NVIDIA and Cisco announced a joint specification for a new pluggable optics module on October 12, called Ultra-Dense Small Form Factor Pluggable (USFP), designed to carry the networking load of the next generation of AI factories. According to NVIDIA's press release, USFP is intended to be the first module to support 400Gb/s-per-lane electrical signals, enabling total capacity of up to 12.8Tb/s — a density jump the companies say will let operators scale Ethernet AI fabrics without adding network tiers or blowing up their power budgets.
The announcement is the latest sign that data center networking, long the unglamorous plumbing of cloud computing, has become a strategic battleground in the AI buildout. As clusters grow from thousands to hundreds of thousands of accelerators, the optics that connect them now shape both the economics and the physics of AI at scale. For the broader context of how infrastructure vendors are repositioning for the agentic era, our AI news section tracks the moves as they land.
What USFP Actually Changes
Today's high-performance pluggable modules largely run at 200Gb/s per electrical lane. USFP doubles that to 400Gb/s per lane while keeping the pluggable form factor that operators know how to deploy and service. According to NVIDIA's published figures, USFP delivers 4x the bandwidth and front-panel density compared with 200Gb/s-per-lane modules, at roughly 1.3x the power efficiency — meaning more data moved per watt.
That combination addresses a compounding problem in AI factory design. When a switch reaches its port capacity, the only way to add more compute is to add another layer of switches and optical links, which increases latency, capital cost, and power consumption with every tier. By doubling both the lane count and the bandwidth per lane, USFP lets more 1.6Tb/s endpoints connect through the same switching footprint, postponing the point at which operators must buy their way out of congestion with more tiers.
"Building the world's largest AI factories starts with rethinking the smallest of connections," said Gilad Shainer, senior vice president at NVIDIA, in the announcement.
Sixteen Lanes in an OSFP-Sized Footprint
The first planned USFP-based transceivers will pack 16 full-duplex lanes into a footprint the size of an OSFP module — the workhorse form factor already common in AI cluster deployments — supporting both switches and SuperNICs. NVIDIA says the design can expand to 32 lanes and beyond in the future, which would push per-module capacity well past today's ceiling.
Mechanically, the higher signal density required new connector thinking: USFP uses a two-dimensional array connector with a two-step insertion design, preserving the front-panel serviceability that makes pluggable optics preferable to soldered-down alternatives. "AI builders need an interconnect architecture that can scale performance without locking them into a more complex operating model," said Bill Gartner, senior vice president and general manager of optical systems and optics at Cisco, emphasizing that the collaboration was aimed at preserving "the serviceability, deployment flexibility and operational simplicity of pluggable optics."
An Ecosystem Play, Not Just a Module
NVIDIA and Cisco are explicitly framing USFP as an open, multi-vendor effort: TE Connectivity and Amphenol are already working with the two companies to build a multi-source hardware ecosystem around the specification. That matters because pluggable optics only succeed as standards — operators buy them by the tens of thousands, and no single vendor can supply (or price) that volume alone.
The multi-source approach also lowers the risk for switch makers and hyperscalers considering adoption. If USFP transceivers are available from multiple suppliers, the module can follow the same trajectory as earlier pluggable generations: rapid cost decline, broad interoperability, and gradual displacement of proprietary or soldered interconnects at the network's edge tiers.
Power Is the New Bottleneck
The power efficiency claims deserve particular attention. NVIDIA's stated goal for USFP is to shrink "networking's share of the overall AI factory power budget" — an acknowledgment that as compute densities rise, the energy cost of moving data is competing with the energy cost of computing. Every gigabit moved between racks costs watts, and at AI factory scale those watts translate directly into constrained grid connections and operating expense.
The optics announcement did not arrive in isolation. The same day, NVIDIA said it is opening its Vera Rubin rack architecture designs to the Open Compute Project community, and coverage from SiliconANGLE and Wccftech described radical simplifications to compute trays — stripping out cables, hoses, and fans — alongside the Cisco optics work. FourWeekMBA reported that Meta is adopting NVIDIA's Spectrum-6 Ethernet platform. Together, the moves sketch out a coordinated push: open the physical designs, densify the optics, and make Ethernet the fabric of choice for AI at hyperscale.
What to Watch
USFP is a specification with planned products, not shipping hardware — NVIDIA's language ('first planned USFP-based transceivers') leaves the timeline deliberately loose. The metrics to watch are adoption by switch vendors beyond Cisco, transceiver pricing relative to 200G-per-lane parts, and whether the 400Gb/s-per-lane electrical signaling meets its reliability targets at volume.
If it delivers, USFP could do for AI-era networking what successive QSFP generations did for cloud data centers: establish a commodity interconnect cadence that the entire industry builds on. And in a market where NVIDIA's proprietary NVLink fabric has been a key lock-in argument, an open, high-density pluggable standard — co-authored with Cisco, the largest networking vendor — is also a competitive statement about how AI factories should be built.
