On October 12, 2026, NVIDIA and Cisco announced plans for USFP (Ultra-Dense Small Form Factor Pluggable), a pluggable optical module built for 400Gb/s-per-lane electrical signals. Compared with today's 200Gb/s-per-lane pluggables, the companies claim 4x the bandwidth and front-panel density but only about 1.3x the power efficiency. So USFP mainly adds capacity per switch port. It does not make each module use less power. No sampling or ship date has been given, and every figure below comes from the companies.
What was announced
According to NVIDIA's newsroom release of October 12, 2026:
- Lane rate: 400Gb/s per lane, against a 200Gb/s-per-lane baseline.
- First modules: 16 full-duplex lanes in a footprint the size of an OSFP (octal small form factor pluggable) module, for both switches and SuperNICs.
- Growth path: the design can expand to 32 lanes and more.
- Capacity: up to 12.8Tb/s.
- Power: up to 100 watts of power dissipation, without a dedicated embedded cooling circuit in the transceiver.
- Partners: TE Connectivity and Amphenol are working on cages, connectors and modules, and more partners are expected.
- Specification: the companies plan an open industry specification through an "upcoming community-led multi-source project."
The announcement came during the OCP Global Summit 2026, according to Network World and StorageReview.
How the 12.8Tb/s figure is counted
The NVIDIA release gives "up to 12.8Tb/s" without saying which module that refers to. Cisco's account, as reported by Network World, is more specific:
- 16 lanes x 400Gb/s = 6.4T for the first module. Cisco's Bill Gartner and Will Eatherton call this "four times the bandwidth density of today's 1.6T pluggables."
- 32 lanes give "a path to 12.8T of aggregate capacity."
StorageReview gives the same split: 6.4Tb/s per first-generation module, with 12.8Tb/s as a design target for the 32-lane version. Read 12.8Tb/s as a ceiling for later modules, not the capacity of the first USFP transceivers.
Why 4x bandwidth is not 4x efficiency
The two multiples measure different things:
- 4x bandwidth and density: how much data fits through one front-panel slot, compared with 200Gb/s-per-lane pluggables.
- About 1.3x power efficiency: how much less energy each bit takes.
If both claims hold, one USFP module moves far more data but draws more total power than the module it replaces. As a simple calculation from the companies' own multiples (not a figure they published), 4x the throughput at 1.3x the efficiency means roughly 3x the power per module. That fits with the companies stressing support for up to 100W of dissipation. The 100W figure is a design limit, not a measured draw.
For network planners, the gain is more bandwidth per switch and fewer ports for the same fabric. Lower total optics power is not part of the claim.
How much more power does one USFP module draw?
Enter a bandwidth multiple and an efficiency multiple to see the power one module draws compared with the 200Gb/s-per-lane module it replaces.
| Bandwidth per module against a 200Gb/s-per-lane pluggable | Power efficiency (energy per bit) against a 200Gb/s-per-lane pluggable | Power per USFP module compared with the module it replaces |
|---|---|---|
| 4 x | 1.3 x | 3.1 x |
| 4 x | 1.3 x | 3.1 x |
| 4 x | 4 x | 1 x |
- The defaults are the companies' own 4x and about 1.3x multiples. The roughly 3x result is our arithmetic, not a figure NVIDIA or Cisco published.
- A result above 1 means each module draws more total power than the one it replaces. A result of 1 would mean efficiency rose as fast as bandwidth, which is not what the companies claim.
This divides two multiples and nothing else. It does not give measured watts per module, cooling or switch-level power, and it cannot check the companies' claims about a product that has no ship date yet.
The connector and cooling
USFP uses a new two-dimensional array connector with two-step insertion. The module seats first, and the high-speed contacts engage after that. The companies say this reduces contact wear, improves signal integrity and makes insertion force more predictable.
Handling up to 100W without a cooling circuit inside the transceiver is meant to simplify switch design and improve reliability, according to the announcement. None of these benefits has been independently tested.
Where it fits with Spectrum-X and Silicon One
NVIDIA says the work builds on interoperability between NVIDIA Spectrum-X Ethernet and Cisco Silicon One-based Ethernet networks. Cisco says USFP "will provide a path supporting both 200G and 400G SerDes based systems and ASICs," so it is pitched as a bridge from today's 1.6T fabrics, not only a 400G-only part.
Brian Sparks, a marketing director at NVIDIA, told Network World that USFP is "not part of the partnership. It's just Cisco and us and others in the ecosystem." In other words, it sits outside the existing NVIDIA–Cisco Spectrum-X deal.
What is still open
- Timing: no sampling schedule, availability date or specification date.
- Independent data: the 4x and about 1.3x figures have not been checked by third-party tests.
- Baseline detail: the release does not say which 200Gb/s-per-lane modules, lane counts or power figures the comparison uses.
- Specification: the multi-source project has not been launched, and its other members have not been named.
Limitations: every specification here is a company claim about a planned product, reported by the vendors and the trade press that covered the announcement.
