Nvidia CPO production begins for Spectrum-X switches to boost AI networking

The transition to full production for Spectrum-X Ethernet Photonics validates co-packaged optics as a viable solution for the massive power and signal integrity challenges facing AI data centers. By moving optics closer to the switch ASIC, Nvidia is addressing the physical limits of traditional pluggable transceivers in Vera Rubin era AI factories. This shift creates a specialized supply chain involving Lumentum, Fabrinet, and TSMC, while competitors like Broadcom and Marvell advance their own CPO roadmaps with products like the Tomahawk 6 Davisson. Watch for whether cloud providers like Oracle and CoreWeave report measurable OpEx savings from these 5x power efficiency gains in upcoming deployment cycles, as Nvidia and Wall Street mobilize $500 billion for AI compute financing to support such infrastructure.

Broadcom is Nvidia's most direct competitor in the co-packaged optics space for AI networking. In October 2025, Broadcom shipped its Tomahawk 6 Davisson CPO Ethernet switch to early access customers, delivering 102.4 Tbps of switching capacity using sixteen 6.4 Tbps silicon photonic engines integrated on a shared package with the switch ASIC. The Davisson chip doubles the bandwidth of Broadcom's prior-generation Tomahawk 5 Bailly CPO switch and migrates lane speed from 100 Gbps to 200 Gbps per lane, with the company already developing a fourth-generation CPO targeting 400 Gbps per lane. Broadcom's VP of marketing Manish Mehta noted that the optical interconnect space for AI networks has seen an order of magnitude growth in volumes compared to earlier front-end cloud deployments. The business case for CPO in AI data centers is being validated by hyperscaler deployment data. Meta presented results showing more than one million cumulative 400 Gbps equivalent port device hours without any data drops using Broadcom's second-generation Tomahawk 5 Bailly CPO switches, along with 65% savings and a 5x improvement in serviceable failures over pluggable optics. Broadcom's Mehta stated that these results put CPO reliability targets, matching traditional silicon chips, within reach. The Davisson platform also introduces field-replaceable ELSFP laser modules accessible through the front panel, addressing a key serviceability concern that had slowed earlier CPO adoption. On the packaging and supply chain side, Broadcom is working with TSMC on its Compact Universal Photonic Engine (COUPE) packaging technique to integrate the optical engines around the switch ASIC, with Corning partnering to deliver fiber harnesses and cable assemblies connecting optical ports to the switch chassis. The Davisson package measures approximately 120 mm by 120 mm, significantly larger than the 75 mm by 75 mm Bailly predecessor. Broadcom reports that an 800 Gbps CPO port consumes about 3.5 watts, representing a 36.4% reduction versus Tomahawk 5 CPO at the same bandwidth and more than 70% lower than pluggable optics. These figures establish a competitive benchmark that Nvidia's Spectrum-X Ethernet Photonics platform must match or exceed as it enters full production.