NVIDIA and Broadcom start mass production of co-packaged optics switches

The shift from pluggable transceivers to integrated co-packaged optics marks a fundamental architectural change for the high-end networking fabric that supports intensive video processing and AI workloads. For data center operators, the 70% reduction in power consumption provides a critical path to scaling density without exceeding thermal limits. However, because these switches now compete for the same specialized packaging capacity as flagship AI GPUs, near-term supply will likely remain tight. Streaming infrastructure players should watch for the pace of 400 Tb/s capacity expansion in late 2026 as a signal for when this technology will move beyond early hyperscale adopters into broader commercial availability.

The transition to co-packaged optics (CPO) is accelerating as traditional copper and pluggable optical links hit physical limits. Per TSMC’s July 2026 earnings report, the company has officially moved its Compact Universal Photonic Engine (COUPE) technology into mass production to support this demand. Industry analysts at TrendForce noted in March 2026 that while CPO currently represents only about 0.5% of AI data center optical modules, it is projected to reach 35% by 2030 as the 'power wall' forces integration closer to the silicon.

Strategic supply chain moves underscore the intensity of this shift. In early 2026, NVIDIA secured future capacity by investing a combined $4 billion into optical leaders Lumentum and Coherent through multi-year procurement deals, per XenoSpectrum. This mirrors a broader trend toward vertical integration where chip designers are competing directly for raw silicon photonics and packaging resources. TSMC’s Chairman C.C. Wei recently confirmed the company is expanding its 'advanced packaging cluster' in Chiayi to address these bottlenecks, with newer Phase II facilities intended to increase 2.5D/3D capacity.

Simultaneously, the competitive landscape for 102.4 Tb/s systems is taking shape. Broadcom’s Tomahawk 6 family, featuring the 'Davisson' CPO variant, entered the market to compete with and complement NVIDIA’s InfiniBand and Ethernet-based Spectrum-X lines. While 800G and 1.6T pluggable modules continue to serve the majority of current deployments, per LightCounting in July 2026, the 70% power efficiency gain seen in CPO is becoming the deciding factor for the next generation of AI-optimized data center fabrics.