, and increasingly struggle to meet the high-bandwidth, low-latency requirements of hyperscale AI systems. Integrating optical components directly into the switch ASIC package, the CPO (Co-Packaged Optics) architecture, is therefore becoming an inevitable step in the technology's evolution.
However, the real value of CPO lies not only in higher bandwidth but in effectively solving the power problem. In large AI systems, transmission energy has become a key bottleneck for overall system efficiency and operating cost. By shortening the electrical-to-optical conversion distance and eliminating traditional I/O bottlenecks, CPO can significantly reduce power per transmitted bit and improve overall data center energy efficiency.
NVIDIA's Strategy and Technical Approach
Flagship product: CPO Switch (internally referred to as the CX7 platform)
Packaging-centric view: NVIDIA starts from system architecture and treats optical interconnect as part of the SoC rather than a bolt-on module. In NVIDIA's photonics architecture, the PIC and EIC are stacked and packaged together, truly achieving co-packaging for a more compact and efficient design. The corresponding challenge is manufacturability of the package.
Better CPO power density: by integrating optical transceiver components, NVIDIA's CPO design achieves a 3x increase in interconnect density at the same power. Put simply, with a CPO design, the same power budget yields three times the compute headroom in an overall AI system.
Prefers an external laser: NVIDIA supports a separated light source architecture to ease thermal management and serviceability.
System benefits first: from a platform integration and deployment perspective, NVIDIA stresses that CPO investment should be evaluated on total TCO and scalability.
🔍 Core philosophy: every watt represents data center value; cutting transmission power means cutting cost.
Broadcom's Strategy and Technical Approach
Flagship platforms: Bailey / Davisson CPO
Compared with NVIDIA, Broadcom's CPO concept and progress are actually more complete and more mature, but the two start from different business positions. NVIDIA focuses on the total compute solution, while Broadcom supplies the full range of chips needed to build networks, including optoelectronic devices such as lasers and photodetectors.
Its CPO design philosophy is as follows:
Modular design orientation: acting as a module supplier, Broadcom provides complete CPO platforms for customers (such as Microsoft and Meta) to integrate into their systems.
Complete ADK (Assembly Design Kit): focused on helping the supply chain scale, with manufacturing-friendly assembly design data.
External laser: also an external laser, but unlike NVIDIA, laser development and sales are Broadcom's core heritage business.
Close collaboration with OSATs: actively building a CPO ecosystem that links optical components, packaging and test.
📦 Core philosophy: build a manufacturable, deliverable CPO platform to accelerate industry adoption.
NVIDIA vs. Broadcom Comparison
Item | NVIDIA | Broadcom |
Technical orientation | System architecture first, platform integration | Module supply oriented, emphasizes manufacturing and delivery capability |
Laser integration | Prefers external laser | Prefers external laser, owns key laser design |
CPO platform | Photonics Switch | Bailey / Davisson |
Power and interconnect density | 3x interconnect density per watt | Emphasizes low power, module thermal management and BER performance |
Business model | In-house platform integration, in-house chips | Supplies modules to third-party customers (e.g., hyperscalers) |
Go-to-market approach | Emphasizes overall data center efficiency and scale-out | Focuses on design for volume production and supply chain build-out |
Conclusion: Who Will Lead the CPO Battle?
Although NVIDIA and Broadcom take very different strategies and market positions, both are core forces pushing CPO technology toward maturity.
Leveraging its dominant position in the AI ecosystem and strong platform integration capabilities, NVIDIA is working to build CPO into its overall compute architecture. While NVIDIA has publicly launched its CPO switch solution, it has not yet released real test data, and the industry is watching its performance closely.
Broadcom, by contrast, focuses on modularity and supply chain scale, driving faster CPO deployment from the manufacturing side and design specifications. As CSPs choose to develop their own ASICs as compute cores, Broadcom's neutrality and complete module offering make it an important non-NVIDIA option. CPO's advantages in power and interconnect density also give it the potential to become the core architecture of future high-performance data centers.
Overall, these two companies are currently the most capable and influential drivers in terms of technology, resources and market clout. The biggest challenges for CPO have been supply chain integration and packaging innovation, and NVIDIA's and Broadcom's aggressive investment promises breakthrough progress that could bring this critical technology to commercial deployment faster.
Comments