NVIDIA and Cisco Unveil USFP: A Tiny Optical Module With 4x the Bandwidth for AI Data Centres

AI data centres are running into a surprisingly physical problem: the small optical modules that link their switches together. As AI models grow and inference demand expands, operators need more bandwidth on every connection and more connections on every switch, without adding extra layers of networking hardware that push up latency, power bills and cost. At the OCP Global Summit in San Francisco on 12 October 2026, NVIDIA and Cisco set out their answer: a new pluggable optics format called USFP.

What is USFP?

USFP stands for Ultra-Dense Small Form Factor Pluggable. It is designed as the first pluggable optical module built for 400 gigabits per second per lane, doubling the 200Gb/s-per-lane rate used by today’s generation of optics. The companies say the design delivers four times the bandwidth density and around 1.3 times the power efficiency of 200Gb/s-per-lane modules.

The first planned USFP transceivers will pack 16 full-duplex lanes into a footprint roughly the size of an existing OSFP module, giving about 6.4 terabits per second per module. The connector is designed to grow to 32 lanes and more, taking overall capacity to 12.8Tb/s. You can read more coverage of new hardware in our Technology section.

Why bandwidth density matters

In an AI factory, the front panel of a switch is prime real estate. Once a switch runs out of ports or bandwidth, the usual fix is another network tier: more switches, more optical links, more latency and more electricity. By doubling both the lane count and the bandwidth per lane, USFP is meant to let a single switch connect far more 1.6Tb/s endpoints within the same footprint, so operators can scale computing capacity without reworking the whole network.

The design also keeps the familiar pluggable model. Modules remain field-replaceable and serviceable, which matters to operators who swap optics routinely. A two-dimensional connector array with a two-step insertion mechanism is designed so that the high-speed electrical contacts engage only once the module is seated, reducing contact wear and protecting signal quality.

An open specification, built with partners

NVIDIA and Cisco are not keeping USFP to themselves. The two companies are working with TE Connectivity and Amphenol on the cages, connectors and modules, and they plan to publish USFP as an open industry specification through a community-led multi-source project, so other suppliers can build compatible hardware. Cisco set out the announcement in its OCP 2026 newsroom update, and StorageReview has a detailed breakdown of the lane counts and capacities.

What happens next

For now, USFP is a specification and an ecosystem plan rather than a product on a shelf: neither company has announced a sampling schedule or a commercial launch date. Even so, the direction of travel is clear. The bottleneck in AI infrastructure is no longer only the GPU; it is the thousands of small connections between machines. A module the size of a matchbox that moves four times as much data could quietly shape how the next generation of AI data centres gets built.