Nvidia Ships Co-Packaged Optics Switches at Scale as AI Data Centers Rethink Connectivity

Mass Production Milestone

Nvidia has entered mass production of co-packaged optics switches, marking a critical inflection point for optical interconnect technology in AI infrastructure. The move represents more than an incremental upgrade; it reflects a structural change in how hyperscale data centers architect networking for compute-intensive workloads.

Gilad Shainer, Senior Vice President of Networking at Nvidia, framed the development within a broader architectural evolution. Speaking ahead of the Open Compute Project APAC Summit scheduled for August 11-12, 2026, Shainer pointed to a fundamental rethinking of the computing unit itself. Where traditional data centers treated individual servers as the atomic unit of compute, AI factories now operate at data-center scale, with entire facilities functioning as unified compute resources.

That shift places unprecedented demands on optical interconnects. Co-packaged optics technology integrates photonic components directly with switch silicon, reducing signal loss, latency, and power consumption compared to pluggable optics modules. The approach has been discussed in industry forums for years, but deployment has lagged due to manufacturing complexity and thermal management challenges.

Nvidia's announcement that CPO switches have moved beyond prototyping into volume production suggests those obstacles are now manageable at commercial scale. The company has been building out its networking portfolio aggressively since acquiring Mellanox Technologies in 2020 for USD 7 billion, and CPO represents the latest layer in that strategy.

Architectural Implications

The transition to CPO carries implications beyond raw bandwidth. Traditional pluggable optics, while modular and field-replaceable, introduce insertion loss and consume rack space. Co-packaging eliminates the physical connector interface, allowing denser switch designs and improved signal integrity at higher data rates.

For AI training clusters, where thousands of GPUs communicate continuously during model runs, every basis point of latency reduction and every watt of power saved compounds across the fabric. Nvidia has positioned its networking hardware as integral to AI factory economics, arguing that interconnect performance directly determines GPU utilization rates and, by extension, return on capital for infrastructure operators.

The shift also pressures the broader optical supply chain. Pluggable optics vendors, including Broadcom, Marvell, and a constellation of specialist firms, have built business models around field-upgradable modules. CPO, by contrast, locks optical components into the switch at manufacture, reducing flexibility but improving performance. The trade-off favors operators running homogeneous, long-lifecycle deployments, a profile that fits hyperscale AI infrastructure.

Regional Context

Asia-Pacific remains a critical theater for AI infrastructure investment. Governments in Singapore, Japan, South Korea, and India have committed billions in subsidies and incentives to attract data center capital, while Chinese operators continue expanding domestic AI capacity despite export controls on advanced chips.

The OCP APAC Summit, where Nvidia's networking strategy will be discussed in detail, has become a key venue for aligning hyperscale operators, original design manufacturers, and component suppliers around open standards. Co-packaged optics has been a recurring topic at OCP forums, with participants debating interoperability standards and thermal management practices.

Nvidia's timing aligns with a broader wave of AI infrastructure build-outs across the region. Taiwan Semiconductor Manufacturing Company recently disclosed capacity expansions for advanced packaging, a capability essential for CPO manufacturing. Meanwhile, optical component suppliers in Taiwan, Japan, and South Korea are retooling production lines to support co-packaged designs, anticipating multi-year demand cycles driven by AI workloads.

Supply Chain and Competitive Dynamics

The CPO production ramp raises questions about Nvidia's supply chain depth. Co-packaging requires tight integration between switch ASICs, photonic dies, and advanced packaging substrates. Nvidia has historically relied on TSMC for silicon fabrication and a mix of outsourced assembly and test partners for final packaging.

The move also intensifies competition with Broadcom, which supplies custom AI networking silicon to Google and other hyperscalers. Broadcom has disclosed its own CPO roadmap, targeting 2027 for volume shipments. Marvell, another merchant silicon vendor, is developing CPO-compatible switch chips for cloud and enterprise customers.

For optical module vendors, the shift presents a strategic dilemma. Some are pivoting toward linear-drive pluggable optics, which offer a middle path between traditional modules and full co-packaging. Others are investing in CPO manufacturing capabilities, seeking partnerships with switch vendors to secure design wins in next-generation platforms.

The optical communications industry has historically cycled through waves of integration and disaggregation. CPO represents a swing back toward integration, driven by performance demands that outweigh modularity benefits. How far that pendulum swings, and how quickly, will depend on deployment economics and the pace at which hyperscale operators refresh infrastructure.

Nvidia's production announcement provides a data point, but the broader market transition will unfold over multiple hardware generations. For now, the signal is clear: optical interconnects are moving closer to the silicon, and the companies that control both layers hold a structural advantage in AI infrastructure.