On October 12, 2026, NVIDIA and Cisco announced design plans for USFP, a new optical module for AI data centers. While keeping the structure that lets modules be inserted and removed from the front of the equipment, the companies aim for four times the bandwidth density and roughly 1.3 times the power efficiency of existing 200Gb/s-per-lane products. Expansion to a maximum of 12.8Tb/s is also in view, but the first configuration they plan is 6.4Tb/s. The aim of the design is not only to increase traffic, but also to limit the footprint of network equipment and the maintenance effort it requires.
The Path to 12.8Tb/s Starts with 16 Lanes
USFP stands for "Ultra-Dense Small Form Factor Pluggable," a new form factor and connection method for optical transceivers, which convert between electrical and optical signals. The first module the companies plan fits 16 full-duplex lanes into a housing close in size to the widely used OSFP. Full duplex means a lane can transmit and receive at the same time. The electrical signal on each lane supports 400Gb/s, and the module is intended for use in both switches and NVIDIA SuperNICs.
The "4x" and "12.8Tb/s" figures in the announcement do not describe the same configuration. An explanation by a Cisco technical executive separates the initial configuration from future expansion, as follows.
| Comparison / configuration | Speed per lane | Lanes | Module capacity | Status at announcement |
|---|---|---|---|---|
| Current pluggable optical module (baseline) | 200Gb/s | 8 | 1.6Tb/s | Existing generation |
| Initial USFP | 400Gb/s | 16 | 6.4Tb/s | First planned configuration |
| Expanded USFP | 400Gb/s | 32 | 12.8Tb/s | Future expansion |
Capacity is the product of lane speed and lane count: 16 × 400Gb/s = 6.4Tb/s, and 32 × 400Gb/s = 12.8Tb/s. The current 8 lanes can also be confirmed by dividing the baseline 1.6Tb/s by 200Gb/s. Because the initial configuration doubles both the lane count and the speed of each lane, the explanation is that it fits four times the 1.6Tb/s bandwidth into a comparable footprint.
In other words, 12.8Tb/s is not the sum of 6.4Tb/s of transmit and receive; it is the capacity at the stage where the module grows to 32 lanes. Assuming 12.8Tb/s is available from the first products would cause a twofold error in the bandwidth expected at deployment. The joint announcement also states that the design can scale to 32 lanes or more in the future.
Contact Layout and Insertion Method Are the Key to More Bandwidth
Through USFP, the two companies aim to accommodate more devices connected at 1.6Tb/s and to scale AI systems without adding network tiers. When a switch's connection capacity runs out, more switches and optical links must be added, which can increase latency, power, and cost. NVIDIA's explanation published the same day also stresses limiting the addition of network tiers while growing compute resources.
Front-panel area is finite. If more bandwidth can be delivered from the same space, the pressure to enlarge equipment just to line up modules is eased. However, this does not increase the processing capacity inside the switch or the cooling capacity of the data center. How many more devices can actually be connected also depends on the design of the overall system that incorporates USFP.
A new two-dimensional array connector handles the 400Gb/s high-speed signals. It uses a two-stage structure in which the high-speed signal contacts mate only after the module is inserted, which the companies say reduces contact wear and improves signal quality. The force needed for insertion also becomes more predictable. The companies explain that this connector will support adding lanes in the future and will also improve performance at 200Gb/s per lane.
What carries over is the operating model of swapping modules in the field. A description of dimensions close to OSFP does not mean the module can be dropped straight into an existing OSFP socket. Because a new connection method is used, compatibility with existing equipment will need to be checked separately once supporting devices and specifications are published.
The 1.3x Efficiency and 100W Are Different Numbers
The power efficiency shown in the joint announcement of October 12 is a value of about 1.3 times that of pluggable optical modules running 200Gb/s per lane. This means more data can be moved with the same energy. Reading it as "power consumption falls by 30%" would change the calculation.
Converting 1.3x efficiency into energy for the same amount of traffic gives a reduction of about 23%. If efficiency is defined as data divided by energy, and a fixed amount of data is transferred within the same comparison scope, the energy required is 1 ÷ 1.3 ≈ 0.769 times. The reduction is therefore (1 − 1 ÷ 1.3) × 100 ≈ 23%.
This is a conditional conversion obtained by taking the reciprocal of the companies' stated multiple, not a power saving measured from USFP. The joint announcement gives no details on the scope of power measurement or test conditions, and the actual power draw of each module cannot be read from it, so it cannot be said that the total power of a data center, including cooling, would fall by 23%. And if traffic itself is increased, whether total power falls is a separate calculation.
The figure of up to 100W, on the other hand, indicates heat-dissipation capability. The companies explain that the module handles this heat load without a dedicated, independent cooling circuit inside the transceiver. This does not mean the equipment's own cooling systems become unnecessary, nor is it a measured value showing that the module constantly draws 100W.
Although they appear side by side in the same announcement, 1.3x concerns transfer efficiency per unit of energy, while 100W concerns thermal design. To size power and cooling capacity, network designers need to verify, for the specific modules and equipment they adopt, how much power is actually used under real load and how that heat is removed.
Field-Swappable Optics Alongside CPO
In explaining USFP, Cisco states that pluggable, near-packaged, and CPO (co-packaged optics, which integrate optical and electronic components in the same package) approaches each have advantages and constraints. It does not take the view that AI networks will converge on a single optical connection method. Its stance is that customers will choose technologies to fit their operational needs and long-term facility plans.
What USFP seeks to carry forward is the ability to replace a failed module individually in the field and to draw on a mature supply chain. Interoperability, meaning the ability to combine products from different suppliers, is also among the advantages Cisco cites. It can be read as a design aimed at users who decide on deployment by considering replacement work and parts procurement, not just power per unit of bandwidth.
To turn this concept into products, NVIDIA and Cisco will work with TE Connectivity and Amphenol to develop the cage that holds the module, the connector, and the module itself. They also set out a plan to draw up an open industry specification through a community-driven, multi-supplier project. However, this is not an announcement that the public specification is complete or that interoperability testing with other vendors' products has been finished.
The October 12 joint announcement and Cisco's explanation include no commercial release date, pricing, or independent performance measurements. The inputs for deployment decisions will be the supporting equipment that appears in the future, along with power consumption and heat-dissipation conditions under real load. Once those are available, AI data center operators will be able to evaluate USFP as an option for increasing communication capacity while keeping a familiar swap-and-maintain approach.