Cisco and Nvidia have unveiled a new optical pluggable module designed to deliver up to 12.8 terabits per second of bandwidth, targeting the growing demands of artificial intelligence data centers. The ultra-dense small form factor pluggable (USFP) standard, announced at the Open Compute Project global summit, aims to consolidate high-speed connectivity into a compact form factor roughly equivalent in size to existing optical modules. Each USFP module supports up to 32 lanes of 400 gigabits per second, collectively enabling 12.8 Tbps of Ethernet throughput. This capacity surpasses current high-end pluggables, which typically max out at around 1.6 Tbps using eight 200 Gbps links.

The collaboration seeks to streamline cabling infrastructure within AI clusters, where individual GPUs often require 800 Gbps or higher network interface connections. By consolidating connectivity for multiple GPUs into a single fiber bundle, the USFP module could reduce cable complexity in dense server environments. For instance, a future switch capable of 204.8 Tbps could support 128 endpoints using just 16 USFP ports, potentially simplifying network architecture.

Technical challenges accompany the performance gains. Cisco’s USFP specification permits up to 100 watts of power consumption per module, significantly higher than the approximately 20 watts typical of contemporary pluggables. The design employs a two-stage insertion mechanism to ensure reliable electrical contact across multiple high-speed lanes, though this increases mechanical complexity. Additionally, a single faulty component in the module could affect a larger portion of the network compared to traditional pluggables, though this risk is considered manageable relative to failures in co-packaged optics, which might disrupt entire server racks.

Nvidia’s involvement marks a notable shift, as the company has largely transitioned to co-packaged optics in its latest switch designs to improve energy efficiency. However, Nvidia continues to rely on pluggables for its network interface cards, citing flexibility for diverse deployment scenarios, such as top-of-rack switches using lower-power copper cabling. The company’s upcoming Vera Rubin Ultra platform, announced in early 2025, targets 115.2 terabits per second of inter-rack bandwidth, aligning with the 6.4 Tbps per GPU bandwidth offered by Cisco’s initial USFP specification.

The USFP standard is being released as an open specification, allowing broader industry adoption. Cisco has also emphasized compatibility with SONiC, a widely used open-source network operating system, extending its existing support for the platform. This includes plans to integrate its Live Protect security patching system into SONiC, with rollout expected in the following year. Meanwhile, Microsoft has indicated use of Cisco’s Silicon One P200 routing chips to enable multi-datacenter AI workloads, reflecting growing interest in distributed computing architectures that bypass local power limitations.

The development underscores a broader industry push to scale network bandwidth while managing power and physical constraints in AI infrastructure. As AI models grow in size and complexity, the demand for high-speed, low-latency interconnects continues to drive innovation in optical networking, with USFP representing one approach to balancing performance, efficiency, and deployability.