Cisco and Nvidia Unveil USFP Pluggable Optic Module for AI Data Centers at OCP Summit

Cisco and Nvidia this week unveiled plans to jointly develop an ultra-dense small form factor pluggable (USFP) optic module designed specifically for AI data centers, marking a significant step forward in the optical interconnect technology that underpins large-scale GPU clusters. The announcement was made at the OCP Global Summit in San Jose, California, underscoring the industry's push toward open, non-proprietary solutions for next-generation AI infrastructure.

The USFP module is capable of delivering 400 Gbps-per-lane, representing a twofold increase over the current 200 Gbps-per-lane standard. More importantly, it is expected to deliver four times the bandwidth density and 1.3 times the power efficiency of existing 200 Gbps-per-lane optics. The development comes in direct response to demand from hyperscalers and neocloud customers who want to scale their AI workloads without overhauling their existing data center infrastructure.

Will Eatherton, Cisco Senior Vice President of Engineering, explained that pluggable modules serve a critical role in AI data centers by connecting servers housing GPUs to switches. "They have the fiber optic pigtail coming out of them, and these plug into both servers as well as into the switches, and then the optic runs between, and that's how you connect up your cluster of hundreds of thousands to millions of GPUs," he said. The current OSFP technology is reaching its performance ceiling, and data centers require an upgrade path for the future.

The "lanes" in question refer to the tiny copper pathways that carry electrical signals between the chip and the optical module. By doubling the per-lane bandwidth from 200 Gbps to 400 Gbps, Cisco and Nvidia aim to give customers a familiar form factor that can support next-generation AI workloads. Eatherton noted that customers specifically requested something resembling the old technology but with next-generation capabilities, rather than a radical architectural shift.

Competing approaches are also in development. Arista and Broadcom are working on a solution called XPO that involves liquid cooling directed to the module, while co-packaged optics (CPO) represents another alternative. However, Eatherton said Cisco's customers have made clear that what they really want is an evolution of existing pluggable technology rather than a complete redesign. "These modules will have massive engineering," Eatherton said. "You have an ecosystem. So, what we're announcing is we have simulated and tested and are ready for this module to move into the next stage. And Cisco and Nvidia will champion it."

Cisco and Nvidia plan to work with connector vendors and other ecosystem partners to form an open standard around USFP. The decision to announce at the Open Compute Project event reflects that commitment to openness, as OCP focuses on technologies that are not proprietary or locked in. Interestingly, even though Cisco and Nvidia are collaborating on USFP while Arista and Broadcom pursue the competing XPO standard, Eatherton suggested that ultimately the industry will likely support both standards to ensure customers can choose whichever product best suits their needs.

While neither company has disclosed specific roadmaps or timelines, Eatherton indicated that development is moving rapidly. "It will take a matter of months, not years," he said. The target market for USFP includes AI hyperscalers, neoclouds, telecommunications operators building sovereign AI clouds worldwide, and even high-end enterprises. The module's ability to quadruple bandwidth density while improving power efficiency addresses one of the most pressing bottlenecks in AI data center design, where the demand for GPU-to-GPU interconnect bandwidth continues to surge with each new generation of AI accelerators.