NVIDIA and Cisco announced plans for the Ultra-Dense Small Form Factor Pluggable (USFP), “a high-density optical module” for AI factories that use pluggable transceivers. (The brief called it “Ultra Small”; the companies’ name is “Ultra-Dense”.) TE Connectivity and Amphenol are named as building the cages, connectors and modules, and “additional partners are expected to join”.

The companies say USFP is the first module designed for “400Gb/s-per-lane electrical signals”, with capacity “up to 12.8Tb/s”. The first planned transceivers have 16 full-duplex lanes in an “OSFP-sized footprint”, for switches and SuperNICs (the network cards in each GPU server), with a path to 32 or more lanes. The specification is to be published through an “upcoming community-led multi-source project”.

An optical transceiver turns the electrical signal from a switch chip or network card into light for the fibre, and back again. In a GPU cluster every GPU needs one or more of them, so a large cluster runs hundreds of thousands. Their size limits how much bandwidth fits on a switch’s front panel, and their power is a real share of the network’s draw.

Each step in switch-chip bandwidth needs either faster lanes or more modules. USFP doubles the electrical lane rate of today’s 200 Gb/s-per-lane modules in the same footprint, so a switch can expose four times the bandwidth per panel without giving up modules that can be pulled and replaced.

USFP is the pluggable route; co-packaged optics (see Spectrum-X Ethernet Photonics) is the other. The industry is pursuing both.

A pluggable module sits in a cage on the switch’s front panel. The switch chip drives electrical signals across the circuit board to the module, where a digital signal processor cleans them up and lasers and photodetectors convert them to and from light. Doubling the lane rate to 400 Gb/s makes that electrical path harder: signals degrade faster, so the connector, cage and board have to be designed for it, which is why the connector makers are part of the announcement.

More bandwidth in the same space means more heat in the same space. The companies say USFP handles up to “100 watts” of power dissipation without a dedicated embedded cooling circuit.

Power & Cooling

Modules of up to 100 W each concentrate heat at the switch front panel; switch cooling (air or liquid) must be checked for the new density.

Networking

Higher radix per switch means fewer switches and fewer tiers for the same cluster, but a new optics generation to qualify. Until the multi-source specification exists, the network team is buying a two-vendor design, so second-source risk belongs on the risk register.

Operations

Modules stay hot-swappable, so the spares and replacement process carries over, but at much higher value per module.

1. Announced (current stage)Described publicly; no date for customers to buy it.
2. PlannedA stated availability window that has not arrived yet.
3. ShippingIn production and reaching early customers or partners.
4. Generally availableOrderable by any customer through normal channels.
5. DeployedRunning in customers’ production environments, confirmed by them or the manufacturer.

The companies “announced plans for” USFP. No availability date, product list or price has been given, and the multi-source project that will publish the specification has not yet been formed.

• The building blocks: networkWhere optics sit between GPUs, switches and fibre.
• Scale-up versus scale-out networkingThe scale-out fabric these modules connect.

1. NVIDIA and Cisco Introduce New Pluggable Optics Form Factor to Scale AI Factories(opens in a new tab)NVIDIA Newsroom · 12 October 2026 · Manufacturer

Figures are as the manufacturer states them unless marked otherwise. None has been independently tested by BSS/OSS Academy.