Co-Packaged Optics (CPO): The Inflection Point in AI-Driven Data Center Infrastructure (2026–2036)

The global data center landscape is undergoing a seismic shift, driven by the exponential growth of artificial intelligence (AI) workloads. At the heart of this transformation lies co-packaged optics (CPO), a technology poised to redefine the economics and performance of high-speed interconnects. As market analysis shows, the CPO market is projected to grow at a compound annual rate of 37% from 2026 to 2036, surpassing $20 billion by 2036. The competition between industry titans NVIDIA and Broadcom has intensified. This article evaluates their divergent ecosystem strategies and foundry roadmaps to assess which is better positioned to dominate the CPO-driven AI infrastructure race.

NVIDIA: Vertical Integration and Full-Stack Optimization

NVIDIA's CPO strategy is rooted in vertical integration, aiming to create end-to-end solutions that optimize performance, power efficiency, and scalability. The company's Spectrum-X Photonics and Quantum-X Photonics platforms exemplify this approach, integrating photonic and electronic components directly onto switch ASICs. By leveraging TSMC's System on Integrated Chips (SoIC) and 3D hybrid bonding technologies, NVIDIA achieves unprecedented integration density, enabling systems with up to 409.6 Tb/s of bandwidth and 512 ports at 800 Gb/s.

A critical enabler of this strategy is the Micro Ring Modulator (MRM) silicon photonics engine, which delivers 200Gbps PAM4 modulation per wavelength with ultra-low power consumption (1–2 picojoules per bit). However, MRMs are sensitive to thermal fluctuations and manufacturing variations, necessitating advanced packaging and direct liquid cooling solutions. This reliance on TSMC's COUPE platform and SoIC-X advanced packaging technology introduces risks related to capacity constraints and premium pricing.

NVIDIA's roadmap emphasizes full-stack optimization, embedding CPO into its broader AI ecosystem. Products like the NVLink and InfiniBand interconnects ensure high performance but raise concerns about interoperability and cost. Commercial availability of CPO-based switches is slated for 2026, with a focus on hyperscale AI training clusters. While this approach maximizes bandwidth density, it may struggle to gain traction in markets prioritizing cost efficiency and modularity.

Broadcom: Modularity, Scalability, and Ecosystem Maturity

Broadcom's CPO strategy contrasts sharply with NVIDIA's, prioritizing modularity and scalability to simplify adoption. Its third-generation 200G/lane CPO product line, including the Bailly and upcoming Davisson platforms, leverages Mach-Zehnder Modulators (MZMs) and 2.5D/3D integration on silicon interposers. This design choice enhances thermal stability and manufacturing maturity, with MZMs already achieving 200Gbps per lane in production.

Broadcom's modular approach allows its CPO solutions to function as "drop-in" enhancements to existing switch families, reducing deployment complexity for hyperscalers and enterprises. Early deployments with partners like Meta highlight its practicality, while collaborations with Corning, Delta Electronics, and Foxconn ensure robust supply chains for fiber, connectors, and component production. The company's 51.2 Tbps Bailly switch demonstrates significant power savings and bandwidth density improvements over traditional pluggable transceivers.

Broadcom's ecosystem strategy also emphasizes total cost of ownership (TCO). By leveraging TSMC's COUPE platform for standardized photonic-electronic integration, it accelerates development cycles while maintaining flexibility. This approach aligns with customer demands for reliability and scalability, particularly in scale-out and scale-up networking scenarios.

Foundry Partnerships: TSMC's Role and Capacity Constraints

Both companies depend heavily on TSMC's advanced packaging capabilities, but their strategies expose different vulnerabilities. NVIDIA's reliance on 3D heterogeneous integration and SoIC-X requires access to TSMC's most advanced nodes, which are capacity-constrained and expensive. This could limit scalability in high-volume markets.

Broadcom, by contrast, benefits from TSMC's COUPE platform, which provides a standardized path for photonic-electronic integration. While its modular design offers some flexibility, it still faces bottlenecks in high-end packaging processes. The ability of TSMC to scale CPO production will be a critical determinant for both firms, but Broadcom's approach may better withstand capacity constraints due to its modular architecture.

Market Projections and Investment Considerations

The CPO market's rapid growth underscores the importance of ecosystem scalability. NVIDIA's focus on high-density, power-efficient solutions targets premium AI training clusters, while Broadcom's modular designs cater to a broader range of applications, including enterprise and hyperscale scale-out networks.

Investors must weigh the trade-offs between performance and practicality. NVIDIA's vertically integrated approach could cement its leadership in cutting-edge AI infrastructure but risks being outpaced by Broadcom's more pragmatic, scalable solutions. Conversely, Broadcom's modular strategy may dominate in markets prioritizing TCO and interoperability, particularly as AI workloads diversify.

Conclusion: The Path to CPO Market Dominance

The CPO race hinges on execution, ecosystem maturity, and customer preferences. NVIDIA's aggressive vertical integration and TSMC partnerships position it as a leader in performance-driven AI infrastructure, but its reliance on advanced packaging and proprietary technologies introduces risks. Broadcom's modular, scalable approach, supported by a mature ecosystem and proven supply chains, offers a compelling alternative for markets prioritizing reliability and cost efficiency.

As the CPO market evolves, investors should monitor TSMC's capacity to scale advanced packaging, customer adoption trends, and technological breakthroughs in optical I/O speeds beyond 12.8T. While both companies are formidable, Broadcom's balanced strategy may provide a more sustainable path to long-term dominance in the AI-driven data center era.