Cisco and Nvidia’s new Ultra-Dense Small Form Factor Pluggable (USFP) module targets a specific constraint in Ethernet AI fabrics: fitting more connection capacity into the space available for optics. The design, presented at the Open Compute Project (OCP) Global Summit 2026, supports electrical signaling at 400 Gb/s per lane, with a scaling path to 12.8T. For network teams, the practical question is how that density changes fabric expansion without disrupting optical maintenance.
Connection Density Affects Fabric Expansion
AI infrastructure puts pressure on the links joining GPUs, switches, and security appliances, Cisco says. Adding bandwidth can also consume more room for connectivity hardware, making physical density part of the network design decision. USFP is intended to raise capacity without requiring space to grow at the same rate.
Nvidia describes the topology consequence: once a switch exhausts its capacity, expansion can require extra switching layers and optical connections. Those additions can increase latency, electricity use, and cost, the company says. Where does the constraint sit? Teams evaluating the design should distinguish a shortage of bandwidth per connection from a shortage of connections available on a switch, because Nvidia identifies both as requirements for scaling AI factories.
Denser optics could give planners more room to expand within that constraint. They do not, on the information provided, establish how many switching layers a particular deployment could avoid. Treat topology simplification as something to evaluate against the intended fabric, rather than an automatic benefit of the module.
Separate the 6.4t Design From the 12.8t Scaling Path
Bill Gartner, senior vice president and general manager of Cisco’s optical systems and optics business, and Will Eatherton, senior vice president of networking engineering, describe USFP as supporting systems and ASICs based on both 200G and 400G SerDes. Their stated starting point is a transition beyond existing 1.6T fabrics. With 16 electrical lanes running at 400 Gb/s each, they say the module would provide aggregate bandwidth of 6.4T.
The distinction is lane count. Gartner and Eatherton describe an architecture that can expand to 32 lanes, taking aggregate capacity to 12.8T. They also claim that the 6.4T arrangement offers four times the bandwidth density of current 1.6T pluggables.
For implementation planning, keep these figures separate:
| Design factor | Stated capability or direction | Planning relevance |
|---|---|---|
| SerDes support | 200G and 400G systems and ASICs | Evaluate the module alongside the intended silicon generation. |
| 16-lane arrangement | 6.4T aggregate capacity | Use this figure when assessing the described configuration. |
| 32-lane expansion | Path to 12.8T aggregate capacity | Treat the higher figure as architectural scaling potential. |
| Power efficiency | Opportunity for approximately 1.3x improvement using 400G SerDes | Keep the vendor’s efficiency estimate distinct from deployment measurements. |
Gartner and Eatherton frame the efficiency opportunity as a way to address increasing power and cooling requirements. That claim is useful for evaluating the direction of the technology, but it is not a measured outcome for a customer installation. Capacity and efficiency should remain separate inputs when comparing a future USFP-based design with an existing fabric.
Keep the Optical Maintenance Model Familiar
USFP is intended to retain field replacement and established optical deployment practices, Gartner and Eatherton say. That gives the proposal an operational appeal beyond its capacity figures. Network teams could assess denser connectivity while retaining a familiar approach to replacing modules.
What changes for the maintenance team? The companies’ stated goal is to carry forward the pluggable model as signaling speeds increase, rather than require a different service workflow. Nvidia also says customers are intended to be able to move from 200 Gb/s signaling per lane to 400 Gb/s with this design without overhauling their data-center infrastructure.
Treat those goals as design intent rather than evidence from deployments. The announcement gives teams a basis for assessing whether USFP can preserve optical servicing practices, but compatibility with every installed platform remains unestablished. Existing maintenance habits and supported hardware combinations need to be considered separately.
Evaluate Openness Alongside Capacity
Cisco and Nvidia plan an open USFP specification through a forthcoming multi-source effort led by the community. Procurement teams have a reason to follow that work: the proposal covers a shared pluggable format as well as a capacity increase. The announcement outlines how that process is intended to work; the specification isn’t complete.
Start by assessing where fabric expansion is constrained: bandwidth per connection, connectivity on the switch, or the space optics occupy. In that assessment, distinguish the described 6.4T configuration from the path to 12.8T, and keep the vendors’ qualifications attached to their efficiency and infrastructure continuity claims. When the design is ready for evaluation, factor the available specification and platform support into the deployment decision.
